{"generated_at":"2026-10-16T14:01:47-05:00","count":283,"source":"https://chatppg.com/research","items":[{"slug":"ecgfm-u-lite-compressed-ecg-fm-ijmi-2026","title":"A lightweight ECG model with teacher-derived integrated medical semantic preservation for cardiac disease screening (ECGFM-U-Lite)","summary":"This study shrinks a unified ECG foundation model so it can run where compute and memory are limited. The result, ECGFM-U-Lite, is a compact 1D-MobileViT student built with a three-stage compression procedure. It reached an overall AUC of 0.862 against 0.918 for the teacher, keeping about 94% of the teacher's performance on rhythm and structure tasks.","venue":"International Journal of Medical Informatics","year":2027,"open_code":null,"updated":"2026-10-12","url":"/research/papers/ecgfm-u-lite-compressed-ecg-fm-ijmi-2026","links":[{"href":"https://doi.org/10.1016/j.ijmedinf.2026.106758","label":"DOI","ariaLabel":"DOI for A lightweight ECG model with teacher-derived integrated medical semantic preservation for cardiac disease screening (ECGFM-U-Lite) (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42849250","label":"PubMed","ariaLabel":"PubMed for A lightweight ECG model with teacher-derived integrated medical semantic preservation for cardiac disease screening (ECGFM-U-Lite) (opens in new tab)"}]},{"slug":"lima-mamba-ppg-biometrics-patcog-2026","title":"LIMA-Mamba: A lightweight and distribution-robust Mamba network for PPG-based biometrics","summary":"From the title only, it is a Mamba (selective state-space) network for PPG-based biometric recognition. It is designed to be lightweight and robust to distribution shift (e.g., cross-session or cross-condition PPG). It is a supervised task model, not a foundation model.","venue":"Pattern Recognition (Elsevier), vol. 183 (Mar 2027 issue), article 114955; online 2026-09-19","year":2027,"open_code":false,"updated":"2026-10-08","url":"/research/papers/lima-mamba-ppg-biometrics-patcog-2026","links":[{"href":"https://doi.org/10.1016/j.patcog.2026.114955","label":"Paper","ariaLabel":"Paper for LIMA-Mamba: A lightweight and distribution-robust Mamba network for PPG-based biometrics (opens in new tab)"}]},{"slug":"signal-representation-comparison-ppg-bp-sensors-2026","title":"A Comparative Study of Signal Representations Methods and Deep Learning Architectures for PPG-Based Cuffless Blood Pressure Estimation","summary":"Seven 1D→2D transforms × several CNN variants on the UCI BP database. Best: CWT, errors reported as 3.80 ± 5.02 (SBP) and 1.65 ± 2.70 mmHg (DBP). Real-world check on 26 people vs an Omron cuff: R = 0.96 (SBP), 0.74 (DBP).","venue":"Sensors (MDPI) 26(9):2847","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/signal-representation-comparison-ppg-bp-sensors-2026","links":[{"href":"https://doi.org/10.3390/s26092847","label":"Paper","ariaLabel":"Paper for A Comparative Study of Signal Representations Methods and Deep Learning Architectures for PPG-Based Cuffless Blood Pressure Estimation (opens in new tab)"}]},{"slug":"pat-actigraphy-mental-health","title":"A Foundation Model for Wearable Movement Data in Mental Health Research (PAT)","summary":"PAT is a masked-autoencoder transformer over one week of minute-level actigraphy (10,080 minutes per participant), pretrained on 21,538 NHANES participants and evaluated on five mental-health labels. It is the first open-source foundation model for wearable movement data aimed at psychiatric outcomes, and the clearest published demonstration that long-context behavioural structure — not short-window signal fidelity — is where movement data pays off for mental health.","venue":"IEEE J Biomed Health Inform","year":2026,"open_code":null,"updated":null,"url":"/research/papers/pat-actigraphy-mental-health","links":[{"href":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11623705/","label":"Paper","ariaLabel":"Paper for A Foundation Model for Wearable Movement Data in Mental Health Research (PAT) (opens in new tab)"},{"href":"https://doi.org/10.1109/JBHI.2026.3694809","label":"DOI","ariaLabel":"DOI for A Foundation Model for Wearable Movement Data in Mental Health Research (PAT) (opens in new tab)"},{"href":"https://arxiv.org/abs/2411.15240","label":"arXiv","ariaLabel":"arXiv for A Foundation Model for Wearable Movement Data in Mental Health Research (PAT) (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42154714","label":"PubMed","ariaLabel":"PubMed for A Foundation Model for Wearable Movement Data in Mental Health Research (PAT) (opens in new tab)"}]},{"slug":"pulseox-fm-wearable-pulse-oximetry-foundation-model-hpp-2026","title":"A foundation model of wearable pulse oximetry reveals physiological signatures of health and cardiometabolic risk (PulseOx-FM)","summary":"PulseOx-FM is a self-supervised model trained on 6,995,558 pulse oximetry segments from 42,282 overnight recordings of 10,704 people in the Human Phenotype Project. It predicted age better than existing feature methods and held up in an external cohort. Its representations predicted 64 cardiometabolic and neuropsychiatric phenotypes and flagged two-year hypertension risk in people with normal blood pressure.","venue":"openRxiv","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/pulseox-fm-wearable-pulse-oximetry-foundation-model-hpp-2026","links":[{"href":"https://doi.org/10.64898/2026.07.01.26356992","label":"DOI","ariaLabel":"DOI for A foundation model of wearable pulse oximetry reveals physiological signatures of health and cardiometabolic risk (PulseOx-FM) (opens in new tab)"}]},{"slug":"metawears-meta-learning-few-shot-af-seizure-wearable-commun-med-2026","title":"A meta-learning method with reduced data requirements for training and updating deep learning models on wearable devices","summary":"This paper proposes a meta-learning method that cuts how much labelled data a new wearable needs before its model works. In two case studies, fine-tuning on just a few samples gave 70% AUC for epileptic seizure detection and 82% for atrial fibrillation, up to 45% better than a conventional approach.","venue":"Communications Medicine","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/metawears-meta-learning-few-shot-af-seizure-wearable-commun-med-2026","links":[{"href":"https://doi.org/10.1038/s43856-026-01740-1","label":"DOI","ariaLabel":"DOI for A meta-learning method with reduced data requirements for training and updating deep learning models on wearable devices (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42850332","label":"PubMed","ariaLabel":"PubMed for A meta-learning method with reduced data requirements for training and updating deep learning models on wearable devices (opens in new tab)"},{"href":"https://github.com/esl-epfl/metawears","label":"Code","ariaLabel":"Code for A meta-learning method with reduced data requirements for training and updating deep learning models on wearable devices (opens in new tab)"}]},{"slug":"geenjaar-multimodal-ppg-fm-2026","title":"A robust PPG foundation model using multimodal physiological supervision","summary":"Pretrains a PPG-only foundation model (1D ResNet-26 encoder) on MIMIC-III Waveform Database Matched Subset (10,282 ICU patients, 125 Hz) using co-recorded ECG and respiration signals only to select which PPG windows count as contrastive positives/negatives during training — five summary metrics from ECG/RESP per 10s window build a \"physiological metric space,\" and windows close in that space become positive pairs. At inference, only PPG is used. Explicitly targets robustness to noisy/motion-corrupted PPG, including an IMU-energy-stratified DaLiA noise analysis.","venue":"ICML 2026 workshop / preprint (Dolby Labs + Georgia Tech)","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/geenjaar-multimodal-ppg-fm-2026","links":[{"href":"https://arxiv.org/abs/2606.07365","label":"Paper","ariaLabel":"Paper for A robust PPG foundation model using multimodal physiological supervision (opens in new tab)"}]},{"slug":"conformer-ppg-to-abp-cuffless-bp-lsens-2026","title":"Accurate Cuffless Blood Pressure Estimation Using Conformer-Based PPG-to-ABP Waveform Reconstruction Framework","summary":"Single-channel PPG → Conformer → ABP waveform → SBP/DBP. Validated on MIMIC-BP (BHS grade A and AAMI met; numbers not in abstract), then applied to 80 PPG signals from the authors' chest-worn red/IR wearable as preliminary feasibility.","venue":"IEEE Sensors Letters","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/conformer-ppg-to-abp-cuffless-bp-lsens-2026","links":[{"href":"https://doi.org/10.1109/LSENS.2026.3714451","label":"Paper","ariaLabel":"Paper for Accurate Cuffless Blood Pressure Estimation Using Conformer-Based PPG-to-ABP Waveform Reconstruction Framework (opens in new tab)"}]},{"slug":"actinet-ssl-harnet-hmm-activity-intensity-sci-rep-2026","title":"ActiNet: an open-source tool for activity intensity classification of wrist-worn accelerometry using self-supervised deep learning","summary":"The use of accurate and reliable open-source human activity recognition (HAR) models on passively collected wrist-accelerometer data is essential in large-scale epidemiological studies that investigate the association between physical activity and health outcomes. While self-supervised learning has generated considerable excitement in improving HAR, the extent to which these models, coupled with hidden Markov models (HMMs), would make a tangible improvement to classification performance and the effect this may have on the predicted daily activity intensity compositions is unknown. Using up to 24 h of wrist-worn accelerometer data from 151 CAPTURE-24 participants (aged 18–91, mean age 42, 66% female), the authors trained the ActiNet model, consisting of HARNet, a self-supervised, 18-layer, modified ResNet-V2 model, followed by hidden Markov model (HMM) smoothing, to classify labels of activity intensity.","venue":"Scientific Reports (published 2026-10-13); preprint arXiv 2510.01712 (v1 2025-10-02, v2 2026-04-30)","year":2026,"open_code":true,"updated":"2026-10-13","url":"/research/papers/actinet-ssl-harnet-hmm-activity-intensity-sci-rep-2026","links":[{"href":"https://doi.org/10.1038/s41598-026-65900-y","label":"Paper","ariaLabel":"Paper for ActiNet: an open-source tool for activity intensity classification of wrist-worn accelerometry using self-supervised deep learning (opens in new tab)"},{"href":"https://arxiv.org/abs/2510.01712","label":"arXiv","ariaLabel":"arXiv for ActiNet: an open-source tool for activity intensity classification of wrist-worn accelerometry using self-supervised deep learning (opens in new tab)"}]},{"slug":"ai-sleep-co-scientist-agentic-discovery-psg-2026","title":"Agentic AI-enabled discovery across large-scale sleep physiology (AI Sleep Co-Scientist)","summary":"AI Sleep Co-Scientist is a setup in which human scientists direct specialist AI agents for hypothesis development, signal preprocessing and statistics, reviewing each step. The authors used it on four cohorts of about 124,000 polysomnography recordings (more than 50 TB) for five case studies on how sleep relates to future disease, clinical phenotypes and sleep regulation.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/ai-sleep-co-scientist-agentic-discovery-psg-2026","links":[{"href":"https://arxiv.org/abs/2607.25175","label":"arXiv","ariaLabel":"arXiv for Agentic AI-enabled discovery across large-scale sleep physiology (AI Sleep Co-Scientist) (opens in new tab)"}]},{"slug":"codas-ai-co-data-scientist-wearable-biomarkers-2026","title":"An AI Co-Data-Scientist for Prioritizing Candidate Biomarkers from Wearable Sensor Data (CoDaS)","summary":"CoDaS is an AI co-data-scientist that proposes and checks candidate biomarkers from wearable data, combining multiple agents, fixed statistical tests, adversarial validation and literature grounding under human oversight. Across three cohorts with 9,279 participant-observations it flagged sleep-variability signals linked to depression and a steps-to-resting-heart-rate fitness index linked to insulin resistance (ρ = -0.374). In a review by 12 clinicians, validity ratings matched CoDaS confidence tiers (ρ = 0.67).","venue":"arXiv 2604.14615 (2026-04-16; v2)","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/codas-ai-co-data-scientist-wearable-biomarkers-2026","links":[{"href":"https://arxiv.org/abs/2604.14615","label":"arXiv","ariaLabel":"arXiv for An AI Co-Data-Scientist for Prioritizing Candidate Biomarkers from Wearable Sensor Data (CoDaS) (opens in new tab)"}]},{"slug":"doomen-grip-open-source-wrist-acc-activity-intensity-chronic-pain-fdgth-2026","title":"An open-source deep learning model for classifying real-world activity intensities in persons with chronic pain using a wrist-worn accelerometer","summary":"In chronic pain (CP) care, patterns of activity and rest are relevant across diagnostic, intervention, and evaluation phases. Public availability of a detailed dataset together with a validated deep learning (DL) model for classifying activity intensities from raw accelerometer data can promote clinical use and standardize activity intensity measurement in CP research. This study aimed to (1) develop, test, and share a DL model that classifies activity intensity from raw accelerometer data, (2) explore whether CP-specific training improves performance in CP, and (3) explore construct validity in real-world settings among individuals with and without CP, using real-world observations as comparator instrument. Seven strategies were applied to train a Convolutional Neural Network, using three datasets comprising 211,864 ten-second windows.","venue":"Frontiers in Digital Health 8:1783507 (published 2026-10-14)","year":2026,"open_code":true,"updated":"2026-10-14","url":"/research/papers/doomen-grip-open-source-wrist-acc-activity-intensity-chronic-pain-fdgth-2026","links":[{"href":"https://doi.org/10.3389/fdgth.2026.1783507","label":"Paper","ariaLabel":"Paper for An open-source deep learning model for classifying real-world activity intensities in persons with chronic pain using a wrist-worn accelerometer (opens in new tab)"}]},{"slug":"anyecg-echo-wearable-single-lead-ecg-structural-heart-disease-2026","title":"AnyECG-Echo: Wearable Single-Lead ECG Detects Fine-Grained Structural Heart Disease Through Echo-Report Supervision","summary":"AnyECG-Echo uses contrastive pretraining between single-lead ECGs and echocardiography reports to screen for structural heart disease, extending earlier single-health-system work toward clinical translation. The authors evaluate it in a geographically independent external cohort (n = 16,621), cover 13 fine-grained structural heart disease subtypes spanning myocardial, chamber, valvular and great-vessel pathologies, and add Shapley-based interpretability. Across validation cohorts totalling 25,222 patients it reached AUROC 0.866-0.924 for reduced left-ventricular systolic function, 0.877-0.931 for global heart enlargement and 0.836-0.906 for mitral stenosis.","venue":"Research Square preprint (posted 2026-10-14; CC BY)","year":2026,"open_code":null,"updated":"2026-10-15","url":"/research/papers/anyecg-echo-wearable-single-lead-ecg-structural-heart-disease-2026","links":[{"href":"https://doi.org/10.21203/rs.3.rs-11313095/v1","label":"DOI","ariaLabel":"DOI for AnyECG-Echo: Wearable Single-Lead ECG Detects Fine-Grained Structural Heart Disease Through Echo-Report Supervision (opens in new tab)"}]},{"slug":"anymo-geometry-aware-setup-agnostic-imu-2026","title":"AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild","summary":"This paper introduces AnyMo, which models human motion from IMUs placed anywhere on the body. It simulates IMU signals over dense body positions, pretrains a graph encoder, turns multi-position IMU data into full-body motion tokens and aligns them with an LLM. Across 14 unseen datasets it improved zero-shot activity recognition accuracy by 11.7%, and it also improved IMU-text retrieval and motion captioning.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-16","url":"/research/papers/anymo-geometry-aware-setup-agnostic-imu-2026","links":[{"href":"https://arxiv.org/abs/2605.22715","label":"Paper","ariaLabel":"Paper for AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild (opens in new tab)"},{"href":"https://github.com/Breezelled/AnyMo","label":"Code","ariaLabel":"Code for AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild (opens in new tab)"}]},{"slug":"anyppg-ecg-guided-ppg-fm-2026","title":"AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling","summary":"AnyPPG pretrains a PPG encoder (Net1D/ResNet-style, 512-dim) using a CLIP-style cross-modal contrastive objective against a frozen ECG foundation model (ECGFounder) instead of PPG-only self-supervised learning. Pretraining data is 109,909 hours / 58,796 subjects of synchronized PPG+ECG from five clinical sources (MC-MED, PulseDB [MIMIC+VitalDB derived], MESA, HSP, CFS) — largely ICU and sleep-lab, not wearable. At inference only PPG is needed (ECG branch is pretraining scaffolding).","venue":"KDD 2026 (ACM SIGKDD V.2)","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/anyppg-ecg-guided-ppg-fm-2026","links":[{"href":"https://arxiv.org/abs/2511.01747","label":"Paper","ariaLabel":"Paper for AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling (opens in new tab)"},{"href":"https://doi.org/10.1145/3770855.3818911","label":"DOI","ariaLabel":"DOI for AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling (opens in new tab)"},{"href":"https://github.com/PKUDigitalHealth/AnyPPG","label":"Code","ariaLabel":"Code for AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling (opens in new tab)"}]},{"slug":"apnea-burden-ppg-ood-generalization-2026","title":"Apnea Burden-Guided Framework: Enhancing Out-of-Distribution Generalization in PPG-Based Sleep Apnea Characterization","summary":"This study predicts apnea burden from PPG as a continuous value, then converts it to the apnea-hypopnea index to sort people into four severity groups. The authors tested whether PPG pulse-shape features help alongside SpO2, especially on out-of-distribution data. On that data, adding PPG features raised macro-F1 by about 16% and macro-accuracy by about 9% compared with SpO2 alone.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/apnea-burden-ppg-ood-generalization-2026","links":[{"href":"https://arxiv.org/abs/2608.12229","label":"Paper","ariaLabel":"Paper for Apnea Burden-Guided Framework: Enhancing Out-of-Distribution Generalization in PPG-Based Sleep Apnea Characterization (opens in new tab)"}]},{"slug":"role-aware-physiological-sensing-ecg-gsr-group-interactions-iswc-2026","title":"Are You Performing, Anticipating, or Observing? Role-Aware Physiological Sensing in Group Interactions","summary":"This ISWC 2026 study looks at whether people in the same group activity have different physiological responses depending on their role. ECG and GSR were recorded from 100 participants in groups of three during sequential speaking tasks, and each moment was labelled as performing, anticipating, observing or baseline. Under leave-one-group-out testing the roles could be told apart, and the best deep learning model reached 0.66 accuracy and 0.87 AUROC.","venue":"ISWC '26 (ACM International Symposium on Wearable Computers), published 2026-10-10","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/role-aware-physiological-sensing-ecg-gsr-group-interactions-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834816","label":"Paper","ariaLabel":"Paper for Are You Performing, Anticipating, or Observing? Role-Aware Physiological Sensing in Group Interactions (opens in new tab)"}]},{"slug":"tcn-bilstm-ppg-to-abp-vitaldb-ahjo-2026","title":"Arterial blood pressure waveform reconstruction estimation from PPG using TCN–BiLSTM","summary":"TCN + BiLSTM on derivative-enhanced PPG reconstructs ABP. VitalDB, 2,226 patients: MAE/RMSE SBP 3.61/5.24, DBP 1.79/2.37 mmHg. Tested on a MIMIC-IV cohort without retraining (external numbers not in abstract).","venue":"American Heart Journal Plus: Cardiology Research and Practice","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/tcn-bilstm-ppg-to-abp-vitaldb-ahjo-2026","links":[{"href":"https://doi.org/10.1016/j.ahjo.2026.100837","label":"Paper","ariaLabel":"Paper for Arterial blood pressure waveform reconstruction estimation from PPG using TCN–BiLSTM (opens in new tab)"}]},{"slug":"chen-rar-gmm-patterns-mortality-nhanes-exp-gerontol-2026","title":"Association of accelerometer-measured rest-activity rhythm patterns and parameters with cardiovascular and all-cause mortality in older adults","summary":"NHANES 2011–2014 wrist accelerometry, 1,710 adults ≥60 y, median 6.67 y follow-up. GMM found 4 RAR patterns (morning-, earlier-, delayed-, evening-type); evening-type vs morning-type CVD mortality HR 12.16 (4.02–36.82), all-cause 3.31 (1.91–5.72); higher IS and RA protective.","venue":"Experimental Gerontology 2026","year":2026,"open_code":null,"updated":"2026-10-15","url":"/research/papers/chen-rar-gmm-patterns-mortality-nhanes-exp-gerontol-2026","links":[{"href":"https://doi.org/10.1016/j.exger.2026.113043","label":"Paper","ariaLabel":"Paper for Association of accelerometer-measured rest-activity rhythm patterns and parameters with cardiovascular and all-cause mortality in older adults (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/41581839","label":"PubMed","ariaLabel":"PubMed for Association of accelerometer-measured rest-activity rhythm patterns and parameters with cardiovascular and all-cause mortality in older adults (opens in new tab)"}]},{"slug":"autoresearchar-multi-agent-imu-har-iswc-2026","title":"AutoResearcHAR: A Multi-Agent Framework for Dataset Adaptive IMU Based Human Activity Recognition","summary":"Treats IMU-HAR as a whole-pipeline optimisation problem run by LLM agents: a Data Agent (source selection, preprocessing, windowing, sampling), a Model Agent (architecture + training choices) and an Augmentation Agent (IMU augmentation and synthetic-data strategies), composed by an Orchestrator, with an LLM-based Judge that filters candidate edits before they are run. Search is iterative hill-climbing over bounded code/config edits, keeping candidates that raise Macro-F1. Evaluated on 10 IMU-HAR datasets (LOSO) and 3 held-out datasets (leave-one-dataset-out): +1.2 to +12.9 Macro-F1 over the best non-agent baselines in LOSO and +2.3 to +8.9 in LODO.","venue":"ISWC '26 (Proceedings of the 2026 ACM International Symposium on Wearable Computers, UbiComp/ISWC 2026, Shanghai), pp. 149–155; published 2026-10-10","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/autoresearchar-multi-agent-imu-har-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834823","label":"DOI","ariaLabel":"DOI for AutoResearcHAR: A Multi-Agent Framework for Dataset Adaptive IMU Based Human Activity Recognition (opens in new tab)"}]},{"slug":"bayahar-bayesian-few-shot-har-personalization-iswc-2026","title":"BayaHAR: Lightweight Bayesian Few-Shot User Adaptation for On-Device Personalized Human Activity Recognition","summary":"BayaHAR personalizes activity recognition on the device without gradient training. It turns a pretrained classifier into a prototype network and updates the prototypes with closed-form Bayesian estimates, using labelled or only weakly labelled calibration data. With 3 seconds of calibration per activity, it improved macro-F1 on new users by 2.76 to 33.44 points across four datasets.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/bayahar-bayesian-few-shot-har-personalization-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834833","label":"Paper","ariaLabel":"Paper for BayaHAR: Lightweight Bayesian Few-Shot User Adaptation for On-Device Personalized Human Activity Recognition (opens in new tab)"},{"href":"https://arxiv.org/abs/2606.04798","label":"arXiv","ariaLabel":"arXiv for BayaHAR: Lightweight Bayesian Few-Shot User Adaptation for On-Device Personalized Human Activity Recognition (opens in new tab)"}]},{"slug":"beat-synchronous-ecg-tokenization-2026","title":"Beat-Synchronous Tokenization for ECG Transformers","summary":"University of Minnesota group asks a narrow, well-controlled question: does tokenizing an ECG waveform at cardiac-cycle (beat) boundaries instead of fixed time patches change Transformer performance, holding everything else (encoder architecture, SSL objective, evaluation protocol) constant? They compare fixed patches against three beat-synchronous tokenizer designs — resampled beats, adaptive-pooled beats, and resampled beats plus explicit R-R interval timing — on two settings: 12-lead diagnostic classification (PTB-XL after MIMIC-IV-ECG masked pretraining) and single-lead rhythm classification (Icentia11k after patient-level contrastive pretraining). No wrist signal, no PPG, no accelerometer — this is a pure ECG tokenization ablation.","venue":"IEEE MLSP 2026 (accepted)","year":2026,"open_code":null,"updated":null,"url":"/research/papers/beat-synchronous-ecg-tokenization-2026","links":[{"href":"https://arxiv.org/abs/2608.30367","label":"Paper","ariaLabel":"Paper for Beat-Synchronous Tokenization for ECG Transformers (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2608.30367","label":"DOI","ariaLabel":"DOI for Beat-Synchronous Tokenization for ECG Transformers (opens in new tab)"},{"href":"https://github.com/muha-0/beat-synchronous-tokenizer","label":"Code","ariaLabel":"Code for Beat-Synchronous Tokenization for ECG Transformers (opens in new tab)"}]},{"slug":"benchecg-xecg-xlstm-ecg-fm-2026","title":"BenchECG and xECG: a benchmark and baseline for ECG foundation models (xLSTM)","summary":"BenchECG is an open, standardised benchmark for ECG foundation models built from eight public datasets covering 421,171 patients and 1,674,704 recordings. It tests ten kinds of task, including classification, detection, regression, survival analysis and segmentation, with explicit out-of-distribution tests. The paper also introduces xECG, an xLSTM-based baseline model.","venue":"npj Digital Medicine","year":2026,"open_code":true,"updated":"2026-10-16","url":"/research/papers/benchecg-xecg-xlstm-ecg-fm-2026","links":[{"href":"https://doi.org/10.1038/s41746-026-03196-y","label":"Paper","ariaLabel":"Paper for BenchECG and xECG: a benchmark and baseline for ECG foundation models (xLSTM) (opens in new tab)"},{"href":"https://arxiv.org/abs/2509.10151","label":"arXiv","ariaLabel":"arXiv for BenchECG and xECG: a benchmark and baseline for ECG foundation models (xLSTM) (opens in new tab)"}]},{"slug":"motion-representations-fall-detection-real-world-2026","title":"Beyond Simulated Benchmarks: Evaluating Motion Representations for Fall Detection Under Real-World Data Scarcity","summary":"Real falls are rare: collecting 100 of them takes about 100,000 days of monitoring. This paper compares interval-based, kernel-based, symbolic and foundation model representations of accelerometer data for fall detection when labelled real-world falls are scarce.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/motion-representations-fall-detection-real-world-2026","links":[{"href":"https://arxiv.org/abs/2608.13197","label":"Paper","ariaLabel":"Paper for Beyond Simulated Benchmarks: Evaluating Motion Representations for Fall Detection Under Real-World Data Scarcity (opens in new tab)"}]},{"slug":"bio-pm-movement-segment-tokenization-2026","title":"Bio-PM: Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data","summary":"UMass Amherst / Google Health / UMD team argues that wearable accelerometer SSL has been leaving performance on the table by treating motion as an unstructured time series and windowing it arbitrarily. They ground a new tokenization in the submovement theory of motor control: continuous wrist motion decomposes into bell-shaped kinematic primitives (\"submovements\"), and the practically extractable unit — the \"movement segment\" bounded by acceleration zero-crossings — functions like a word made of phoneme-like submovements. They pretrain a Transformer via masked reconstruction over these movement-segment tokens (not raw time windows) on NHANES (~28k hours, ~11k participants, public US accelerometer corpus) and call the result Bio-PM.","venue":"arXiv preprint (ICML tag in keywords; venue not otherwise stated)","year":2026,"open_code":null,"updated":null,"url":"/research/papers/bio-pm-movement-segment-tokenization-2026","links":[{"href":"https://arxiv.org/abs/2603.10961","label":"Paper","ariaLabel":"Paper for Bio-PM: Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2603.10961","label":"DOI","ariaLabel":"DOI for Bio-PM: Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data (opens in new tab)"},{"href":"https://prithvitarale.github.io/biopm-site/","label":"Code","ariaLabel":"Code for Bio-PM: Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data (opens in new tab)"}]},{"slug":"m2ae-biosignal-fingerprinting-ppg-ecg-2026","title":"Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model","summary":"This paper introduces biosignal fingerprints, compact embeddings from M2AE, a multimodal masked autoencoder trained on more than 3.4 million paired ECG and PPG signals. Used frozen across 7 tasks, they matched or beat specialist foundation models, with AUROC of 0.974 for five-class cardiovascular disease and 0.877 for hypertension.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-28","url":"/research/papers/m2ae-biosignal-fingerprinting-ppg-ecg-2026","links":[{"href":"https://arxiv.org/abs/2605.09579","label":"Paper","ariaLabel":"Paper for Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2605.09579","label":"DOI","ariaLabel":"DOI for Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model (opens in new tab)"}]},{"slug":"ppg-vs-ecg-mediated-bp-pipelines-2026","title":"Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines","summary":"This study compares estimating blood pressure directly from PPG with pipelines that first reconstruct ECG from PPG. In MIMIC-III, PPG was more strongly linked to arterial pressure (|r| = 0.247) than ECG was (r = 0.018). Across 1.74M segments from 3,127 patients, direct PPG-to-BP prediction met British Hypertension Society grade A, with MAE of 4.82 mmHg systolic and 4.31 mmHg diastolic, beating every ECG-mediated pipeline.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ppg-vs-ecg-mediated-bp-pipelines-2026","links":[{"href":"https://arxiv.org/abs/2607.23406","label":"Paper","ariaLabel":"Paper for Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines (opens in new tab)"}]},{"slug":"ecg-ppg-jepa-latent-space-prediction-kdd-2026","title":"Bridging ECG and PPG: Latent-Space Prediction for Robust Physiological Analysis","summary":"Zhaoliang Chen, Saurabh Kataria, Minxiao Wang, Xiao Hu. Nell Hodgson Woodruff School of Nursing, Emory University. KDD '26, DOI 10.1145/3770855.3819000, open access CC BY 4.0.","venue":"KDD '26 (Proc. 32nd ACM SIGKDD, V.2), August 2026","year":2026,"open_code":false,"updated":"2026-10-12","url":"/research/papers/ecg-ppg-jepa-latent-space-prediction-kdd-2026","links":[{"href":"https://doi.org/10.1145/3770855.3819000","label":"Paper","ariaLabel":"Paper for Bridging ECG and PPG: Latent-Space Prediction for Robust Physiological Analysis (opens in new tab)"}]},{"slug":"calibrated-ppg-hrv-ecg-stress-aud-relapse-2026","title":"Calibrated photoplethysmography-derived heart rate variability approximates electrocardiographic gold-standard performance for stress-state classification","summary":"This study tested whether heart rate variability from PPG can stand in for ECG when classifying stress, as a proxy for relapse risk in alcohol use disorder. ECG-derived features gave 80% accuracy and an ROC AUC of 0.85. A simple linear calibration that maps PPG features to ECG raised the PPG AUC from 0.79 to 0.85, matching ECG.","venue":"Springer Science and Business Media LLC","year":2026,"open_code":false,"updated":"2026-10-05","url":"/research/papers/calibrated-ppg-hrv-ecg-stress-aud-relapse-2026","links":[{"href":"https://doi.org/10.21203/rs.3.rs-11204325/v1","label":"DOI","ariaLabel":"DOI for Calibrated photoplethysmography-derived heart rate variability approximates electrocardiographic gold-standard performance for stress-state classification (opens in new tab)"}]},{"slug":"dilated-residual-attention-unet-calibrated-bp-information-2026","title":"Calibration-Based Cuffless Blood Pressure Estimation Using a Dilated-Residual Attention U-Net","summary":"Dual-stream 1D encoder-decoder on raw PPG + ECG: dilated residual convs, FiLM cross-signal modulation at the bottleneck, learned scale-wise fusion, attention-gated skips, mean-attention pooling head. PulseDB calibration-based subset (SBP 57–180, DBP 25–100 mmHg). MAE SBP 3.81 (ME −0.28 ± 5.54), DBP 2.03 (ME 0.38 ± 3.13), R² 0.92/0.93; BHS and IEEE-1708 grade A.","venue":"Information (MDPI) 17(9):878","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/dilated-residual-attention-unet-calibrated-bp-information-2026","links":[{"href":"https://doi.org/10.3390/info17090878","label":"Paper","ariaLabel":"Paper for Calibration-Based Cuffless Blood Pressure Estimation Using a Dilated-Residual Attention U-Net (opens in new tab)"}]},{"slug":"cap-ppg-clinical-anchored-pretraining-2026","title":"CAP: Towards PPG Universal Representation Learning with Patient-level Supervision","summary":"CAP tables compare mainly to SiamQuality and GPT-PPG (plus PaPaGei-P on HR row without LOSO MAE). CAP wins on reported AF/HR/RR/BP numbers but is clinical-EHR-anchored closed-weight; Pulse-PPG/PaPaGei remain open wrist baselines.","venue":"KDD 2026","year":2026,"open_code":null,"updated":"2026-09-28","url":"/research/papers/cap-ppg-clinical-anchored-pretraining-2026","links":[{"href":"https://arxiv.org/abs/2606.15284","label":"Paper","ariaLabel":"Paper for CAP: Towards PPG Universal Representation Learning with Patient-level Supervision (opens in new tab)"},{"href":"https://doi.org/10.1145/3770855.3818881","label":"DOI","ariaLabel":"DOI for CAP: Towards PPG Universal Representation Learning with Patient-level Supervision (opens in new tab)"},{"href":"https://github.com/gody123gody/CAP","label":"Code","ariaLabel":"Code for CAP: Towards PPG Universal Representation Learning with Patient-level Supervision (opens in new tab)"}]},{"slug":"capfm-capacitive-touch-fm-uist-2026","title":"CapFM: A Foundation Model for Capacitive Touch Sensing","summary":null,"venue":"ACM UIST 2026 (to appear, November)","year":2026,"open_code":false,"updated":"2026-10-02","url":"/research/papers/capfm-capacitive-touch-fm-uist-2026","links":[{"href":"https://siplab.org/","label":"Paper","ariaLabel":"Paper for CapFM: A Foundation Model for Capacitive Touch Sensing (opens in new tab)"},{"href":"https://github.com/eth-siplab/CapFM","label":"Code","ariaLabel":"Code for CapFM: A Foundation Model for Capacitive Touch Sensing (opens in new tab)"}]},{"slug":"capwear-footwear-capillary-refill-time-iswc-2026","title":"CapWear: Smart Footwear for Peripheral Perfusion Monitoring by Continuous Plantar Capillary Refill Time Estimation","summary":"A footwear system that estimates capillary refill time (CRT) continuously during everyday walking. The heel pressure of each step restricts blood flow and its release triggers a refill transient. The abstract demonstrates feasibility; the sensing modality and quantitative agreement are not stated in the abstract (ACM DL full text blocked, 403).","venue":"ISWC '26 (ACM International Symposium on Wearable Computers), published 2026-10-10","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/capwear-footwear-capillary-refill-time-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834814","label":"Paper","ariaLabel":"Paper for CapWear: Smart Footwear for Peripheral Perfusion Monitoring by Continuous Plantar Capillary Refill Time Estimation (opens in new tab)"}]},{"slug":"cardionets-ecg-cmr-2026","title":"CardioNets: ECG-CMR Cross-Modal Foundation Model","summary":"CardioNets translates 12-lead ECG into cardiac MRI-level measurements and synthetic MRI images. It combines cross-modal contrastive learning with a masked autoregressive image generator and was trained on 159,819 samples from five cohorts, including UK Biobank, with external validation on 3,767 cases. In UK Biobank it improved cardiac phenotype R² by 24.8% and cardiomyopathy AUC by up to 39.3% over baselines, and in a reader study it was 13.9% more accurate than physicians using ECG plus real MRI.","venue":"NEJM AI","year":2026,"open_code":null,"updated":null,"url":"/research/papers/cardionets-ecg-cmr-2026","links":[{"href":"https://doi.org/10.1056/AIoa2500549","label":"DOI","ariaLabel":"DOI for CardioNets: ECG-CMR Cross-Modal Foundation Model (opens in new tab)"}]},{"slug":"cardiostate-jepa-cross-modal-cardiac-fm-2026","title":"CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation","summary":"CardioState-JEPA is a cardiac foundation model that learns one shared representation across ECG, PPG and heart sounds. Signals are mapped into a common token space and processed by a single Transformer that learns by predicting masked latent cardiac states. As a frozen encoder across 25 tasks, it improved PPG classification by 8.2 AUROC points, murmur detection by 18.8 and ECG classification by 15.5 over the best baselines.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-27","url":"/research/papers/cardiostate-jepa-cross-modal-cardiac-fm-2026","links":[{"href":"https://arxiv.org/abs/2608.12944v2","label":"Paper","ariaLabel":"Paper for CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation (opens in new tab)"},{"href":"https://arxiv.org/abs/2608.12944","label":"arXiv","ariaLabel":"arXiv for CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation (opens in new tab)"}]},{"slug":"har-performance-gap-older-adults-iswc-2026","title":"Characterizing the Performance Gap in Human Activity Recognition for Older Adults","summary":"This study asks whether activity recognition progress on young-adult benchmarks carries over to older adults. Using MyMove, an annotated free-living dataset of older adults (mean age 71), the authors found that gains on younger-adult data did not transfer equally, leaving a persistent and often widening gap.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-05","url":"/research/papers/har-performance-gap-older-adults-iswc-2026","links":[{"href":"https://arxiv.org/abs/2610.02711","label":"Paper","ariaLabel":"Paper for Characterizing the Performance Gap in Human Activity Recognition for Older Adults (opens in new tab)"},{"href":"https://doi.org/10.1145/3830727.3834835","label":"DOI","ariaLabel":"DOI for Characterizing the Performance Gap in Human Activity Recognition for Older Adults (opens in new tab)"}]},{"slug":"chear-confusion-hierarchy-evidence-aware-llm-har-2026","title":"CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition","summary":"Training-free LLM HAR in the ZARA line. CHEAR builds a confusion hierarchy over activity labels from seen-subject data, retrieves only the small candidate set relevant to a query window and attaches local signal evidence, so the LLM decides among a few confusable classes instead of all labels.","venue":"arXiv 2610.15298 (v1 2026-10-11, cs.AI)","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/chear-confusion-hierarchy-evidence-aware-llm-har-2026","links":[{"href":"https://arxiv.org/abs/2610.15298","label":"Paper","ariaLabel":"Paper for CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition (opens in new tab)"}]},{"slug":"citrus-ring-hr-selective-reporting-2026","title":"CITRUS: Candidate Inference and Temporal-tracking for Reliable, Unobtrusive Sensing of Wearable Heart Rate under Motion","summary":"A thesis (not a peer-reviewed venue paper) building a two-layer wearable HR system for a custom finger-ring PPG platform (τ-Ring, 100 Hz IR/red + accelerometer): an estimation layer that generates ~140 candidate HR values per window from multiple DSP estimators/channels/harmonics, scores them with a small MLP, and decodes a temporally-consistent path with causal Viterbi (HMM-style dynamic programming, past-only); and a reporting layer that estimates reliability and applies a learned 3-way policy — accept the new estimate, hold the last trusted value, or reject the window outright. Directly and independently replicates the \"selective reporting reduces motion-window MAE substantially\" finding on an external wrist-BVP cohort: PPG-DaLiA, with proper subject-disjoint LOSO, no ring-trained weights carried over.","venue":"MSc/PhD thesis, Tsinghua University (cs.HC preprint)","year":2026,"open_code":null,"updated":null,"url":"/research/papers/citrus-ring-hr-selective-reporting-2026","links":[{"href":"https://arxiv.org/abs/2607.18818","label":"Paper","ariaLabel":"Paper for CITRUS: Candidate Inference and Temporal-tracking for Reliable, Unobtrusive Sensing of Wearable Heart Rate under Motion (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2607.18818","label":"DOI","ariaLabel":"DOI for CITRUS: Candidate Inference and Temporal-tracking for Reliable, Unobtrusive Sensing of Wearable Heart Rate under Motion (opens in new tab)"}]},{"slug":"cogadapt-ecg-fm-wearable-cognitive-load-2026","title":"CogAdapt: Adapting Clinical ECG Foundation Models for Wearable Cognitive Load Assessment","summary":"This paper presents CogAdapt, which adapts a clinical ECG foundation model to estimate cognitive load from wearable ECG. Its ProFine strategy unfreezes encoder layers in stages to limit drift. On CLARE and CL-Drive with leave-one-subject-out testing it reached macro-F1 of 0.626 and 0.768, 11.2 and 16.1 points above models trained from scratch.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-26","url":"/research/papers/cogadapt-ecg-fm-wearable-cognitive-load-2026","links":[{"href":"https://arxiv.org/abs/2605.22774","label":"Paper","ariaLabel":"Paper for CogAdapt: Adapting Clinical ECG Foundation Models for Wearable Cognitive Load Assessment (opens in new tab)"}]},{"slug":"comodo-video-to-imu-distillation-egocentric-har-imwut-2026","title":"COMODO: Cross-Modal Video-to-IMU Distillation for Efficient Egocentric Human Activity Recognition","summary":"Teacher/student-agnostic (video: VideoMAE, TimeSformer; time-series: MOMENT, Mantis). Reported to improve downstream egocentric HAR across several datasets, matching or beating fully supervised models, with strong cross-dataset generalisation.","venue":"Proc. ACM IMWUT 10(2), published 2026-06-15 (CC BY 4.0); presented at UbiComp/ISWC 2026 (Shanghai, week of 2026-10-12)","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/comodo-video-to-imu-distillation-egocentric-har-imwut-2026","links":[{"href":"https://doi.org/10.1145/3810218","label":"DOI","ariaLabel":"DOI for COMODO: Cross-Modal Video-to-IMU Distillation for Efficient Egocentric Human Activity Recognition (opens in new tab)"},{"href":"https://arxiv.org/abs/2503.07259","label":"arXiv","ariaLabel":"arXiv for COMODO: Cross-Modal Video-to-IMU Distillation for Efficient Egocentric Human Activity Recognition (opens in new tab)"}]},{"slug":"clmt-compact-latent-manifold-translation-ppg-ecg-rvq-2026","title":"Compact Latent Manifold Translation: A Parameter-Efficient Foundation Model for Cross-Modal and Cross-Frequency Physiological Signal Synthesis","summary":"CLMT is a small (0.09B parameter) model that translates between physiological signals, such as PPG to ECG, and between sampling rates using a two-stage discrete latent approach. In PPG-to-ECG synthesis it raised R-peak detection F1 from 0.37 to 0.83, and upsampling from 25 Hz to 100 Hz reached a Pearson correlation of 0.9956.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-13","url":"/research/papers/clmt-compact-latent-manifold-translation-ppg-ecg-rvq-2026","links":[{"href":"https://arxiv.org/abs/2605.13248","label":"Paper","ariaLabel":"Paper for Compact Latent Manifold Translation: A Parameter-Efficient Foundation Model for Cross-Modal and Cross-Frequency Physiological Signal Synthesis (opens in new tab)"}]},{"slug":"compact-kd-ppg-bp-fpga-mcu-lsens-2026","title":"Compact PPG-Based Blood Pressure Estimation Using Hardware-Efficient Knowledge Distillation","summary":"Students trained on BP labels + teacher predictions + feature-level supervision from PPG-only or hybrid ECG–PPG teachers. PhysioNet PTT and VitalDB. A 1-depth, 16-channel residual U-Net student (20.7k parameters) keeps accuracy with minimal penalty; fits a VCU709 FPGA; nRF5340 MCU deployment is RAM-limited for full-window inference.","venue":"IEEE Sensors Letters","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/compact-kd-ppg-bp-fpga-mcu-lsens-2026","links":[{"href":"https://doi.org/10.1109/LSENS.2026.3720577","label":"Paper","ariaLabel":"Paper for Compact PPG-Based Blood Pressure Estimation Using Hardware-Efficient Knowledge Distillation (opens in new tab)"}]},{"slug":"blpm-eeg-language-continuous-latent-2026","title":"Continuous-Latent Predictive Modeling with Semantic Alignment for EEG-Language Foundation Models","summary":"BLPM is an EEG-language foundation model that treats different EEG decoding tasks as predicting a continuous semantic embedding. Its encoder learns by predicting latent targets, and it generalized consistently across several benchmarks.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/blpm-eeg-language-continuous-latent-2026","links":[{"href":"https://arxiv.org/abs/2608.11656","label":"Paper","ariaLabel":"Paper for Continuous-Latent Predictive Modeling with Semantic Alignment for EEG-Language Foundation Models (opens in new tab)"}]},{"slug":"coverage-aware-virtual-imu-augmentation-2026","title":"Coverage-Aware Virtual IMU Augmentation for Low-Resource Human Activity Recognition","summary":"This paper adds virtual IMU data to activity recognition training only where it fills gaps in the real data. The method picks where to add samples, generates candidates, ranks them by closeness to real anchors and label consistency, and weights them by reliability during training.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-29","url":"/research/papers/coverage-aware-virtual-imu-augmentation-2026","links":[{"href":"https://arxiv.org/abs/2609.16768","label":"Paper","ariaLabel":"Paper for Coverage-Aware Virtual IMU Augmentation for Low-Resource Human Activity Recognition (opens in new tab)"}]},{"slug":"cowear-collaborative-wearable-inertial-localization-2026","title":"CoWear: Towards Collaborative Wearable Inertial Localization with Multiple Smart Devices","summary":"Dataset + reference framework for collaborative inertial localization with synchronized smartphone, smartwatch, and smart glasses. >30 participants, >300 walking sessions, >30 h and >100 km aggregate trajectories; each device tracked by motion capture. Role-specific estimators predict displacement of a common reference device over synced windows; gyroscope-integrated world-frame fusion with learned uncertainty.","venue":"arXiv 2610.20596 (submitted 2026-10-14)","year":2026,"open_code":null,"updated":"2026-10-15","url":"/research/papers/cowear-collaborative-wearable-inertial-localization-2026","links":[{"href":"https://arxiv.org/abs/2610.20596","label":"arXiv","ariaLabel":"arXiv for CoWear: Towards Collaborative Wearable Inertial Localization with Multiple Smart Devices (opens in new tab)"}]},{"slug":"cross-modal-ecg-ppg-stress-wesad-2026","title":"Cross-Modal Correlation Fusion for End-to-End ECG–PPG Stress Classification","summary":"Dual parallel encoders extract morphological features from ECG and PPG; a cross-modal correlation fusion layer models electro-mechanical alignment (cardiac electrical triggers ↔ peripheral BVP). Evaluated on WESAD with subject-independent 5-fold CV: mean Accuracy 90.34%, AUC-ROC 97.29%, Stress Recall 94.07%.","venue":"Frontiers in Artificial Intelligence and Applications (book chapter)","year":2026,"open_code":null,"updated":"2026-10-02","url":"/research/papers/cross-modal-ecg-ppg-stress-wesad-2026","links":[{"href":"https://doi.org/10.3233/faia261041","label":"Paper","ariaLabel":"Paper for Cross-Modal Correlation Fusion for End-to-End ECG–PPG Stress Classification (opens in new tab)"}]},{"slug":"cm-ssl-cross-modal-token-vocabulary-ecg-ppg-icassp-2026","title":"Cross-Modality Self-Supervision for Unimodal Physiological Signals Representation Learning","summary":"Xiaomao Fan, Shuangdu Li, Jikui Liu, Jinzhou Cao, Jingyan Jiang, Yumeng Liu (+ others; Shenzhen Polytechnic University, Shenzhen Technology University). DOI 10.1109/ICASSP55912.2026.11461644. Code: github.com/ximosss/CM-SSL (no licence file; checkpoints/ is empty).","venue":"ICASSP 2026","year":2026,"open_code":true,"updated":"2026-10-11","url":"/research/papers/cm-ssl-cross-modal-token-vocabulary-ecg-ppg-icassp-2026","links":[{"href":"https://doi.org/10.1109/ICASSP55912.2026.11461644","label":"Paper","ariaLabel":"Paper for Cross-Modality Self-Supervision for Unimodal Physiological Signals Representation Learning (opens in new tab)"}]},{"slug":"csfm-cardiac-signal-fm-2026","title":"CSFM: Cardiac Signal Foundation Model","summary":"CSFM is a cardiac sensing foundation model that uses transformers and generative masked pretraining to learn from mixed ECG and PPG health records. It beat single-modality, single-task approaches on diagnosis, demographics, vital-sign measurement, clinical outcome prediction and ECG question answering.","venue":"Nature Machine Intelligence","year":2026,"open_code":null,"updated":null,"url":"/research/papers/csfm-cardiac-signal-fm-2026","links":[{"href":"https://doi.org/10.1038/s42256-026-01180-5","label":"DOI","ariaLabel":"DOI for CSFM: Cardiac Signal Foundation Model (opens in new tab)"}]},{"slug":"windkessel-regularized-ppg-bp-lsens-2026","title":"Cuffless Blood Pressure Estimation From Wearable Photoplethysmography and Windkessel Model-Regularized Learning Approach","summary":"Adds a bounded Windkessel loss: arterial pressure must decay exponentially with a PPG-derived diastolic time constant. 10 healthy subjects, wearable finger PPG, rest + affective stimulation + Stroop. MAE SBP 4.10 ± 1.41, DBP 2.90 ± 1.20 mmHg; authors claim AAMI compliance (n = 10 can't meet AAMI's 85-subject requirement).","venue":"IEEE Sensors Letters","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/windkessel-regularized-ppg-bp-lsens-2026","links":[{"href":"https://doi.org/10.1109/LSENS.2026.3724173","label":"Paper","ariaLabel":"Paper for Cuffless Blood Pressure Estimation From Wearable Photoplethysmography and Windkessel Model-Regularized Learning Approach (opens in new tab)"}]},{"slug":"eit-beat-beat-ppg-bp-ppgbp-features-2026","title":"Cuffless Photoplethysmogram Based Beat-Beat AI Powered Continuous Blood Pressure Estimation Towards Precision Medicine","summary":"PPG-BP (Guilin, 219 subjects): top-5 features from each of five domains (time, frequency, entropy, recurrence, statistical) via SelectKBest; Huber / TheilSen / SVR / RANSAC / linear regressors. Reported SBP MAE 0.208 (R² 0.996), DBP MAE 0.173 (R² 0.996).","venue":"IEEE eIT 2026","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/eit-beat-beat-ppg-bp-ppgbp-features-2026","links":[{"href":"https://doi.org/10.1109/eIT68936.2026.11670374","label":"Paper","ariaLabel":"Paper for Cuffless Photoplethysmogram Based Beat-Beat AI Powered Continuous Blood Pressure Estimation Towards Precision Medicine (opens in new tab)"}]},{"slug":"few-shot-fm-embeddings-cuffless-bp-aha-bsn-2026","title":"Data Efficient Cuffless Blood Pressure Classification from PPG using Foundation Model Embeddings and Few-Shot Learning","summary":"A pretrained wearable-physiology FM (not named in the abstract) is used as a frozen feature extractor; a lightweight few-shot classifier conditioned on 3 labeled PPG–SBP context pairs predicts the AHA BP category (6 classes) of a query segment. Adaptation pool 2,400 pairs; held-out test 15,000 pairs from PulseDB/VitalDB. Accuracy 82.4%, weighted F1 0.832 from PPG alone.","venue":"IEEE BSN 2026 (OpenReview)","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/few-shot-fm-embeddings-cuffless-bp-aha-bsn-2026","links":[{"href":"https://openreview.net/forum?id=yvYELPabvv","label":"Paper","ariaLabel":"Paper for Data Efficient Cuffless Blood Pressure Classification from PPG using Foundation Model Embeddings and Few-Shot Learning (opens in new tab)"}]},{"slug":"deepconvcontext-multi-scale-har-iswc-2026","title":"DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition","summary":"DeepConvContext is a multi-scale activity recognition model that learns across neighbouring windows instead of classifying each window alone. On six common benchmarks it improved F1 by 5% on average and mAP by up to 18 points, with latency and throughput similar to earlier methods. Learning across windows also gave more coherent activity segments in online prediction.","venue":"arXiv preprint","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/deepconvcontext-multi-scale-har-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834810","label":"Paper","ariaLabel":"Paper for DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition (opens in new tab)"},{"href":"https://arxiv.org/abs/2505.20894","label":"arXiv","ariaLabel":"arXiv for DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition (opens in new tab)"},{"href":"https://github.com/mariusbock/deepconvcontext","label":"Code","ariaLabel":"Code for DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition (opens in new tab)"}]},{"slug":"george-pulsewatch-smartwatch-ppg-acc-arrhythmia-rf-vs-cnn-bilstm-frai-2026","title":"Development of an efficient deep learning model for arrhythmia detection from consumer smartwatches","summary":"Compares a feature-based Random Forest with two end-to-end models (CNN-GRU, CNN-BiLSTM) for 3-class rhythm (NSR, AF, PAC/PVC) on 30 s windows of real-life smartwatch data from the Pulsewatch trial. Input is a 2-channel series: wrist PPG at 50 Hz plus a scalar accelerometer magnitude (the release ships ACC as one value per sample). Devices per the paper: Samsung Gear S3 or Galaxy Watch 3 with a chest ECG patch (250 Hz) as reference.","venue":"Frontiers in Artificial Intelligence 9:1882739 (2026-09-28)","year":2026,"open_code":false,"updated":"2026-10-13","url":"/research/papers/george-pulsewatch-smartwatch-ppg-acc-arrhythmia-rf-vs-cnn-bilstm-frai-2026","links":[{"href":"https://doi.org/10.3389/frai.2026.1882739","label":"DOI","ariaLabel":"DOI for Development of an efficient deep learning model for arrhythmia detection from consumer smartwatches (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42865579","label":"PubMed","ariaLabel":"PubMed for Development of an efficient deep learning model for arrhythmia detection from consumer smartwatches (opens in new tab)"}]},{"slug":"luminaecg-doctor-grounded-ecg-agent-2026","title":"Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations","summary":"LuminaECG treats ECG interpretation the way cardiologists read it: as measurement-based visual reading. A general 2B-parameter vision-language model is fine-tuned with low-rank adapters to link measurement primitives to diagnostic reasoning, without changing its architecture.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/luminaecg-doctor-grounded-ecg-agent-2026","links":[{"href":"https://arxiv.org/abs/2608.09053","label":"Paper","ariaLabel":"Paper for Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations (opens in new tab)"}]},{"slug":"diffusion-synthetic-pretraining-har-2026","title":"Diffusion-Based Synthetic Data Pretraining for Enhancing Activity Recognition","summary":"This paper improves detection of eating and drinking from wearable motion data by pretraining a CaBiGRU model on synthetic sensor windows made with a diffusion model, then fine-tuning on real data. On the DEO dataset it reached a balanced accuracy of 90.6%, beating a strong supervised baseline and helping with the rare class.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-05","url":"/research/papers/diffusion-synthetic-pretraining-har-2026","links":[{"href":"https://arxiv.org/abs/2610.02292","label":"Paper","ariaLabel":"Paper for Diffusion-Based Synthetic Data Pretraining for Enhancing Activity Recognition (opens in new tab)"}]},{"slug":"soley-wearable-digital-phenotypes-hypertensive-disorders-pregnancy-2026","title":"Discovering wearable sensor-derived temporal digital phenotypes and their association with hypertensive disorders in pregnancy","summary":"Retrospective cohort: pregnant participants in the NIH All of Us Research Program with Fitbit and EHR data. 14-day wearable windows before the end of each trimester and postpartum. An attention-based recurrent autoencoder per modality (activity, sleep, pulse rate variability), with GMM clustering into digital phenotypes.","venue":"Research Square preprint (posted 2026-10-14; CC BY)","year":2026,"open_code":null,"updated":"2026-10-15","url":"/research/papers/soley-wearable-digital-phenotypes-hypertensive-disorders-pregnancy-2026","links":[{"href":"https://doi.org/10.21203/rs.3.rs-9452274/v1","label":"DOI","ariaLabel":"DOI for Discovering wearable sensor-derived temporal digital phenotypes and their association with hypertensive disorders in pregnancy (opens in new tab)"}]},{"slug":"dna-ppg-dual-neighborhood-alignment-2026","title":"DNA-PPG: A Foundation Model for Photoplethysmography via Dual Neighborhood Alignment","summary":"A direct attack on PaPaGei-S, from the loss function rather than the architecture. The claim is that the two standard ways of building a PPG contrastive objective are both wrong in a specific, fixable way. First, instance discrimination (Apple's 141k-participant model, SimCLR-style) declares every segment from a different subject a negative — which means two people with near-identical pulse morphology get actively pushed apart, distorting the manifold.","venue":"IJCAI-2026 (Thirty-Fifth IJCAI, Special Track on AI and Health), pp. 6984-6992","year":2026,"open_code":null,"updated":"2026-09-27","url":"/research/papers/dna-ppg-dual-neighborhood-alignment-2026","links":[{"href":"https://www.ijcai.org/proceedings/2026/777","label":"Paper","ariaLabel":"Paper for DNA-PPG: A Foundation Model for Photoplethysmography via Dual Neighborhood Alignment (opens in new tab)"},{"href":"https://doi.org/10.24963/ijcai.2026/777","label":"DOI","ariaLabel":"DOI for DNA-PPG: A Foundation Model for Photoplethysmography via Dual Neighborhood Alignment (opens in new tab)"},{"href":"https://github.com/yyzyyz001/DNA-PPG","label":"Code","ariaLabel":"Code for DNA-PPG: A Foundation Model for Photoplethysmography via Dual Neighborhood Alignment (opens in new tab)"}]},{"slug":"dual-accelerometry-fm-phenome-risk-2026","title":"Dual foundation models for accelerometry predict future health","summary":"This medRxiv study encoded one week of UK Biobank wrist accelerometry from 97,696 people with two frozen self-supervised models and trained one survival model for 390 health outcomes. In 5,253 held-out participants mean concordance was 0.688, and one shared component explained 76% of predicted risk variance. Movement data also flagged early signs of neurodegenerative disease, strongest for Parkinson's (five-year AUROC 0.90).","venue":"medRxiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/dual-accelerometry-fm-phenome-risk-2026","links":[{"href":"https://www.medrxiv.org/content/10.64898/2026.07.24.26358894v3","label":"Paper","ariaLabel":"Paper for Dual foundation models for accelerometry predict future health (opens in new tab)"}]},{"slug":"obermeyer-ecg-scd-generative-explainability-2026","title":"ECG Mortality/SCD Prediction with Generative Explainability","summary":"This Nature paper uses ECG to predict sudden cardiac death, where the standard marker, reduced ejection fraction, misses most cases. The model picked out a high-risk group (2.2% of patients) with a 7.0% yearly rate of sudden cardiac death, higher than people with reduced LVEF (4.6%). Most of the model's high-risk patients (86.1%) were not flagged by LVEF.","venue":"Nature","year":2026,"open_code":null,"updated":null,"url":"/research/papers/obermeyer-ecg-scd-generative-explainability-2026","links":[{"href":"https://doi.org/10.1038/s41586-026-10674-6","label":"DOI","ariaLabel":"DOI for ECG Mortality/SCD Prediction with Generative Explainability (opens in new tab)"}]},{"slug":"ecg-gpt-image-report-2026","title":"ECG-GPT: ECG Image-to-Report Foundation Model","summary":"ECG-GPT treats ECG as a visual image rather than raw voltage time-series and generates clinical text reports using a BEiT vision encoder + GPT-2 decoder (VLM architecture). Trained on ~2.9M development / ~4.1M validation ECGs from 7 sites. Reports accuracy 0.93–0.99 on benchmark report generation tasks.","venue":"European Heart Journal – Digital Health","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ecg-gpt-image-report-2026","links":[{"href":"https://academic.oup.com/ehjdh/article/7/3/ztag031/8489956","label":"Paper","ariaLabel":"Paper for ECG-GPT: ECG Image-to-Report Foundation Model (opens in new tab)"}]},{"slug":"ecg-interpbench-interpretability-2026","title":"ECG-InterpBench: Benchmarking Interpretability of ECG Foundation Models with Matched-Scale Sparse Autoencoders","summary":"ECG-InterpBench measures how interpretable the learned representations inside ECG foundation models are, alongside how well they predict. It uses sparse autoencoders of matched capacity as standard measuring tools. The authors tested six frozen ECG foundation models across five encoder depths, five dictionary widths and three seeds, producing a 450-cell interpretability atlas.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ecg-interpbench-interpretability-2026","links":[{"href":"https://arxiv.org/abs/2607.27404","label":"arXiv","ariaLabel":"arXiv for ECG-InterpBench: Benchmarking Interpretability of ECG Foundation Models with Matched-Scale Sparse Autoencoders (opens in new tab)"}]},{"slug":"ecg-rsi-recursive-self-improvement-waveform-evolution-2026","title":"ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution","summary":"ECG models need supervision that targets evolving weaknesses and stays faithful to waveform evidence. Across seven benchmarks and three rounds, Qwen3.5-35B-A3B mean question-answering accuracy rises from 34.57% → 56.04% and mean classification F1 from 24.27 → 34.03.","venue":"arXiv 2610.17718 (2026-10-13, cs.AI / cs.LG / eess.SP / q-bio.QM)","year":2026,"open_code":true,"updated":"2026-10-14","url":"/research/papers/ecg-rsi-recursive-self-improvement-waveform-evolution-2026","links":[{"href":"https://arxiv.org/abs/2610.17718","label":"Paper","ariaLabel":"Paper for ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution (opens in new tab)"}]},{"slug":"edgehar-edge-native-sensor-fm-har-2026","title":"EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition","summary":"Disentangles activity semantics / motion dynamics / acquisition context for edge-deployable HAR (~1M params). Strong few-shot transfer under user/device/placement shift; PAMAP2 emerging-activity Macro-F1 0.9186 at K=100.","venue":"UbiComp Companion 2026","year":2026,"open_code":null,"updated":"2026-09-29","url":"/research/papers/edgehar-edge-native-sensor-fm-har-2026","links":[{"href":"https://arxiv.org/abs/2609.14498","label":"Paper","ariaLabel":"Paper for EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition (opens in new tab)"}]},{"slug":"ppg-artifacts-uncertainty-af-detection-2026","title":"Effect of Photoplethysmogram Artifacts on Uncertainty in Atrial Fibrillation Detection","summary":"This study measured how different PPG artifacts affect the uncertainty of atrial fibrillation detectors. Two detectors, one on raw 25-second PPG and one on rhythm-irregularity features, were trained on wrist PPG from cardiac rehabilitation and tested on 260,000 signals with controlled artifacts. On clean signals the feature-based detector did better (0.94/0.94 sensitivity/specificity vs 0.92/0.86), and device displacement hurt the raw-signal detector most.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ppg-artifacts-uncertainty-af-detection-2026","links":[{"href":"https://arxiv.org/abs/2608.09204","label":"Paper","ariaLabel":"Paper for Effect of Photoplethysmogram Artifacts on Uncertainty in Atrial Fibrillation Detection (opens in new tab)"}]},{"slug":"end-to-end-nas-ppg-bp-2026","title":"End-to-end Automated Deep Neural Network Optimization for PPG-based Blood Pressure Estimation on Wearables","summary":"This work builds compact neural networks for estimating blood pressure from PPG on low-power wearable chips. It combines hardware-aware neural architecture search, pruning and mixed-precision search into one automated pipeline. Starting from top models on four public datasets, it produced networks with up to 7.99% lower error and 7.5 times fewer parameters, or 83 times fewer parameters with almost no loss, and patient-specific fine-tuning improved accuracy by up to 64%.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/end-to-end-nas-ppg-bp-2026","links":[{"href":"https://arxiv.org/abs/2604.10117","label":"Paper","ariaLabel":"Paper for End-to-end Automated Deep Neural Network Optimization for PPG-based Blood Pressure Estimation on Wearables (opens in new tab)"}]},{"slug":"text2imu-diversity-sources-ojemb-2026","title":"Evaluating Sources of Diversity in Text-Driven IMU Synthesis for Wearable Human Activity Recognition","summary":"Freiburg (Amft lab) systematically stress-tests Text2IMU: text → motion synthesis → body surface models → virtual wearable ACC/gyro, then trains HAR on synthetic-only data. They ablate diversity sources — number of textual prompts, surface-model morphologies, virtual sensor locations — and compare / combine motion-synthesis backbones. Headline: for correctly synthesised activities, best config reaches 89.2% balanced accuracy under synthetic-only training.","venue":"IEEE Open Journal of Engineering in Medicine and Biology","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/text2imu-diversity-sources-ojemb-2026","links":[{"href":"https://ieeexplore.ieee.org/document/11570132","label":"Paper","ariaLabel":"Paper for Evaluating Sources of Diversity in Text-Driven IMU Synthesis for Wearable Human Activity Recognition (opens in new tab)"},{"href":"https://doi.org/10.1109/OJEMB.2026.3705244","label":"DOI","ariaLabel":"DOI for Evaluating Sources of Diversity in Text-Driven IMU Synthesis for Wearable Human Activity Recognition (opens in new tab)"}]},{"slug":"expertorhythm-morphology-aware-ppg-abp-bp-2026","title":"ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG","summary":"Attention-enhanced (CPCA) 1D U-Net reconstructs full arterial blood pressure (ABP) waveform from single-channel PPG, then derives SBP/DBP from waveform extrema. Composite morphology-aware loss: range-weighted SmoothL1 + window-range amplitude regularizer (λAR=0.6 peak). Evaluated on UCI cuff-less BP dataset (MIMIC-derived; 942 subjects; 10 s @ 125 Hz).","venue":"arXiv preprint; Version of Record in LNAI 16748 (AIME 2026)","year":2026,"open_code":true,"updated":"2026-10-01","url":"/research/papers/expertorhythm-morphology-aware-ppg-abp-bp-2026","links":[{"href":"https://arxiv.org/abs/2609.34146","label":"Paper","ariaLabel":"Paper for ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2609.34146","label":"DOI","ariaLabel":"DOI for ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG (opens in new tab)"},{"href":"https://github.com/amirarjmand93/aime2026_loss_function","label":"Code","ariaLabel":"Code for ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG (opens in new tab)"}]},{"slug":"fallnest-dual-domain-condensation-imu-fall-detection-2026","title":"FALLNEST: Dual-Domain Dataset Condensation of Imbalanced IMU Data for Wearable Fall Detection","summary":"Wearable fall detection is hampered by rare fall events, scarce labeled IMU data, and the cost of training deep models on large, imbalanced corpora. Existing methods either rely on hand-crafted features with shallow classifiers, which limits learning from raw signals, or train deep networks on full datasets that require re-sampling to counter class skew and demand substantial data and compute. To address this, the authors propose a novel Fall detection framework via Normalized-gradient and Embedding matched Synthetic Time-frequency condensation, dubbed FALLNEST, a data-centric approach that distills large, imbalanced IMU datasets into tiny, class-balanced synthetic training sets.","venue":"arXiv 2610.18260v1 (submitted 2026-10-13; eess.SP; comment: '10 pages, 2 figures, for Conference')","year":2026,"open_code":null,"updated":"2026-10-14","url":"/research/papers/fallnest-dual-domain-condensation-imu-fall-detection-2026","links":[{"href":"https://arxiv.org/abs/2610.18260","label":"Paper","ariaLabel":"Paper for FALLNEST: Dual-Domain Dataset Condensation of Imbalanced IMU Data for Wearable Fall Detection (opens in new tab)"}]},{"slug":"femwear-womens-health-wearable-fm-2026","title":"FemWear: A Parameter-Efficient Wearable Foundation Model for Women's Health","summary":"CUHK-Shenzhen team takes Stanford's open OpenMHC wearable backbone (frozen, 21.5M params) and adds a tiny (1.11%, 239K param) low-rank adapter plus task-family heads to specialize it for women's-health outcomes — menstrual phase/onset, symptom severity (cramps, mood, sleep), pregnancy, and general ability retention. It is explicitly not a new pretraining effort: it's a parameter-efficient fine-tuning (PEFT) recipe layered on someone else's foundation model, evaluated on six cohorts including the hormone-verified mcPHASES dataset (42 participants, Fitbit + urinary LH/estrogen/progesterone).","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/femwear-womens-health-wearable-fm-2026","links":[{"href":"https://arxiv.org/abs/2608.08244","label":"Paper","ariaLabel":"Paper for FemWear: A Parameter-Efficient Wearable Foundation Model for Women's Health (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2608.08244","label":"DOI","ariaLabel":"DOI for FemWear: A Parameter-Efficient Wearable Foundation Model for Women's Health (opens in new tab)"}]},{"slug":"filmamba-film-conditioned-mamba-cuffless-bp-2026","title":"FiLMamba: A FiLM-Conditioned Mamba Architecture with an Energy-Efficient FPGA Accelerator for Generalizable Cuffless Blood Pressure Estimation","summary":"Addresses PPG morphology–BP ambiguity (similar waveforms → different pressures across subjects → mean regression) with a selective state-space (Mamba) model conditioned via Feature-wise Linear Modulation (FiLM) on a 4-D hemodynamic state vector, plus Cross-Domain Hemodynamic Alignment (CDHA) (unlabeled BN recalibration + per-calibration-unit bias correction). Evaluated on MIMIC-BP (1,524 patients): SBP/DBP SDE 6.45 / 4.96 mmHg, >2.5× better than unconditioned baselines; claims AAMI compliance. Training-free transfer to UCI-BP: 4.29 / 3.24 mmHg SDE — abstract claims first AAMI-compliant training-free cross-dataset cuffless BP result.","venue":"IEEE Transactions on Biomedical Circuits and Systems (TBCAS)","year":2026,"open_code":false,"updated":"2026-10-01","url":"/research/papers/filmamba-film-conditioned-mamba-cuffless-bp-2026","links":[{"href":"https://doi.org/10.1109/TBCAS.2026.3736710","label":"Paper","ariaLabel":"Paper for FiLMamba: A FiLM-Conditioned Mamba Architecture with an Energy-Efficient FPGA Accelerator for Generalizable Cuffless Blood Pressure Estimation (opens in new tab)"}]},{"slug":"finger-to-ear-ecg-earbud-iswc-2026","title":"Finger-to-Ear ECG: Systematic Evaluation of an Earbud-Based Cardiac Monitoring Approach","summary":"This study tests an earbud that records ECG between an in-ear electrode and a finger electrode on the back of the housing. Thirty participants tried four electrode layouts against a chest-reference ECG, including during music playback, talking and walking. Left-finger layouts detected R peaks with F1 above 99% and closely matched key ECG features.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/finger-to-ear-ecg-earbud-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834819","label":"Paper","ariaLabel":"Paper for Finger-to-Ear ECG: Systematic Evaluation of an Earbud-Based Cardiac Monitoring Approach (opens in new tab)"},{"href":"https://arxiv.org/abs/2609.24530","label":"arXiv","ariaLabel":"arXiv for Finger-to-Ear ECG: Systematic Evaluation of an Earbud-Based Cardiac Monitoring Approach (opens in new tab)"}]},{"slug":"found-af-benchmark-ecg-fm-2026","title":"FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection","summary":"FOUND-AF benchmarks ECG foundation models for atrial fibrillation detection under identical, leakage-controlled conditions. It evaluates nine public models from five families, including HuBERT-ECG, CLEF, ST-MEM, ECG-JEPA and ECGFounder, on four datasets: AFDB, CinC2017, CPSC2021 and LTAFDB.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/found-af-benchmark-ecg-fm-2026","links":[{"href":"https://arxiv.org/abs/2608.03597","label":"arXiv","ariaLabel":"arXiv for FOUND-AF: Benchmarking ECG Foundation Models for Atrial Fibrillation Detection (opens in new tab)"}]},{"slug":"rms-ecg-fm-reduced-lead-rheumatic-mitral-stenosis-medrxiv-2026","title":"Foundation Model Artificial Intelligence for Electrocardiogram Detection of Rheumatic Mitral Stenosis","summary":"This study tested whether an ECG foundation model can detect rheumatic mitral stenosis from a limited set of indirectly labelled ECGs. The model was fine-tuned on 804 positive ECGs per fold and validated externally on clinical and wearable lead setups. AUROC was 0.89, 0.87 and 0.87 in held-out testing for 12, 6 and 1 leads, and 0.82, 0.80 and 0.77 in external validation.","venue":"openRxiv","year":2026,"open_code":null,"updated":"2026-10-14","url":"/research/papers/rms-ecg-fm-reduced-lead-rheumatic-mitral-stenosis-medrxiv-2026","links":[{"href":"https://doi.org/10.64898/2026.10.12.26365300","label":"Paper","ariaLabel":"Paper for Foundation Model Artificial Intelligence for Electrocardiogram Detection of Rheumatic Mitral Stenosis (opens in new tab)"}]},{"slug":"har-fm-survey-2026-review","title":"Foundation Models Defining A New Era In Sensor-based HAR — survey review","summary":"So: there is no benchmark to scrutinize, because they didn't run one. Item 4 of the brief (which reported numbers survive scrutiny?) has a trivial answer for this artifact — none, because none are reported. The leakage critique has to be applied to the cited primaries instead, which I did below.","venue":"IMWUT Vol 10 Issue 2, Art. 31","year":2026,"open_code":null,"updated":null,"url":"/research/papers/har-fm-survey-2026-review","links":[{"href":"https://doi.org/10.1145/3810230","label":"DOI","ariaLabel":"DOI for Foundation Models Defining A New Era In Sensor-based HAR — survey review (opens in new tab)"},{"href":"https://arxiv.org/abs/2604.02711","label":"arXiv","ariaLabel":"arXiv for Foundation Models Defining A New Era In Sensor-based HAR — survey review (opens in new tab)"}]},{"slug":"movement-foundation-models-prime-time-benchmark-2026","title":"Foundation models for movement data: Are they ready for prime-time?","summary":"This study compares four open-source accelerometer foundation models with supervised baselines across 19 tasks in activity recognition, clinical monitoring and physiological inference. Supervised models stayed competitive on activity recognition, while some foundation models led on fall and stress detection and coped best with changes in sensor placement.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/movement-foundation-models-prime-time-benchmark-2026","links":[{"href":"https://arxiv.org/abs/2608.13316","label":"Paper","ariaLabel":"Paper for Foundation models for movement data: Are they ready for prime-time? (opens in new tab)"}]},{"slug":"glucofm-continuous-glucose-monitoring-fm","title":"GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring","summary":"GlucoFM is a lightweight foundation model for continuous glucose monitoring. It aligns readings to a 24-hour grid, splits glucose into slow trend and short-term deviation streams, and was pretrained on 109,066 hours of CGM data from 477 people. It gave the best linear-probe results across four cohorts and seven tasks, with the biggest gains on diabetes risk and beta-cell dysfunction.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/glucofm-continuous-glucose-monitoring-fm","links":[{"href":"https://arxiv.org/abs/2605.30865","label":"Paper","ariaLabel":"Paper for GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring (opens in new tab)"}]},{"slug":"glucoprism-protocol-supervised-cgm-pretraining-2026","title":"GlucoPRISM: Protocol-Supervised Pretraining for Cross-Cohort Glucose Transfer","summary":"A CGM encoder built on the GlucoFM backbone whose 128-d representation is split by \"protocol supervision\" into Trait (64) / State (48) / Sensor (16) bands. Paired Dexcom↔Libre recordings, different-day pairs, synthetic sensor views and glucose-summary targets tell each band what it may encode. At deployment the Sensor band is dropped (a fixed filter that needs no target-cohort data).","venue":"arXiv 2610.12528v1 (cs.LG), submitted 2026-10-08","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/glucoprism-protocol-supervised-cgm-pretraining-2026","links":[{"href":"https://arxiv.org/abs/2610.12528","label":"arXiv","ariaLabel":"arXiv for GlucoPRISM: Protocol-Supervised Pretraining for Cross-Cohort Glucose Transfer (opens in new tab)"}]},{"slug":"halo-imu-language-aligned-open-set-har-2026","title":"HALO: A Heterogeneity-Aware Language-Aligned IMU Foundation Model for Open-Set Human Activity Recognition","summary":"HALO is an IMU foundation model aligned with language so it can recognise activities it was not trained on, while handling different sampling rates, channels and sensor positions. Trained on 10 public datasets and tested on 7 held-out ones, it beat five strong baselines on all 8 summary metrics. With about 35M trainable parameters, ten times fewer than MOMENT, it raised zero-shot open-set accuracy by 13.7 points.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-29","url":"/research/papers/halo-imu-language-aligned-open-set-har-2026","links":[{"href":"https://arxiv.org/abs/2608.27233","label":"Paper","ariaLabel":"Paper for HALO: A Heterogeneity-Aware Language-Aligned IMU Foundation Model for Open-Set Human Activity Recognition (opens in new tab)"}]},{"slug":"healthloopqa-context-aware-qa-benchmark-diabetes-wearables-2026","title":"HealthLoopQA: A Context-Aware Question Answering Benchmark for Interpreting Wearable Monitoring Data in Diabetes Care","summary":"HealthLoopQA is a benchmark for testing how well large language models reason over continuous diabetes monitoring data. It is built on eleven basic reasoning abilities and contains 127 tasks and more than 1,500 questions covering process mining, anomaly detection and prediction over 30-day windows.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/healthloopqa-context-aware-qa-benchmark-diabetes-wearables-2026","links":[{"href":"https://arxiv.org/abs/2609.06976","label":"arXiv","ariaLabel":"arXiv for HealthLoopQA: A Context-Aware Question Answering Benchmark for Interpreting Wearable Monitoring Data in Diabetes Care (opens in new tab)"}]},{"slug":"hearts-llm-reasoning-health-time-series","title":"HEARTS: Benchmarking LLM Reasoning on Health Time Series","summary":"HEARTS is a benchmark for testing how large language models reason over health time series at several levels. The authors evaluated 16 leading LLMs on more than 20K test samples and offer it as a living benchmark for health-signal agents.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/hearts-llm-reasoning-health-time-series","links":[{"href":"https://arxiv.org/abs/2603.06638","label":"Paper","ariaLabel":"Paper for HEARTS: Benchmarking LLM Reasoning on Health Time Series (opens in new tab)"}]},{"slug":"hippt-physiology-grounded-wearable-ppg-multi-scope-text-alignment-2026","title":"HiPPT: Physiology-Grounded Wearable PPG Representation Learning via Multi-Scope Text Alignment","summary":"HiPPT trains a wrist PPG encoder by aligning recordings with text at different scopes. The authors built a corpus of more than 4 million wrist recordings from 202K users, each paired with descriptions of pulse shape and context. Individual recordings are aligned with their morphology, and groups of recordings that share the same context are aligned together.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/hippt-physiology-grounded-wearable-ppg-multi-scope-text-alignment-2026","links":[{"href":"https://arxiv.org/abs/2610.14489","label":"Paper","ariaLabel":"Paper for HiPPT: Physiology-Grounded Wearable PPG Representation Learning via Multi-Scope Text Alignment (opens in new tab)"}]},{"slug":"hypertension-awake-sleep-ppg-aurora-ehjdh-2026","title":"Hypertension screening via awake–sleep differences in photoplethysmogram signals","summary":"Hypothesis: normotensives show bigger awake-vs-sleep PPG changes than unmanaged hypertensives. 20 PPG indices; per-subject Mann–Whitney p-values (awake vs sleep) discretized as features. Aurora-BP (n = 180; 138 normotensive, 42 hypertensive; 18 held out): SVM accuracy 81.1 ± 8.4%, F1 82.8%.","venue":"European Heart Journal - Digital Health","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/hypertension-awake-sleep-ppg-aurora-ehjdh-2026","links":[{"href":"https://doi.org/10.1093/ehjdh/ztag131","label":"Paper","ariaLabel":"Paper for Hypertension screening via awake–sleep differences in photoplethysmogram signals (opens in new tab)"}]},{"slug":"inertia-1-motion-fm-study-2026","title":"Inertia-1: An Open Exploration of Wearable Motion Foundation Models","summary":"Inertia-1 is a fully open study of foundation models for wearable motion data. Using more than 18.2M hours of accelerometer data from around the world, the authors built a controlled framework to test how sensor type, placement, sampling rate, window length, architecture, model size, pretraining objective and data scale affect results.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/inertia-1-motion-fm-study-2026","links":[{"href":"https://arxiv.org/abs/2607.06617","label":"Paper","ariaLabel":"Paper for Inertia-1: An Open Exploration of Wearable Motion Foundation Models (opens in new tab)"}]},{"slug":"inertia1","title":"Inertia-1: Open Exploration of Wearable Motion Foundation Models","summary":"Inertia-1 is a fully open study of foundation models for wearable motion data. Using more than 18.2M hours of accelerometer data from around the world, the authors built a controlled framework to test how sensor type, placement, sampling rate, window length, architecture, model size, pretraining objective and data scale affect results.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/inertia1","links":[{"href":"https://doi.org/10.48550/arXiv.2607.06617","label":"DOI","ariaLabel":"DOI for Inertia-1: Open Exploration of Wearable Motion Foundation Models (opens in new tab)"},{"href":"https://arxiv.org/abs/2607.06617","label":"arXiv","ariaLabel":"arXiv for Inertia-1: Open Exploration of Wearable Motion Foundation Models (opens in new tab)"}]},{"slug":"insulin-resistance-wearables-biomarker","title":"Insulin Resistance Prediction from Wearables and Routine Blood Biomarkers","summary":"The WEAR-ME study, published in Nature, predicted insulin resistance in 1,165 remote participants using wearable time series and routine blood tests, with HOMA-IR as ground truth. The multimodal model reached AUROC 0.80, and a wearable foundation model pretrained on 40 million hours of sensor data added value. In a separate cohort of 72, adding wearable features to demographics, fasting glucose and lipids raised AUROC from 0.76 to 0.88.","venue":"Nature 2026","year":2026,"open_code":null,"updated":null,"url":"/research/papers/insulin-resistance-wearables-biomarker","links":[{"href":"https://www.nature.com/articles/s41586-026-10179-2","label":"Paper","ariaLabel":"Paper for Insulin Resistance Prediction from Wearables and Routine Blood Biomarkers (opens in new tab)"}]},{"slug":"zhou-language-supervised-ecg-vision-transformer-hiss-2026","title":"Language-supervised electrocardiogram interpretation with vision transformer","summary":"This study investigated whether vision-language learning from paired ECG recordings and summary reports could enable transferable ECG representations without task-specific supervised training. The authors developed ECG-VLM, a contrastive vision-language model consisting of a Transformer-based image encoder and a text encoder, trained on 800,034 paired 10-s ECG recordings and summary reports from the MIMIC-IV-ECG database. Cardiac condition detection was performed using natural-language prompts without condition-specific retraining or fine-tuning.","venue":"Health Information Science and Systems (2026-10-13)","year":2026,"open_code":null,"updated":"2026-10-14","url":"/research/papers/zhou-language-supervised-ecg-vision-transformer-hiss-2026","links":[{"href":"https://doi.org/10.1007/s13755-026-00482-7","label":"DOI","ariaLabel":"DOI for Language-supervised electrocardiogram interpretation with vision transformer (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42876871","label":"PubMed","ariaLabel":"PubMed for Language-supervised electrocardiogram interpretation with vision transformer (opens in new tab)"}]},{"slug":"lapras-latent-reasoning-time-series-language-models-2026","title":"Lapras: Latent Reasoning for Time Series Language Models","summary":"Post-training framework for time-series language models (TSLMs): instead of verbalising the signal in an explicit chain-of-thought, the model reasons through continuous \"thoughts\" in the joint time-series–language space and emits text only for the answer. Trained by teacher–student self-distillation: a CoT-SFT teacher reasons in text, and the student aligns its answer-stage hidden states to the teacher's. Evaluated on four TSLM backbones (ChatTS-7B, OpenTSLM-1B Flamingo variant, SLIP-1B, ITFormer-0.5B) and five QA tasks: ECG (ECG-Reasoning benchmark, Oh et al.","venue":"arXiv 2610.11111 (cs.CL; v1 submitted 2026-10-08)","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/lapras-latent-reasoning-time-series-language-models-2026","links":[{"href":"https://arxiv.org/abs/2610.11111","label":"arXiv","ariaLabel":"arXiv for Lapras: Latent Reasoning for Time Series Language Models (opens in new tab)"}]},{"slug":"bilateral-ppg-cross-site-attention-kd-hr-sd4h-2026","title":"Learning from Bilateral PPG via Cross-Site Attention and Knowledge Distillation","summary":"Dual PaPaGei encoders (one per wrist), SQI-conditioned temporal gating and cross-wrist self-attention. Bilateral sports dataset: 6 activities, 30 subjects, ECG truth. Bilateral model 2.85 bpm MAE, r = 0.96.","venue":"ICML 2026 Workshop SD4H (spotlight)","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/bilateral-ppg-cross-site-attention-kd-hr-sd4h-2026","links":[{"href":"https://openreview.net/forum?id=oKm0d3S5DM","label":"Paper","ariaLabel":"Paper for Learning from Bilateral PPG via Cross-Site Attention and Knowledge Distillation (opens in new tab)"}]},{"slug":"sleepfm-2-transferable-sleep-physiology-2026","title":"Learning transferable human physiology from two million hours of sleep with SleepFM-2","summary":"Rahul Thapa, Christopher Sun, William Theodor Lehn-Schiøler, Sophia Claire Kivelson, Umaer Hanif, Hyatt Moore IV, Harrison G. Zhang, Hafsa Ahmed, Marcus Dige, Niels R. Lorenzen, Elisabeth Roxane M.","venue":"arXiv preprint, cs.AI","year":2026,"open_code":null,"updated":"2026-09-27","url":"/research/papers/sleepfm-2-transferable-sleep-physiology-2026","links":[{"href":"https://arxiv.org/abs/2609.06849","label":"Paper","ariaLabel":"Paper for Learning transferable human physiology from two million hours of sleep with SleepFM-2 (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2609.06849","label":"DOI","ariaLabel":"DOI for Learning transferable human physiology from two million hours of sleep with SleepFM-2 (opens in new tab)"}]},{"slug":"wearable-1lead-ecg-lvf-van-der-valk-2026","title":"Leveraging the wearable 1-lead ECG signal. From cardiac rhythm to cardiac function assessment","summary":"Home 30-second 1-lead ECG from a Withings ECG Smartwatch (300 Hz), recorded in a clinical telemonitoring program after acute coronary syndrome at LUMC (2018–2023), is used to predict left ventricular function. Matched clinical 12-lead ECGs are 10 s at 500 Hz. The model is a partial beta-VAE with seven recurrent convolutional blocks, the same strategy as van der Valk et al., IEEE TBME 2025 (explainable ECG by VAE).","venue":"npj Cardiovascular Health (Article in Press)","year":2026,"open_code":false,"updated":"2026-10-04","url":"/research/papers/wearable-1lead-ecg-lvf-van-der-valk-2026","links":[{"href":"https://doi.org/10.1038/s44325-026-00153-2","label":"Paper","ariaLabel":"Paper for Leveraging the wearable 1-lead ECG signal. From cardiac rhythm to cardiac function assessment (opens in new tab)"}]},{"slug":"liteway-lightweight-har-edge-2026","title":"LITEWAY: LIghtweight HAR via Temporal Efficient highWAY","summary":"LITEWAY is a fully convolutional activity recognition model that replaces recurrent layers with structured convolutions. Tested on 16 datasets against TinyHAR, TinierHAR and MLP-HAR, it kept competitive macro-F1 while being 3.9 to 9.5 times smaller. On-device tests showed 1.5 to 3.1 times lower energy use.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/liteway-lightweight-har-edge-2026","links":[{"href":"https://arxiv.org/abs/2608.09421","label":"Paper","ariaLabel":"Paper for LITEWAY: LIghtweight HAR via Temporal Efficient highWAY (opens in new tab)"},{"href":"https://doi.org/10.1145/3830727.3834818","label":"DOI","ariaLabel":"DOI for LITEWAY: LIghtweight HAR via Temporal Efficient highWAY (opens in new tab)"}]},{"slug":"aharoni-menstrual-cycle-cgm-glycemic-variability-hpp-medrxiv-2026","title":"Menstrual cycle-related dynamics of glycemic variability in women without diabetes","summary":"209 naturally cycling women without diabetes (age 43.2 ± 5.2, BMI 24.3 ± 3.7) from the Human Phenotype Project. CGM worn up to 14 days with real-time diet logging, 3,815 meals. Cycle position normalized 0–100 % with ovulation estimated at 50 %.","venue":"medRxiv (endocrinology), v1 posted 2026-10-15","year":2026,"open_code":null,"updated":"2026-10-16","url":"/research/papers/aharoni-menstrual-cycle-cgm-glycemic-variability-hpp-medrxiv-2026","links":[{"href":"https://doi.org/10.64898/2026.10.14.26365513","label":"DOI","ariaLabel":"DOI for Menstrual cycle-related dynamics of glycemic variability in women without diabetes (opens in new tab)"}]},{"slug":"mimic-abp-s2s","title":"MIMIC ABP Waveform-to-Waveform Translation (s2s)","summary":"Use MIMIC-III-Ext-PPG (30-second WFDB segments with co-registered PPG, ECG, ABP) to train a sequence-to-sequence model translating PPG waveforms to ABP waveforms. The idea: if the FM encoder captures hemodynamic information, fine-tuning a decoder to reconstruct ABP would demonstrate a stronger form of blood pressure signal than PTT-based methods.","venue":"Exploratory concept","year":2026,"open_code":null,"updated":null,"url":"/research/papers/mimic-abp-s2s","links":[{"href":"https://physionet.org/content/mimic-iii-ext-ppg/1.0.0","label":"Paper","ariaLabel":"Paper for MIMIC ABP Waveform-to-Waveform Translation (s2s) (opens in new tab)"}]},{"slug":"mimic-iii-ext-ppg-dataset","title":"MIMIC-III-Ext-PPG: A PPG Benchmark Dataset for Cardiorespiratory Analysis","summary":"MIMIC-III-Ext-PPG is a quality-assessed PPG dataset taken from the MIMIC-III Waveform Database Matched Subset, published on PhysioNet by Moulaeifard, Charlton and Strodthoff. It has about 6.3 million 30-second segments from 6,189 ICU patients, labelled with heart rhythms such as sinus rhythm, atrial fibrillation, flutter, heart block and pacing. Where available it adds ECG, arterial blood pressure and respiration, with derived blood pressure, heart rate, respiratory rate and signal quality labels.","venue":"PhysioNet [dataset]","year":2026,"open_code":null,"updated":"2026-09-28","url":"/research/papers/mimic-iii-ext-ppg-dataset","links":[{"href":"https://physionet.org/content/mimic-iii-ext-ppg/1.1.0/","label":"Paper","ariaLabel":"Paper for MIMIC-III-Ext-PPG: A PPG Benchmark Dataset for Cardiorespiratory Analysis (opens in new tab)"}]},{"slug":"ml4h-2026-workshop","title":"ML4H 2026 Symposium (Machine Learning for Health)","summary":"ML4H 2026 is the premier NeurIPS workshop for machine learning in health, co-located in Sydney Dec 6–7 2026. Paper deadline: Sep 10 2026 (imminent). Demo deadline: Sep 14 2026.","venue":"NeurIPS workshop — Sydney, Australia","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ml4h-2026-workshop","links":[{"href":"https://ml4h.ahli.cc/","label":"Paper","ariaLabel":"Paper for ML4H 2026 Symposium (Machine Learning for Health) (opens in new tab)"}]},{"slug":"mmae-ecg-multiscale-mae-2026-review","title":"MMAE-ECG: multi-scale masked autoencoder for ECG anomaly detection","summary":"MMAE-ECG is a multi-scale masked autoencoder for detecting and locating anomalies in ECG without R-peak detection or heartbeat segmentation. It combines multi-scale masking with multi-scale attention so that a lightweight Transformer encoder captures both global and local patterns. It reached state-of-the-art anomaly detection and localization while using about 1/78 of the inference FLOPs and 1/18 of the trainable parameters of the previous leading method.","venue":"PLoS ONE 21(3):e0343571","year":2026,"open_code":null,"updated":null,"url":"/research/papers/mmae-ecg-multiscale-mae-2026-review","links":[{"href":"https://pubmed.ncbi.nlm.nih.gov/41843891","label":"PubMed","ariaLabel":"PubMed for MMAE-ECG: multi-scale masked autoencoder for ECG anomaly detection (opens in new tab)"}]},{"slug":"models-got-talent-zero-cost-proxies-har-nas-iswc-2026","title":"Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training","summary":"This paper tests zero-cost proxies, quick scores computed from one forward and backward pass, for picking good activity recognition architectures without training them. Across six datasets, the top-ranked architectures came within 7% of the best of 2,000 fully trained ones, and training just the top 10 came within 2%.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/models-got-talent-zero-cost-proxies-har-nas-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834826","label":"Paper","ariaLabel":"Paper for Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training (opens in new tab)"},{"href":"https://arxiv.org/abs/2511.06157","label":"arXiv","ariaLabel":"arXiv for Models Got Talent: Identifying High Performing Wearable Human Activity Recognition Models Without Training (opens in new tab)"}]},{"slug":"morphcl-morphological-contrastive-imu-har-2026","title":"MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition","summary":"MorphCL is a self-supervised pretraining method for inertial data that groups samples by structure, using motion motifs and domain feature descriptors, instead of relying on random batches. This matters for in-the-wild data, which is mostly still, low-variance movement. It substantially improved linear-probe and fine-tuning results.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-08","url":"/research/papers/morphcl-morphological-contrastive-imu-har-2026","links":[{"href":"https://arxiv.org/abs/2610.10245","label":"Paper","ariaLabel":"Paper for MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition (opens in new tab)"}]},{"slug":"morphologyfm-ecg-spo2-morphology-aware-fm-2026","title":"MorphologyFM: A Foundation Model for Morphology-Aware Representation Learning from ECG and Pulse Oximetry Waveforms","summary":"MorphologyFM is a foundation model pretrained on paired ECG and SpO2 waveforms from the MIMIC critical care database with an objective that preserves clinically meaningful waveform shape. It beat masked autoencoders, contrastive learning, Barlow Twins and JEPA on arrhythmia classification, hypoxemia prediction, mortality and length of stay.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/morphologyfm-ecg-spo2-morphology-aware-fm-2026","links":[{"href":"https://arxiv.org/abs/2607.09749","label":"arXiv","ariaLabel":"arXiv for MorphologyFM: A Foundation Model for Morphology-Aware Representation Learning from ECG and Pulse Oximetry Waveforms (opens in new tab)"}]},{"slug":"ms-ecg-fm-multi-source-contrastive-ecg-fm-2026","title":"MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning","summary":"MS-ECG-FM is an ECG foundation model trained by contrastive alignment with several kinds of clinical notes: ECG, echocardiography, radiology and discharge reports. The idea is that ECG reports alone miss information in the waveform that other notes capture.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-07","url":"/research/papers/ms-ecg-fm-multi-source-contrastive-ecg-fm-2026","links":[{"href":"https://arxiv.org/abs/2610.07662","label":"Paper","ariaLabel":"Paper for MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning (opens in new tab)"}]},{"slug":"lee-multimodal-ssl-ppg-acc-sleep-icassp-2026","title":"Multimodal Self-Supervised Learning for Wearable Sleep Staging Using Photoplethysmography and Accelerometer Signals","summary":"Joint SSL pretraining of PPG and accelerometer encoders on 23,000 hours of unlabeled wearable data, then a causal Mamba sequence model for staging so it could run in real time. On DREAMT (dreamt), 4-class staging improves over fully supervised baselines: AUROC 0.59 → 0.72, Cohen's κ 0.06 → 0.30. The pretraining corpus source isn't named in the abstract (authors are Samsung; likely Galaxy Watch data, unconfirmed).","venue":"ICASSP 2026","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/lee-multimodal-ssl-ppg-acc-sleep-icassp-2026","links":[{"href":"https://doi.org/10.1109/ICASSP55912.2026.11462724","label":"Paper","ariaLabel":"Paper for Multimodal Self-Supervised Learning for Wearable Sleep Staging Using Photoplethysmography and Accelerometer Signals (opens in new tab)"}]},{"slug":"lee-gao-multiplex-sweat-estradiol-progesterone-aptamer-device-2026","title":"Multiplex ovarian hormone monitoring in sweat using a dual-aptamer wearable sensor","summary":"Wearable sweat sensor for multiplex E2 and P4. Two redox-orthogonal aptamer–reporter systems (methylene blue, ferrocene) sit on one inkjet-printed gold electrode and are read by sequential square-wave voltammetry at picomolar concentrations. A passive microfluidic module normalises sweat pH and ionic strength, so no extra correction channels are needed, across sweat-induction methods and body sites.","venue":"Device (Cell Press), 2026","year":2026,"open_code":false,"updated":"2026-10-15","url":"/research/papers/lee-gao-multiplex-sweat-estradiol-progesterone-aptamer-device-2026","links":[{"href":"https://doi.org/10.1016/j.device.2026.101299","label":"DOI","ariaLabel":"DOI for Multiplex ovarian hormone monitoring in sweat using a dual-aptamer wearable sensor (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42871933","label":"PubMed","ariaLabel":"PubMed for Multiplex ovarian hormone monitoring in sweat using a dual-aptamer wearable sensor (opens in new tab)"}]},{"slug":"omniimu-bench-wrist-imu-representations-iclr2027-sub","title":"OmniIMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks","summary":"Benchmark of ten tasks from eight datasets for wrist IMU multi-task performance and cross-task transfer (activity, gesture, pose & velocity estimation, odometry, clinical motor assessment, IMU–language retrieval). Compares six SSL methods pretrained on unlabelled recorded and simulated wrist motion. Masked autoencoding is best/joint-best frozen-probe on 9/10 tasks; fine-tuning beats scratch on 9/10.","venue":"ICLR 2027 submission (under review); public release as IMU-Bench (Li, Bradshaw, Gao, Ahuja 2026; arXiv 2610.19481, submitted 2026-10-14)","year":2026,"open_code":true,"updated":"2026-10-15","url":"/research/papers/omniimu-bench-wrist-imu-representations-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=6UOlH8mAqO","label":"Paper","ariaLabel":"Paper for OmniIMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks (opens in new tab)"},{"href":"https://arxiv.org/abs/2610.19481","label":"arXiv","ariaLabel":"arXiv for OmniIMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks (opens in new tab)"}]},{"slug":"psg-fm-vs-sleep-scorers-sleep-wake-disorder-identification-2026","title":"On Sleep-Wake Disorder Identification from Polysomnographic Representations","summary":"Compared: self-supervised PSG FMs SleepFM, OSF, Hypnos vs sleep-stage-supervised scorers U-Sleep and SleepTransformer (pretrained in the SLEEPYLAND protocol on ~220k h of expert-annotated PSG from 17 NSRR cohorts; EMG/ECG/RESP counterparts trained the same way). Downstream: linear probing and a missingness-aware coupled-Mamba state-space model; fixed patient-grouped 5-fold CV.","venue":"arXiv 2610.17451 (v1 2026-10-12, cs.LG)","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/psg-fm-vs-sleep-scorers-sleep-wake-disorder-identification-2026","links":[{"href":"https://arxiv.org/abs/2610.17451","label":"Paper","ariaLabel":"Paper for On Sleep-Wake Disorder Identification from Polysomnographic Representations (opens in new tab)"}]},{"slug":"ontoar-har-ontology-guided-agentic-rag-har-2026","title":"OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition","summary":"A tree-structured activity ontology turns HAR into a sequence of local decisions among related activities (e.g. stationary → sit/stand/lie). At each node an agentic LLM combines (a) node-constrained examples from a static population store, (b) executable signal features discovered offline by an LLM-propose / rule-execute / error-feedback loop (≤5 rounds), and (c) a personalized memory that collects the target user's own confidently-labelled windows during inference.","venue":"arXiv 2610.15292 (v1 2026-10-11, cs.AI)","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/ontoar-har-ontology-guided-agentic-rag-har-2026","links":[{"href":"https://arxiv.org/abs/2610.15292","label":"Paper","ariaLabel":"Paper for OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition (opens in new tab)"}]},{"slug":"openmhc-open-wearable-fm-benchmark-2026","title":"OpenMHC: Accelerating the Science of Wearable Foundation Models","summary":"OpenMHC is a wearable health dataset from more than a decade of the My Heart Counts study app, released to qualified researchers with open-source code for recent wearable foundation models. It has more than 60 million hours of data across 19 channels, such as steps, heart rate, sleep and workouts, and up to 169 linked health and lifestyle variables from 11,894 participants.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/openmhc-open-wearable-fm-benchmark-2026","links":[{"href":"https://arxiv.org/abs/2607.16235","label":"Paper","ariaLabel":"Paper for OpenMHC: Accelerating the Science of Wearable Foundation Models (opens in new tab)"}]},{"slug":"osf-open-sleep-foundation-models","title":"OSF: On Pre-training and Scaling of Sleep Foundation Models","summary":"OSF studies how pretraining and scale affect sleep foundation models across different polysomnography devices and cohorts. The authors gathered 166,500 hours of sleep recordings from nine public sources and released SleepBench, a fully open benchmark.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/osf-open-sleep-foundation-models","links":[{"href":"https://arxiv.org/abs/2603.00190","label":"Paper","ariaLabel":"Paper for OSF: On Pre-training and Scaling of Sleep Foundation Models (opens in new tab)"}]},{"slug":"p2e-vq-ecg-linked-representation-retrieval-ppg-2026","title":"P2E-VQ: ECG-linked representation augmentation for PPG via discrete patch retrieval","summary":"P2E-VQ helps PPG models pick up cardiac information that normally needs ECG. It turns PPG patches into discrete tokens and retrieves ECG-linked information from a memory bank built from training data, instead of reconstructing ECG waveforms. Across five datasets and six tasks it beat pretrained baselines under frozen linear probing.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/p2e-vq-ecg-linked-representation-retrieval-ppg-2026","links":[{"href":"https://arxiv.org/abs/2608.14656","label":"arXiv","ariaLabel":"arXiv for P2E-VQ: ECG-linked representation augmentation for PPG via discrete patch retrieval (opens in new tab)"}]},{"slug":"passive-heart-rate-smartphone-rppg","title":"Passive Heart-Rate Monitoring During Smartphone Use in Everyday Life","summary":"This Nature paper presents passive heart-rate monitoring (PHRM), a deep learning system that measures heart rate from the front camera's face video while people use their smartphones. It was developed on 192,353 videos from 485 participants and validated on 162,546 videos from 211 participants in the lab and in daily life. Against ECG, heart-rate error stayed below 10% (MAPE) for light, medium and dark skin tones, and daily resting heart rate was within 5 bpm of a wearable tracker.","venue":"Nature 2026","year":2026,"open_code":null,"updated":null,"url":"/research/papers/passive-heart-rate-smartphone-rppg","links":[{"href":"https://www.nature.com/articles/s41586-026-10507-6","label":"Paper","ariaLabel":"Paper for Passive Heart-Rate Monitoring During Smartphone Use in Everyday Life (opens in new tab)"}]},{"slug":"pat","title":"PAT: Pretrained Actigraphy Transformer","summary":"PAT is an open-source transformer foundation model for week-long, minute-level actigraphy from smartwatches. It uses patch embeddings and masked autoencoder pretraining and was evaluated on psychiatric outcomes. On predicting benzodiazepine use it beat an LSTM by 55.6%.","venue":"IEEE Journal of Biomedical and Health Informatics","year":2026,"open_code":null,"updated":null,"url":"/research/papers/pat","links":[{"href":"https://doi.org/10.1109/JBHI.2026.3694809","label":"DOI","ariaLabel":"DOI for PAT: Pretrained Actigraphy Transformer (opens in new tab)"},{"href":"https://arxiv.org/abs/2411.15240","label":"arXiv","ariaLabel":"arXiv for PAT: Pretrained Actigraphy Transformer (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42154714","label":"PubMed","ariaLabel":"PubMed for PAT: Pretrained Actigraphy Transformer (opens in new tab)"}]},{"slug":"peak-detector-instruction-tuned-llm-explainable-peak-detection-2026","title":"Peak-Detector: Explainable Peak Detection via Instruction-Tuned Large Language Models in Physiological Signal","summary":"Peak-Detector uses instruction-tuned large language models to find peaks in cardiac signals and explain its choices. Across ECG, PPG, ballistocardiography and bodyseismography in seven datasets, it gave the best or tied-best detection within clinical timing tolerances, and its explanations help spot failure modes.","venue":"Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/peak-detector-instruction-tuned-llm-explainable-peak-detection-2026","links":[{"href":"https://doi.org/10.1145/3810224","label":"DOI","ariaLabel":"DOI for Peak-Detector: Explainable Peak Detection via Instruction-Tuned Large Language Models in Physiological Signal (opens in new tab)"},{"href":"https://arxiv.org/abs/2605.16452","label":"arXiv","ariaLabel":"arXiv for Peak-Detector: Explainable Peak Detection via Instruction-Tuned Large Language Models in Physiological Signal (opens in new tab)"}]},{"slug":"personalized-calibration-hybrid-ppg-ecg-bp-sensors-2026","title":"Personalized Calibration and Hybrid Feature Fusion for Continuous Blood Pressure Estimation Using PPG and ECG Signals","summary":"Handcrafted PTT/HRV/PPG-morphology features + masked-reconstruction SSL embeddings + subject prior BP (alpha-weighted base BP from the earliest calibration file) → XGBoost. MAE 7.25/3.98 (private set) and 7.85/4.02 mmHg (UCI). Small-sample feasibility per the authors.","venue":"Sensors (MDPI) 26(15):4676","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/personalized-calibration-hybrid-ppg-ecg-bp-sensors-2026","links":[{"href":"https://doi.org/10.3390/s26154676","label":"Paper","ariaLabel":"Paper for Personalized Calibration and Hybrid Feature Fusion for Continuous Blood Pressure Estimation Using PPG and ECG Signals (opens in new tab)"}]},{"slug":"stepfm-step-count-foundation-model-2026","title":"Physical activities enable scalable foundation modelling for broad-spectrum health prediction (StepFM)","summary":"StepFM is a foundation model built only on step-counter data for broad health prediction. Its pretraining learns temporal and behavioural patterns from large-scale step sequences, and it transfers to more than 20 health risk tasks across devices, regions and diseases.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/stepfm-step-count-foundation-model-2026","links":[{"href":"https://arxiv.org/abs/2607.06954","label":"Paper","ariaLabel":"Paper for Physical activities enable scalable foundation modelling for broad-spectrum health prediction (StepFM) (opens in new tab)"}]},{"slug":"pchs-harmonic-ppg-hr-rr-2026","title":"Physically-Constrained Harmonic Separation (PCHS) for Robust HR/RR from Wrist PPG","summary":"Small Morocco-based group (UIR / UM6P / Leeds) proposes an analysis-by-synthesis framework for wrist-PPG HR and RR estimation. Instead of regressing HR/RR directly from raw PPG+accel, the model predicts a small set of physiologically meaningful parameters (fundamental frequency, 3 harmonic amplitudes, gain, baseline, plus a residual and an uncertainty term) and synthesizes the clean PPG waveform from them via an explicit sinusoidal generative model. HR comes from the recovered fundamental frequency; RR comes from slow modulations of the harmonic parameters.","venue":"IEEE EMBC 2026 (accepted)","year":2026,"open_code":null,"updated":null,"url":"/research/papers/pchs-harmonic-ppg-hr-rr-2026","links":[{"href":"https://arxiv.org/abs/2606.30156","label":"Paper","ariaLabel":"Paper for Physically-Constrained Harmonic Separation (PCHS) for Robust HR/RR from Wrist PPG (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2606.30156","label":"DOI","ariaLabel":"DOI for Physically-Constrained Harmonic Separation (PCHS) for Robust HR/RR from Wrist PPG (opens in new tab)"}]},{"slug":"pockettslm-on-device-soft-prompt-tslm-2026","title":"PocketTSLM: Deploying an On-Device Soft-Prompt Time-Series Language Model for Privacy-Preserving Smartphone Reasoning","summary":"Ports OpenTSLM-SP (soft-prompt variant: a time-series encoder + projector inject signal windows as embeddings into a LoRA-adapted Llama-3.2-1B) to MLX-Swift, so inference runs entirely on a consumer iPhone (iPhone Air, iOS 26.4.1) with no health data leaving the device. Parity protocol: encoder, projector and rotary embedding match the Python MLX reference within 10⁻⁴ max element-wise. Greedy-decoded token sequences are identical between Swift and Python MLX; final answers match PyTorch.","venue":"arXiv 2610.21308v1 (cs.LG, cs.CL; submitted 2026-10-15); a version accepted at the Workshop on Deploying AI in Healthcare (DAIH), COLM 2026","year":2026,"open_code":true,"updated":"2026-10-16","url":"/research/papers/pockettslm-on-device-soft-prompt-tslm-2026","links":[{"href":"https://arxiv.org/abs/2610.21308","label":"arXiv","ariaLabel":"arXiv for PocketTSLM: Deploying an On-Device Soft-Prompt Time-Series Language Model for Privacy-Preserving Smartphone Reasoning (opens in new tab)"}]},{"slug":"ppg-heart-age-ecg-guided-2026","title":"PPG Heart Age via ECG-Guided Cross-Modal Pretraining (arXiv:2608.01620)","summary":"This study estimates heart age from smartwatch PPG, using synchronized smartwatch ECG to guide pretraining so that only PPG is needed in use. It draws on three OPPO cohorts in China with 581,804 participants and 7,452,131 recordings. The PPG-only model estimated age with errors of 5.9 years (r = 0.82) and 4.3 years (r = 0.80) in two external cohorts, and a larger heart-age gap was linked to higher pulse wave velocity.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ppg-heart-age-ecg-guided-2026","links":[{"href":"https://arxiv.org/abs/2608.01620","label":"arXiv","ariaLabel":"arXiv for PPG Heart Age via ECG-Guided Cross-Modal Pretraining (arXiv:2608.01620) (opens in new tab)"}]},{"slug":"ppg-bench-unified-benchmark-ppg-foundation-models-2026","title":"PPG-Bench: A Unified Benchmark for PPG Foundation Models","summary":"PPG-Bench is an open benchmark that compares five public PPG foundation models on nine datasets and 26 tasks under four levels of adaptation, from linear probes to full fine-tuning. In 35.3% of the 68 task-regime combinations the foundation models showed no real advantage over a trivial baseline. In 83.8%, no PPG-specific model beat both simple pulse-rate-variability features and MOMENT.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/ppg-bench-unified-benchmark-ppg-foundation-models-2026","links":[{"href":"https://arxiv.org/abs/2610.13435","label":"arXiv","ariaLabel":"arXiv for PPG-Bench: A Unified Benchmark for PPG Foundation Models (opens in new tab)"},{"href":"https://github.com/mannas-ai/ppg-bench","label":"Code","ariaLabel":"Code for PPG-Bench: A Unified Benchmark for PPG Foundation Models (opens in new tab)"}]},{"slug":"ppg-fusionnet-dual-branch-cuffless-bp-computers-2026","title":"PPG-FusionNet: A Dual-Branch Neural Architecture for Cuffless Blood Pressure Estimation from Photoplethysmography","summary":"Shared patch embedding feeds a local dilated-conv encoder and a global AutoCorrelation encoder, fused by cross-attention, with a constrained dual-head SBP/DBP loss. Bayesian hyperparameter search and ablations on PulseDB. SBP/DBP MAE 7.46 / 4.72 mmHg, near-zero mean error; AAMI-compliant and BHS grade B for DBP only.","venue":"Computers (MDPI) 15(9):566","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/ppg-fusionnet-dual-branch-cuffless-bp-computers-2026","links":[{"href":"https://doi.org/10.3390/computers15090566","label":"Paper","ariaLabel":"Paper for PPG-FusionNet: A Dual-Branch Neural Architecture for Cuffless Blood Pressure Estimation from Photoplethysmography (opens in new tab)"}]},{"slug":"ppg-lm-multi-level-clinical-alignment-2026","title":"PPG-LM: A Photoplethysmography–Language Model with Multi-Level Clinical Alignment","summary":"PPG-LM is a family of PPG-language models that learn from both signal-derived physiological labels and the clinical context in electronic health records. The authors argue that task-specific heads in earlier PPG models and the medical knowledge of LLMs each miss part of the picture, and PPG-LM combines them.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-29","url":"/research/papers/ppg-lm-multi-level-clinical-alignment-2026","links":[{"href":"https://arxiv.org/abs/2609.33516","label":"Paper","ariaLabel":"Paper for PPG-LM: A Photoplethysmography–Language Model with Multi-Level Clinical Alignment (opens in new tab)"}]},{"slug":"practical-optimization-dl-ppg-bp-sampling-physiolmeas-2026","title":"Practical optimization of deep learning models for cuffless blood pressure estimation from photoplethysmography","summary":"205,850 ten-second PPG–ABP segments from the MIMIC waveform database; multiresolution U-Net PPG→ABP. 62.5 Hz beats 125 Hz (SBP 5.24 vs 6.34; DBP 2.87 vs 3.27 mmHg MAE); 31.25 Hz fails (16.09 / 9.10). Averaging the top-4 peaks / bottom-4 troughs beats min–max extraction.","venue":"Physiological Measurement","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/practical-optimization-dl-ppg-bp-sampling-physiolmeas-2026","links":[{"href":"https://doi.org/10.1088/1361-6579/ae944f","label":"Paper","ariaLabel":"Paper for Practical optimization of deep learning models for cuffless blood pressure estimation from photoplethysmography (opens in new tab)"}]},{"slug":"pretraining-on-sleep-data-improves-non-sleep-biosignal-tasks-2026","title":"Pretraining on Sleep Data Improves non-Sleep Biosignal Tasks","summary":"This paper asks whether pretraining on sleep recordings helps with biosignal tasks outside sleep. The authors pretrained only on multimodal sleep data with a leave-one-out contrastive objective and tested on non-sleep EEG and ECG. Across eight tasks, sleep pretraining consistently beat training from scratch.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/pretraining-on-sleep-data-improves-non-sleep-biosignal-tasks-2026","links":[{"href":"https://arxiv.org/abs/2605.02500","label":"arXiv","ariaLabel":"arXiv for Pretraining on Sleep Data Improves non-Sleep Biosignal Tasks (opens in new tab)"}]},{"slug":"prism-retrieval-augmented-multimodal-lm-clinical-prognosis-2026","title":"PRISM: Retrieval-Augmented Multimodal Language Model with Heterogeneous Evidence Integration for Clinical Prognosis","summary":"PRISM is a multimodal language model for clinical prognosis that encodes numeric features as tokens, weights recent ECGs more heavily, and retrieves averaged embeddings of similar past patients. No individual retrieved record goes into the prompt. On the 1,443-task MDS-ED benchmark it reached AUROCs of 86.2% for diagnosis, 93.1% for deterioration, 91.8% for ICU admission and 92.8% for mortality.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/prism-retrieval-augmented-multimodal-lm-clinical-prognosis-2026","links":[{"href":"https://arxiv.org/abs/2610.13510","label":"arXiv","ariaLabel":"arXiv for PRISM: Retrieval-Augmented Multimodal Language Model with Heterogeneous Evidence Integration for Clinical Prognosis (opens in new tab)"}]},{"slug":"pulseanchor-personalizing-ppg-fm-bp-bsn-2026","title":"PulseAnchor: Personalizing PPG Foundation Models for Cuffless Blood Pressure Estimation","summary":"Reformulates BP as relative prediction: contrast a test sample against subject-specific calibration examples from a PPG FM's embedding space. On Aurora-BP: SBP 7.84 / DBP 5.47 mmHg MAE, 15% / 22% better than existing personalization approaches. Which FM backbone and how many calibration readings aren't in the abstract.","venue":"IEEE BSN 2026 (OpenReview)","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/pulseanchor-personalizing-ppg-fm-bp-bsn-2026","links":[{"href":"https://openreview.net/forum?id=EkS1OhIfIJ","label":"Paper","ariaLabel":"Paper for PulseAnchor: Personalizing PPG Foundation Models for Cuffless Blood Pressure Estimation (opens in new tab)"}]},{"slug":"quality-aware-wrist-ppg-stress-wesad-zenodo-2026","title":"Quality-Aware Wrist Photoplethysmography for Subject Independent Stress-State Monitoring: ECG-Referenced Validation and Selective Prediction","summary":"WESAD wrist BVP (64 Hz) + wrist ACC (32 Hz) + chest ECG (700 Hz), 15 participants, 865 60-s windows (30-s step), baseline vs stress, LOSO. Raw PRV best macro-F1 0.838; PPG+ACC best AUROC 0.956 / AUPRC 0.933; full model AUROC 0.952 with lowest Brier 0.116. Selective prediction: error 13.33% at 79.8% coverage → 7.48% at 61.9%.","venue":"Zenodo record (Text; part of ISSN 2581-7175)","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/quality-aware-wrist-ppg-stress-wesad-zenodo-2026","links":[{"href":"https://doi.org/10.5281/zenodo.22858828","label":"Paper","ariaLabel":"Paper for Quality-Aware Wrist Photoplethysmography for Subject Independent Stress-State Monitoring: ECG-Referenced Validation and Selective Prediction (opens in new tab)"}]},{"slug":"rae-ppg-duration-retain-extend-2026","title":"RAE-PPG: Duration-Grounded Retain-and-Extend Pretraining for PPG Foundation Models","summary":"RAE-PPG trains one Transformer PPG encoder in stages on 10 s, 30 s and 240 s inputs, adding targets for features that need longer recordings at each stage. Features learned early stay recoverable, and learning earlier stages first made later features easier to learn.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-09-30","url":"/research/papers/rae-ppg-duration-retain-extend-2026","links":[{"href":"https://arxiv.org/abs/2609.36794","label":"Paper","ariaLabel":"Paper for RAE-PPG: Duration-Grounded Retain-and-Extend Pretraining for PPG Foundation Models (opens in new tab)"}]},{"slug":"crm-1d-cnn-lbnp-compensatory-reserve-2025","title":"Real-time comparison of ML-enabled devices for compensatory reserve (CRM vs FDA-cleared CRI)","summary":"This study compared two PPG-based algorithms that estimate compensatory reserve, a measure of how close the body is to decompensating during blood loss. Twenty participants went through a lower body negative pressure (LBNP) chamber protocol that simulates progressive central hypovolemia, while a new compensatory reserve measurement (CRM) device and an FDA-cleared compensatory reserve index (CRI) device ran side by side on noninvasive PPG sensors. Pooled CRM and CRI values agreed closely (R² = 0.859) with a near-zero mean bias of 0.13%. The authors conclude that both devices track compensatory changes in real time, which could help with monitoring and triage in prehospital trauma care.","venue":"Bioengineering","year":2026,"open_code":null,"updated":"2026-10-13","url":"/research/papers/crm-1d-cnn-lbnp-compensatory-reserve-2025","links":[{"href":"https://doi.org/10.3390/bioengineering13070817","label":"DOI","ariaLabel":"DOI for Real-time comparison of ML-enabled devices for compensatory reserve (CRM vs FDA-cleared CRI) (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42510482","label":"PubMed","ariaLabel":"PubMed for Real-time comparison of ML-enabled devices for compensatory reserve (CRM vs FDA-cleared CRI) (opens in new tab)"}]},{"slug":"llm-home-ecg-report-explanation-real-world-interactions-npj-2026","title":"Real-world user interactions with large language models for explaining home electrocardiogram reports","summary":"Retrospective analysis of 109,017 conversational sessions / 323,470 dialogue turns from 26,419 users of the Wen-Xin-Wu-Yang AI-ECG system over 16 months, where an LLM explains portable/home ECG reports. 60.01% of interactions happened outside working hours and 57.90% of users came from lower-tier cities. Users showed relatively high retention and moderate depth; their needs moved from result interpretation to medical knowledge, action guidance and health management.","venue":"npj Digital Medicine (published 2026-10-07; open access CC BY-NC-ND 4.0)","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/llm-home-ecg-report-explanation-real-world-interactions-npj-2026","links":[{"href":"https://doi.org/10.1038/s41746-026-03330-w","label":"Paper","ariaLabel":"Paper for Real-world user interactions with large language models for explaining home electrocardiogram reports (opens in new tab)"}]},{"slug":"ppg-glucose-feasibility-reassessment-2026","title":"Reassessing the Feasibility of PPG-Based Non-Invasive Blood Glucose Level Estimation","summary":"This paper re-tests five published methods for estimating blood glucose from PPG with a reproducible pipeline and three increasingly strict ways of splitting the data. The models looked competitive when windows were split at random but collapsed once the same person could not appear in both training and test sets, with R² near zero or negative, no better than predicting the mean. Over 90% of predictions still landed in the Clarke Error Grid A+B zones, including the mean baseline, which shows that metric alone is misleading.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/ppg-glucose-feasibility-reassessment-2026","links":[{"href":"https://arxiv.org/abs/2608.01820","label":"Paper","ariaLabel":"Paper for Reassessing the Feasibility of PPG-Based Non-Invasive Blood Glucose Level Estimation (opens in new tab)"}]},{"slug":"refit-label-free-imu-placement-shift-correction-2026","title":"REFIT: Recognize, Fix, and Test Wearable Sensor Placement Shifts without Labels","summary":"A gradient-free, label-free input calibration that sits in front of a frozen IMU activity classifier. It models a watch moved to the other wrist (a reflection) or a strap sensor put back turned (one of 24 axis-aligned rotations) as orthogonal axis transforms. The gyroscope picks up an extra det(A) sign because it measures an axial vector.","venue":"arXiv 2610.08991 (submitted 2026-10-06 18:49 UTC; in the 2026-10-07 evening ET listing)","year":2026,"open_code":false,"updated":"2026-10-08","url":"/research/papers/refit-label-free-imu-placement-shift-correction-2026","links":[{"href":"https://arxiv.org/abs/2610.08991","label":"Paper","ariaLabel":"Paper for REFIT: Recognize, Fix, and Test Wearable Sensor Placement Shifts without Labels (opens in new tab)"}]},{"slug":"reliability-evidence-fusion-selective-hr-wearable-ppg-2026","title":"Reliability-Evidence Fusion for Selective Heart-Rate Estimation from Wearable PPG: A Transformation-Aware Cross-Cohort Study","summary":"Frames PPG HR reliability as fusion of reliability evidence around a fixed HR estimate: signal-quality scores, learned heteroscedastic scale, and structured-output dispersion are fused to rank windows for retain/abstain without changing the estimate. Proves weak oracle dominance and decomposes AURC gain. Shows that for reciprocal pulse-interval→HR conversion the scale ratio is (ĥ/h)², which can cause operating-range-dependent ranking reversal — the propagated model-scale channel reversed ordering at a reference HR split of 110 bpm.","venue":"SSRN preprint","year":2026,"open_code":false,"updated":"2026-10-05","url":"/research/papers/reliability-evidence-fusion-selective-hr-wearable-ppg-2026","links":[{"href":"https://doi.org/10.2139/ssrn.7565250","label":"Paper","ariaLabel":"Paper for Reliability-Evidence Fusion for Selective Heart-Rate Estimation from Wearable PPG: A Transformation-Aware Cross-Cohort Study (opens in new tab)"}]},{"slug":"reliagate-reliability-routing-stress-2026","title":"ReliaGate: Reliability Routing for Low-Stakes Wearable Stress Prediction","summary":"It locks an upstream classifier (per-modality TCNs on BVP/EDA/TEMP, or BVP/EDA for UBFC-Phys), then trains a correctness scorer on ~90 reliability features: confidence, per-modality signal quality, cross-modality agreement, train- standardized atypicality, and train-fitted embedding geometry (Mahalanobis / k-NN distances). It evaluates on four datasets (WESAD, UBFC-Phys, EmpaticaE4Stress, PhysioNetE4) with subject-disjoint folds and reports not just coverage/risk but a per-subject \"Reach\" metric — explicitly because a selective system can look better in pooled numbers while starving specific people of any output.","venue":"UbiComp Companion 2026 (WellComp workshop)","year":2026,"open_code":null,"updated":"2026-09-27","url":"/research/papers/reliagate-reliability-routing-stress-2026","links":[{"href":"https://arxiv.org/abs/2608.15951","label":"Paper","ariaLabel":"Paper for ReliaGate: Reliability Routing for Low-Stakes Wearable Stress Prediction (opens in new tab)"},{"href":"https://github.com/jadenmoon27/reliagate-wellcomp2026","label":"Code","ariaLabel":"Code for ReliaGate: Reliability Routing for Low-Stakes Wearable Stress Prediction (opens in new tab)"}]},{"slug":"rethinking-ppg-sleep-staging-benchmark-2026","title":"Rethinking PPG-based Sleep Staging: Datasets, Metrics, and Benchmarks","summary":"This paper revisits sleep staging from wearable PPG, an alternative to lab polysomnography. The authors use a hidden semi-Markov model to expand 30-second epoch labels into second-level labels. Training with those finer labels on MESA improved standard four-class staging by 3.7 to 5.7 accuracy points across four different model types.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/rethinking-ppg-sleep-staging-benchmark-2026","links":[{"href":"https://arxiv.org/abs/2608.00943","label":"Paper","ariaLabel":"Paper for Rethinking PPG-based Sleep Staging: Datasets, Metrics, and Benchmarks (opens in new tab)"}]},{"slug":"retrieval-augmented-personalization-fm-wearable-stress-2026","title":"Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection","summary":"This paper personalizes wearable stress detection without user labels by having frozen foundation models retrieve similar moments from the same user's history. The retrieved items become a compact embedding that adjusts a small transformer. On WESAD with 15 users it improved accuracy by 3.92% and macro-F1 by 4.76% over a non-personalized model, and retrieving from another dataset gave no significant gain.","venue":null,"year":2026,"open_code":false,"updated":"2026-10-05","url":"/research/papers/retrieval-augmented-personalization-fm-wearable-stress-2026","links":[{"href":"https://doi.org/10.1145/3776591.3836729","label":"DOI","ariaLabel":"DOI for Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection (opens in new tab)"}]},{"slug":"review-open-foundation-models-datasets-ecg-ppg-npj-2026","title":"Review of open foundation models and datasets for ECG and PPG waveforms","summary":"Clinician-oriented review of open ECG/PPG FMs and datasets: data-efficiency and generalisability advantages; challenges are homogeneous pretraining data and no standardised evaluation.","venue":"npj Digital Medicine (2026-08-07), DOI 10.1038/s41746-026-03101-7","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/review-open-foundation-models-datasets-ecg-ppg-npj-2026","links":[{"href":"https://doi.org/10.1038/s41746-026-03101-7","label":"DOI","ariaLabel":"DOI for Review of open foundation models and datasets for ECG and PPG waveforms (opens in new tab)"}]},{"slug":"tri-har-rotation-invariant-multi-imu-har-iswc-2026","title":"Rotation-Invariant Multi-IMU Activity Recognition under Independent Per-Location Orientation Shifts","summary":"TRI-HAR builds rotation invariance into multi-IMU activity recognition, so each sensor can be mounted at a different orientation. On four multi-IMU benchmarks it kept macro-F1 under such rotations and beat rotation-augmented baselines without needing rotation augmentation.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/tri-har-rotation-invariant-multi-imu-har-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834824","label":"Paper","ariaLabel":"Paper for Rotation-Invariant Multi-IMU Activity Recognition under Independent Per-Location Orientation Shifts (opens in new tab)"},{"href":"https://arxiv.org/abs/2608.15621","label":"arXiv","ariaLabel":"arXiv for Rotation-Invariant Multi-IMU Activity Recognition under Independent Per-Location Orientation Shifts (opens in new tab)"}]},{"slug":"samsung-research-health-foundation-models-2026","title":"Samsung Research Health Foundation Models: xMAE and HiMAE","summary":"This August 2026 Samsung Newsroom article describes two health foundation models from Samsung Research trained on wearable data. xMAE learns the timing relationship between different biosignals and was accepted at ICML, and HiMAE learns health patterns across time scales in wearable time series and was accepted at ICLR.","venue":"Samsung Research (news); xMAE at ICML 2026; HiMAE at ICLR 2026","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/samsung-research-health-foundation-models-2026","links":[{"href":"https://news.samsung.com/global/from-biosignals-to-health-insights-samsung-researchs-work-on-health-foundation-models","label":"Paper","ariaLabel":"Paper for Samsung Research Health Foundation Models: xMAE and HiMAE (opens in new tab)"}]},{"slug":"llm-affective-personal-informatics-scoping-review-2026","title":"Scoping the Landscape of Large Language Model-based Affective Personal Informatics: Review, Reflection, Future Direction","summary":"Affective personal informatics (PI) involves capturing, tracking, analyzing, and reflecting on personal data to support emotional self-awareness and mental well-being. Recent advances in large language models (LLMs) are reshaping affective PI, yet a comprehensive survey of LLM-based affective PI systems is lacking. Their synthesis reveals an overreliance on multi-stage, text-centric pipelines, a predominance of active, screen-bound interactions, and a severe demographic skew toward young, university-campus-based cohorts in empirical evaluations.","venue":"arXiv 2610.20084 (submitted 2026-10-14)","year":2026,"open_code":false,"updated":"2026-10-15","url":"/research/papers/llm-affective-personal-informatics-scoping-review-2026","links":[{"href":"https://arxiv.org/abs/2610.20084","label":"Paper","ariaLabel":"Paper for Scoping the Landscape of Large Language Model-based Affective Personal Informatics: Review, Reflection, Future Direction (opens in new tab)"}]},{"slug":"opensla-sensor-language-action-models-2026","title":"Sensor-Language-Action Models (OpenSLA)","summary":"OpenSLA links sensor data, language and actions in one model, using language as the bridge between sensing and acting. It predicts actions at several levels, describes states and explains its predictions while staying grounded in the sensor data.","venue":"arXiv preprint","year":2026,"open_code":true,"updated":"2026-10-08","url":"/research/papers/opensla-sensor-language-action-models-2026","links":[{"href":"https://arxiv.org/abs/2610.08244","label":"Paper","ariaLabel":"Paper for Sensor-Language-Action Models (OpenSLA) (opens in new tab)"}]},{"slug":"sensorfm-google-2026","title":"SensorFM: Toward a General Intelligence and Interface for Wearable Health Data","summary":"Google's SensorFM is a wearable health foundation model pretrained on more than one trillion minutes of unlabelled sensor data from five million participants. Scaling model size and data together steadily improved results across 35 health prediction tasks in cardiovascular, metabolic, sleep and mental health, lifestyle and demographics.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/sensorfm-google-2026","links":[{"href":"https://arxiv.org/abs/2605.22759","label":"arXiv","ariaLabel":"arXiv for SensorFM: Toward a General Intelligence and Interface for Wearable Health Data (opens in new tab)"}]},{"slug":"sensorgen-signal-or-noise","title":"Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series (SensorGen)","summary":"SensorGen is a large study of generative models for real-world sensor data, covering 14 settings, 4 domains, 7 datasets and 12 signal types. Flow-matching models did well in most settings, demographic covariates and time-frequency modelling helped, and synthetic data improved downstream performance.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/sensorgen-signal-or-noise","links":[{"href":"https://arxiv.org/abs/2607.04245","label":"Paper","ariaLabel":"Paper for Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series (SensorGen) (opens in new tab)"}]},{"slug":"simauthor-fm-authored-biosignal-simulators-2026","title":"SimAuthor: Harnessing Foundation Models for Persistent Scientific Simulator Authoring","summary":"SimAuthor uses foundation models to write and keep revising scientific simulators, programs that encode how a mechanism produces observed signals. Revisions are guided by comparing simulated and real signal distributions, and the harness keeps reusable implementation pieces. It was tested on six biomedical tasks, including heart and breathing sounds, PPG and ECG, and 111 of 138 audited revisions changed program structure.","venue":"arXiv preprint","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/simauthor-fm-authored-biosignal-simulators-2026","links":[{"href":"https://arxiv.org/abs/2610.06257","label":"Paper","ariaLabel":"Paper for SimAuthor: Harnessing Foundation Models for Persistent Scientific Simulator Authoring (opens in new tab)"}]},{"slug":"single-beat-ear-ppg-cuffless-bp-2026","title":"Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework","summary":"This study estimates blood pressure from single heartbeats using ear PPG and ECG. A 1D CNN turns each PPG beat into a 64-number embedding, which is combined with 30 physiological features such as pulse transit time and HRV and fed to LightGBM. Tested on a 10-person stress protocol and PulseDB, it reached MAEs of 4.02 mmHg systolic and 1.79 mmHg diastolic, 28.2% better than baselines.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/single-beat-ear-ppg-cuffless-bp-2026","links":[{"href":"https://arxiv.org/abs/2607.27076","label":"Paper","ariaLabel":"Paper for Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework (opens in new tab)"}]},{"slug":"sleeplm-sleep-language-model","title":"SleepLM: Natural-Language Intelligence for Human Sleep","summary":"SleepLM is a family of sleep-language foundation models that link sleep signals with natural language. The authors built a sleep-text dataset of more than 100K hours from over 10,000 people and trained with contrastive alignment, caption generation and signal reconstruction. It beat earlier methods on zero-shot and few-shot tasks, retrieval and sleep captioning.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/sleeplm-sleep-language-model","links":[{"href":"https://arxiv.org/abs/2602.23605","label":"Paper","ariaLabel":"Paper for SleepLM: Natural-Language Intelligence for Human Sleep (opens in new tab)"}]},{"slug":"slip-sensor-language-informed-pretraining-2026","title":"SLIP: Learning Transferable Sensor Models via Language-Informed Pretraining","summary":"Pretraining: ~600K sensor–caption pairs (~1 B time points). Community time-series corpora were captioned with the SensorLM recipe and augmented with ChatTS synthetic pairs. Domains include health, energy, IoT and more.","venue":"arXiv 2603.11950 (cs.AI; v1 2026-03-12, v2 2026-09-27) — preprint","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/slip-sensor-language-informed-pretraining-2026","links":[{"href":"https://arxiv.org/abs/2603.11950","label":"arXiv","ariaLabel":"arXiv for SLIP: Learning Transferable Sensor Models via Language-Informed Pretraining (opens in new tab)"}]},{"slug":"carat-reliance-guided-tta-multimodal-wearable-corruption-2026","title":"Source-Learned Reliance for Selective Test-Time Adaptation of Multimodal Time Series (CARAT)","summary":"Test-time adaptation for multimodal wearables when individual streams (wrist BVP/PPG, ACC, EDA, temp, chest ECG…) become noisy or drop out. During source training an asymmetric modality-dropout curriculum drops each modality in proportion to its windowed input-projection gradient norm; the terminal per-modality dropout rate is frozen as a reliance profile. At deployment a lightweight one-class detector (FITS reconstruction residuals + log-spectral distance vs.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-07","url":"/research/papers/carat-reliance-guided-tta-multimodal-wearable-corruption-2026","links":[{"href":"https://arxiv.org/abs/2610.07499","label":"Paper","ariaLabel":"Paper for Source-Learned Reliance for Selective Test-Time Adaptation of Multimodal Time Series (CARAT) (opens in new tab)"},{"href":"https://github.com/payalmohapatra/MAESTRO","label":"Code","ariaLabel":"Code for Source-Learned Reliance for Selective Test-Time Adaptation of Multimodal Time Series (CARAT) (opens in new tab)"}]},{"slug":"specbram-masked-band-power-eeg-fm-2026","title":"SpecBraM: What Should an EEG Foundation Model Predict? Masked Band-Power Prediction versus Waveform Reconstruction","summary":"SpecBraM compares two pretraining targets for EEG models: predicting the band power of masked patches versus reconstructing the masked waveform, with everything else held equal (2,388 hours of data). On ISRUC and HMC sleep staging, band-power prediction beat waveform reconstruction by 1.6 to 2.8 balanced-accuracy points with all labels and 4.7 to 7.3 points with 1% of labels.","venue":"arXiv preprint","year":2026,"open_code":false,"updated":"2026-10-07","url":"/research/papers/specbram-masked-band-power-eeg-fm-2026","links":[{"href":"https://arxiv.org/abs/2610.07484","label":"Paper","ariaLabel":"Paper for SpecBraM: What Should an EEG Foundation Model Predict? Masked Band-Power Prediction versus Waveform Reconstruction (opens in new tab)"}]},{"slug":"spotr-one-token-reconstruction-physiological-ssl-ijcai-2026","title":"SPOTR: Spatio-temporal Pooling One-Token Reconstruction for Universal Physiological Signal Self-supervised Learning","summary":"SPOTR is a pretraining method for physiological signals that squeezes each recording into one global token and reconstructs from it, with a module that cuts compute and memory. Pretrained on 20 datasets of EEG, iEEG, ECG and PPG, it beat the strongest baseline in linear probing by 4.64% AUC on PPG and 17.86% on ECG, with about 78% lower latency than a general time-series model.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/spotr-one-token-reconstruction-physiological-ssl-ijcai-2026","links":[{"href":"https://arxiv.org/abs/2606.21973","label":"arXiv","ariaLabel":"arXiv for SPOTR: Spatio-temporal Pooling One-Token Reconstruction for Universal Physiological Signal Self-supervised Learning (opens in new tab)"},{"href":"https://github.com/5GYYYYY/SPOTR","label":"Code","ariaLabel":"Code for SPOTR: Spatio-temporal Pooling One-Token Reconstruction for Universal Physiological Signal Self-supervised Learning (opens in new tab)"}]},{"slug":"stitchsense-smart-garment-imu-source-domain-watch-har-iswc-2026","title":"StitchSense: Smart-Garment Sensing as a Source Domain for Watch-Based Activity Recognition","summary":"StitchSense is a smart vest with five IMUs at the shoulders, chest and hips, built to study activity recognition from sensors sewn into clothing. Garment motion, fit and sensor shifts made fine-grained recognition from the vest alone hard, though it captured coarse movement well. Using vest data as a source domain together with a small amount of smartwatch calibration improved watch-based recognition.","venue":"ISWC '26 (ACM International Symposium on Wearable Computers), published 2026-10-10","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/stitchsense-smart-garment-imu-source-domain-watch-har-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834837","label":"Paper","ariaLabel":"Paper for StitchSense: Smart-Garment Sensing as a Source Domain for Watch-Based Activity Recognition (opens in new tab)"}]},{"slug":"take-it-personally-rl-ppg-ssl-emotion-acii-2026","title":"Take it Personally: The Limits of General SSL Representations for Real-Life PPG Emotion Detection","summary":"This paper tests self-supervised PPG representations for detecting intense emotions in daily life. They transferred very well to activity recognition, with almost five times the baseline performance in leave-one-subject-out tests, but did not beat simple baselines for emotion detection. Including a person's own data in fine-tuning mattered more than population-level pretraining.","venue":"arXiv preprint","year":2026,"open_code":true,"updated":"2026-10-12","url":"/research/papers/take-it-personally-rl-ppg-ssl-emotion-acii-2026","links":[{"href":"https://arxiv.org/abs/2608.14675","label":"Paper","ariaLabel":"Paper for Take it Personally: The Limits of General SSL Representations for Real-Life PPG Emotion Detection (opens in new tab)"}]},{"slug":"teaching-ppg-how-not-who-fixed-effects-ecg-distillation-2026","title":"Teaching PPG How not Who: Fixed-Effects Distillation from ECG","summary":"ECG is widely used as a privileged teacher for PPG-only models (AnyPPG-style alignment, ECG reconstruction). The authors show much of what is transferred is identity (trait), not state: a per-recording mean holds 40–59% of a frozen ECG teacher's target, pooled students memorise it, and the raw alignment cosine is misleading (a constant predictor scores 0.793). Across 34 runs, more identity memorised → less state learned.","venue":"arXiv 2610.10662 (v1 2026-10-07, cs.LG)","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/teaching-ppg-how-not-who-fixed-effects-ecg-distillation-2026","links":[{"href":"https://arxiv.org/abs/2610.10662","label":"Paper","ariaLabel":"Paper for Teaching PPG How not Who: Fixed-Effects Distillation from ECG (opens in new tab)"}]},{"slug":"tf-mambanet-bidirectional-mamba-ppg-fm-icassp-2026","title":"TF-MambaNet: A Temporal and Frequency Fused Bidirectional Mamba Architecture for PPG Foundation Model","summary":"Backbone: stacked bidirectional Mamba blocks (Gu & Dao 2023 selective SSM). A forward Mamba runs on x and a backward Mamba on the time-flipped x. The backward output is flipped back, and the two are fused as LayerNorm(hfw + hbw) + x.","venue":"IEEE ICASSP 2026 (Barcelona), poster MLSP-P25.3","year":2026,"open_code":false,"updated":"2026-10-08","url":"/research/papers/tf-mambanet-bidirectional-mamba-ppg-fm-icassp-2026","links":[{"href":"https://doi.org/10.1109/ICASSP55912.2026.11461042","label":"Paper","ariaLabel":"Paper for TF-MambaNet: A Temporal and Frequency Fused Bidirectional Mamba Architecture for PPG Foundation Model (opens in new tab)"}]},{"slug":"temporal-context-encoding-ecg-ssl-2026","title":"The Impact of Temporal Context Length and Encoding Strategies on Self-Supervised ECG Representation Learning","summary":"This study tests how input length and input encoding affect self-supervised ECG models, using single-lead Icentia11k with the same Transformer throughout. It compared 16-second, 1-minute, 5-minute and 10-minute inputs, and continuous patch embeddings against fixed quantized tokens. Longer context beat 16-second snapshots for rhythm detection and patient retrieval, with the 5- and 10-minute models best.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/temporal-context-encoding-ecg-ssl-2026","links":[{"href":"https://arxiv.org/abs/2608.12695","label":"Paper","ariaLabel":"Paper for The Impact of Temporal Context Length and Encoding Strategies on Self-Supervised ECG Representation Learning (opens in new tab)"}]},{"slug":"dataset-integrity-calibration-nn-bp-scirep-2026","title":"The role of dataset integrity, calibration and signal quality in neural network-based blood pressure estimation","summary":"MLP (PAT/HR/HRV scalars), ResNet and TCN on curated MIMIC-III and VitalDB sets that differ mainly in signal synchronization; subject-wise splits; calibration with 3–9 initial reference readings. VitalDB-trained models beat MIMIC-III ones (synchronization matters, especially for SBP). Calibration helps, with diminishing returns after 3 readings.","venue":"Scientific Reports","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/dataset-integrity-calibration-nn-bp-scirep-2026","links":[{"href":"https://doi.org/10.1038/s41598-026-66882-7","label":"Paper","ariaLabel":"Paper for The role of dataset integrity, calibration and signal quality in neural network-based blood pressure estimation (opens in new tab)"}]},{"slug":"timecap-mhcap-captioning-wearable-time-series-2026","title":"TimeCap: A Scalable Open-Source Pipeline for Captioning Wearable Health Time-Series Data","summary":"Problem: time-series language models (TSLMs) lack open, signal-grounded, reproducible text–time-series pairs. LLM-written captions cannot be regenerated exactly. Pipeline: deterministic captioner with three extractor tiers (statistical, structural, semantic) plus a cross-channel synthesis layer.","venue":"arXiv 2610.21274v1 (cs.LG, cs.CL; submitted 2026-10-15; 46 pp incl. supplement)","year":2026,"open_code":true,"updated":"2026-10-16","url":"/research/papers/timecap-mhcap-captioning-wearable-time-series-2026","links":[{"href":"https://arxiv.org/abs/2610.21274","label":"arXiv","ariaLabel":"arXiv for TimeCap: A Scalable Open-Source Pipeline for Captioning Wearable Health Time-Series Data (opens in new tab)"}]},{"slug":"normwear-2-world-model-physiological-2026","title":"Toward World Modeling of Physiological Signals with Chaos-Theoretic Balancing and Latent Dynamics","summary":"This paper introduces NormWear-2, a world model that puts physiological signals and clinical interventions into one latent space and models how they evolve together. It combines pretrained knowledge with fast non-parametric adaptation of latent transitions, so it can forecast over several time scales under different interventions.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-01","url":"/research/papers/normwear-2-world-model-physiological-2026","links":[{"href":"https://arxiv.org/abs/2605.15465","label":"Paper","ariaLabel":"Paper for Toward World Modeling of Physiological Signals with Chaos-Theoretic Balancing and Latent Dynamics (opens in new tab)"},{"href":"https://huggingface.co/mosaic-laboratory/normwear2","label":"Code","ariaLabel":"Code for Toward World Modeling of Physiological Signals with Chaos-Theoretic Balancing and Latent Dynamics (opens in new tab)"}]},{"slug":"distilling-ppg-fm-emotion-aware-wearables-mipr-2026","title":"Towards Emotion-Aware Wearables: Distilling PPG Foundation Models for Real-Time Affective and Physiological Monitoring","summary":"This paper compresses the Pulse-PPG foundation model into small 1D-MobileNetV1 and V2 networks with knowledge distillation so that emotion and physiology models can run on wearables. It was tested on WESAD, PPG-DaLiA and SDB across five tasks. Students with as few as 0.15M parameters were up to 190 times smaller and used up to 300 times less compute, and on WESAD 4-class affect they reached AUROC 0.90 versus 0.81 for the teacher.","venue":null,"year":2026,"open_code":false,"updated":"2026-10-05","url":"/research/papers/distilling-ppg-fm-emotion-aware-wearables-mipr-2026","links":[{"href":"https://doi.org/10.1109/MIPR70517.2026.00077","label":"DOI","ariaLabel":"DOI for Towards Emotion-Aware Wearables: Distilling PPG Foundation Models for Real-Time Affective and Physiological Monitoring (opens in new tab)"}]},{"slug":"mw-ppg-hr-fair-trustworthy-hr-2026","title":"Towards fair and trustworthy heart rate estimation from wrist-worn PPG (MW-PPG-HR)","summary":"This study aimed to make wrist PPG heart rate estimates fairer across skin tones and sexes and to flag unreliable readings. The authors collected a multi-wavelength PPG dataset (blue, green, red, infrared) across activities and built a PPG-accelerometer network that estimates its own uncertainty. Error was higher for darker skin and for women at baseline, and rejecting high-uncertainty predictions removed the skin-tone gap (1.2 bpm to 0.0 bpm).","venue":"Physiological Measurement","year":2026,"open_code":null,"updated":"2026-09-27","url":"/research/papers/mw-ppg-hr-fair-trustworthy-hr-2026","links":[{"href":"https://doi.org/10.1088/1361-6579/ae56ae","label":"Paper","ariaLabel":"Paper for Towards fair and trustworthy heart rate estimation from wrist-worn PPG (MW-PPG-HR) (opens in new tab)"},{"href":"https://github.com/danielray54/Towards-Fair-and-Trustworthy-Heart-Rate-Estimation-from-Wrist-worn-Photoplethysmography","label":"Code","ariaLabel":"Code for Towards fair and trustworthy heart rate estimation from wrist-worn PPG (MW-PPG-HR) (opens in new tab)"}]},{"slug":"tpf-bp-physiology-guided-ssl-personalized-cuffless-bp-crhe-2026","title":"TPF-BP: A Three-Stage Physiology-Guided Self-Supervised Transfer Learning Framework for Personalized Blood Pressure Measurement","summary":"TPF-BP is a three-stage transfer learning method for personalized cuffless blood pressure from PPG. It pretrains with a physiology-guided masked autoencoder, then does supervised pretraining and personal fine-tuning with a three-branch model. Applied to four base models, it cut systolic and diastolic errors by 7.66% and 5.78% on average in few-shot tests across four datasets.","venue":"IEEE Journal of Biomedical and Health Informatics","year":2026,"open_code":false,"updated":"2026-10-09","url":"/research/papers/tpf-bp-physiology-guided-ssl-personalized-cuffless-bp-crhe-2026","links":[{"href":"https://doi.org/10.1109/JBHI.2026.3740802","label":"DOI","ariaLabel":"DOI for TPF-BP: A Three-Stage Physiology-Guided Self-Supervised Transfer Learning Framework for Personalized Blood Pressure Measurement (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/42848570","label":"PubMed","ariaLabel":"PubMed for TPF-BP: A Three-Stage Physiology-Guided Self-Supervised Transfer Learning Framework for Personalized Blood Pressure Measurement (opens in new tab)"}]},{"slug":"trace-ts-sensor-language-reasoning-2026","title":"TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding","summary":"TRACE-TS trains a compact language model to recognise activities from wearable data and explain its answer with an evidence-linked reasoning trace. Training traces come from attribution maps of an expert classifier, but none of that is needed at inference. Across seven benchmarks it had the best average accuracy and F1 (84.43% / 81.24%), 17.96 F1 points above the best LLM baseline.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/trace-ts-sensor-language-reasoning-2026","links":[{"href":"https://arxiv.org/abs/2608.00200","label":"Paper","ariaLabel":"Paper for TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2608.00200","label":"DOI","ariaLabel":"DOI for TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding (opens in new tab)"},{"href":"https://github.com/SparshRastogi/TRACE-TS","label":"Code","ariaLabel":"Code for TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding (opens in new tab)"}]},{"slug":"aether-compute-optimal-physiological-waveform-fm-scaling-2026","title":"Training and Scaling Compute-Optimal Physiological Waveform Foundation Models","summary":"This paper studies scaling laws and compute-optimal training for physiological waveform foundation models. The authors trained Aether, more than 100 models from 20M to 2.1B parameters, on up to 36.3M hours of waveforms and tested them on eight clinical tasks built from MIMIC-III. The 720M model beat all existing baseline foundation models on all eight tasks.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-06","url":"/research/papers/aether-compute-optimal-physiological-waveform-fm-scaling-2026","links":[{"href":"https://openreview.net/forum?id=hRTpFhuTuq","label":"Paper","ariaLabel":"Paper for Training and Scaling Compute-Optimal Physiological Waveform Foundation Models (opens in new tab)"},{"href":"https://arxiv.org/abs/2610.05649","label":"arXiv","ariaLabel":"arXiv for Training and Scaling Compute-Optimal Physiological Waveform Foundation Models (opens in new tab)"}]},{"slug":"transfhar-self-supervised-wrist-imu-on-demand-har-uist-2026","title":"TransfHAR: Self-Supervised Wrist Representations for On-Demand Activity Recognition","summary":"TransfHAR pretrains on coarse wrist IMU activities like sitting and walking and transfers to fine-grained ones like snapping or stirring that were never in pretraining. It runs as a smartwatch app where users add their own activities from a few demonstrations. In a lab study with 10 participants it reached 86.7% balanced accuracy with five examples per activity.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/transfhar-self-supervised-wrist-imu-on-demand-har-uist-2026","links":[{"href":"https://arxiv.org/abs/2608.15861","label":"arXiv","ariaLabel":"arXiv for TransfHAR: Self-Supervised Wrist Representations for On-Demand Activity Recognition (opens in new tab)"}]},{"slug":"sensorclr-dualmoco-unsupervised-har-jiot-2026","title":"Unsupervised Sensor-Based Human Activity Recognition via Contrastive Learning of Representation (SensorCLR)","summary":null,"venue":"IEEE Internet of Things Journal, 2026 (S2 publication date 2026-10-15)","year":2026,"open_code":null,"updated":"2026-10-15","url":"/research/papers/sensorclr-dualmoco-unsupervised-har-jiot-2026","links":[{"href":"https://doi.org/10.1109/JIOT.2026.3719110","label":"DOI","ariaLabel":"DOI for Unsupervised Sensor-Based Human Activity Recognition via Contrastive Learning of Representation (SensorCLR) (opens in new tab)"}]},{"slug":"liu-3m-bpnet-multi-site-wrist-ppg-bp-jbhi-2025","title":"Utilizing Multi-PPG-Sensor Site Information in a Localized Wrist Area for Improving Cuffless Blood Pressure Estimation","summary":"The PPG Atlas says 13 features per channel, 27,214 beats and a 511-subset search. None of that appears in the abstract, and the abstract says 9 features were analysed. The train/test split (subject-disjoint or not) and calibration amount are unknown.","venue":"IEEE J. Biomed. Health Inform. 30(4):3579–3592 (issue 2026-04; early access 2025)","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/liu-3m-bpnet-multi-site-wrist-ppg-bp-jbhi-2025","links":[{"href":"https://doi.org/10.1109/JBHI.2025.3619070","label":"Paper","ariaLabel":"Paper for Utilizing Multi-PPG-Sensor Site Information in a Localized Wrist Area for Improving Cuffless Blood Pressure Estimation (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/41060852","label":"PubMed","ariaLabel":"PubMed for Utilizing Multi-PPG-Sensor Site Information in a Localized Wrist Area for Improving Cuffless Blood Pressure Estimation (opens in new tab)"}]},{"slug":"varmae-variable-input-mae-wrist-accelerometry-iclr2027-sub","title":"Variable Input Masked Autoencoders for Wrist-Worn Accelerometry (VarMAE)","summary":"VarMAE learns a single representation across sampling rates and temporal contexts: random crop + downsample creates complementary bottlenecks, and a cross-resolution masked reconstruction loss ties the views. Pretrained on NHANES (~1 week of wrist accelerometry from ~16k subjects); a single encoder matches or outperforms controlled FM baselines (each pretrained on NHANES in the same setup) on eight HAR datasets.","venue":"ICLR 2027 submission (under review); arXiv 2610.17383 (2026-10-12)","year":2026,"open_code":false,"updated":"2026-10-14","url":"/research/papers/varmae-variable-input-mae-wrist-accelerometry-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=otzhV2A17M","label":"Paper","ariaLabel":"Paper for Variable Input Masked Autoencoders for Wrist-Worn Accelerometry (VarMAE) (opens in new tab)"},{"href":"https://arxiv.org/abs/2610.17383","label":"arXiv","ariaLabel":"arXiv for Variable Input Masked Autoencoders for Wrist-Worn Accelerometry (VarMAE) (opens in new tab)"}]},{"slug":"vitalagent-tool-augmented-ecg-ppg-monitoring-agent-2026","title":"VitalAgent: A Tool-Augmented Agent for Reactive and Proactive Physiological Monitoring over Wearable Health Data","summary":"An LLM agent framework, not a foundation model and not a trained network. It answers user questions about ECG/PPG recordings (reactive mode) and streams long recordings to raise alerts (proactive mode), using one shared memory and tool set.","venue":"arXiv preprint (cs.AI), v1 2026-05-28 (v2 exists)","year":2026,"open_code":true,"updated":"2026-10-10","url":"/research/papers/vitalagent-tool-augmented-ecg-ppg-monitoring-agent-2026","links":[{"href":"https://arxiv.org/abs/2605.29483","label":"Paper","ariaLabel":"Paper for VitalAgent: A Tool-Augmented Agent for Reactive and Proactive Physiological Monitoring over Wearable Health Data (opens in new tab)"}]},{"slug":"vsmp-imu-synthetic-imu-generation-2026","title":"VSMP-IMU: Video-Grounded Semantic Motion Programs for Sensor-Aware Synthetic IMU Generation","summary":"VSMP-IMU generates synthetic IMU data from videos using structured semantic motion programs that separate what defines an activity from harmless variation. On five public datasets with leave-one-person-out testing, it reached an average macro-F1 of 78.33%, 9.77 points above real data alone and 4.04 above the best earlier synthetic method.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":null,"url":"/research/papers/vsmp-imu-synthetic-imu-generation-2026","links":[{"href":"https://arxiv.org/abs/2608.05782","label":"Paper","ariaLabel":"Paper for VSMP-IMU: Video-Grounded Semantic Motion Programs for Sensor-Aware Synthetic IMU Generation (opens in new tab)"}]},{"slug":"w-jepa-wavelet-jepa-wearable-har-2026","title":"W-JEPA: Wavelet Joint-Embedding Predictive Architecture for Wearable Activity Recognition","summary":"Inertial windows become a fixed grid of wavelet tokens indexed by scale × sensor axis × time. These tokens are the units for masked latent prediction against an EMA target encoder (JEPA). The aim is to keep scale-specific motion detail and learn cross-scale, cross-axis structure, so the model transfers across datasets with different placement, channel count and sampling rate.","venue":"arXiv 2610.19411 (submitted 2026-10-14; 11 pages)","year":2026,"open_code":null,"updated":"2026-10-15","url":"/research/papers/w-jepa-wavelet-jepa-wearable-har-2026","links":[{"href":"https://arxiv.org/abs/2610.19411","label":"Paper","ariaLabel":"Paper for W-JEPA: Wavelet Joint-Embedding Predictive Architecture for Wearable Activity Recognition (opens in new tab)"}]},{"slug":"wavesfm-hierarchical-longitudinal-wearable-waveforms-2026","title":"WavesFM: Hierarchical Representation Learning for Longitudinal Wearable Sensor Waveforms","summary":"WavesFM is a foundation model for weeks-long wearable waveforms that uses two stages of self-supervised learning. It aims to keep fine waveform detail and long-term behavioural patterns together, which earlier methods usually traded off.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/wavesfm-hierarchical-longitudinal-wearable-waveforms-2026","links":[{"href":"https://arxiv.org/abs/2605.09173","label":"arXiv","ariaLabel":"arXiv for WavesFM: Hierarchical Representation Learning for Longitudinal Wearable Sensor Waveforms (opens in new tab)"}]},{"slug":"large-sensor-models-wearable-ai-position-2026","title":"Wearable AI in the Era of Large Sensor Models","summary":"This position paper argues for Large Sensor Models: foundation models trained on large, multimodal wearable data. The authors say current sensor models are siloed by modality and task, and that LSMs could give wearable AI a more general and scalable base.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/large-sensor-models-wearable-ai-position-2026","links":[{"href":"https://arxiv.org/abs/2604.10172","label":"Paper","ariaLabel":"Paper for Wearable AI in the Era of Large Sensor Models (opens in new tab)"}]},{"slug":"wearableqa-health-reasoning-benchmark-real-wearable-data-2026","title":"WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data","summary":"WearableQA is a benchmark of 4,084 ten-option questions built from the wearable data, blood biomarkers and demographics of 200 real users, each with up to 500 days of daily measurements. Its 16 question types separate data computation from health interpretation and single-signal from cross-signal reasoning. Across 14 LLMs, accuracy ranged from 19.6% to 72.9%, and most scored below 60%.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/wearableqa-health-reasoning-benchmark-real-wearable-data-2026","links":[{"href":"https://openreview.net/forum?id=wZEk6U3vds","label":"Paper","ariaLabel":"Paper for WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data (opens in new tab)"},{"href":"https://arxiv.org/abs/2609.05405","label":"arXiv","ariaLabel":"arXiv for WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data (opens in new tab)"}]},{"slug":"wearicg-chest-only-icg-stroke-volume-iswc-2026","title":"WearICG: Towards A Wearable Approach to Continuous Stroke Volume Monitoring","summary":"WearICG tests whether impedance cardiography electrodes placed only on the chest can estimate stroke volume, which would suit a chest patch. Three chest-only layouts were compared with the standard neck-thorax layout in 17 participants across 5 hemodynamic states, against an FDA-cleared continuous blood pressure reference. The Inline-Chest layout tracked the standard with r = 0.91, and a simple bias correction raised stroke-volume correlation from 0.44 to 0.74.","venue":"ISWC '26 (ACM International Symposium on Wearable Computers), published 2026-10-10","year":2026,"open_code":null,"updated":"2026-10-12","url":"/research/papers/wearicg-chest-only-icg-stroke-volume-iswc-2026","links":[{"href":"https://doi.org/10.1145/3830727.3834832","label":"Paper","ariaLabel":"Paper for WearICG: Towards A Wearable Approach to Continuous Stroke Volume Monitoring (opens in new tab)"}]},{"slug":"weqa-wearable-health-qa-query-adaptive-agentic-reasoning-2026","title":"WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning","summary":"WEQA is an agent framework that answers health questions about wearable data by combining LLM reasoning with specialized analysis and modelling tools, choosing the workflow per question. The authors also built a benchmark from four open wearable datasets. WEQA was 24% more accurate than LLM and agent baselines, and 12 medical experts and 8 users rated it more useful and clinically sound.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/weqa-wearable-health-qa-query-adaptive-agentic-reasoning-2026","links":[{"href":"https://arxiv.org/abs/2606.18147","label":"arXiv","ariaLabel":"arXiv for WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning (opens in new tab)"}]},{"slug":"wonderwall-virtual-to-real-imu-har-2026","title":"Wonderwall: A Virtual-to-Real Foundation Model for IMU-based HAR","summary":"Wonderwall is an activity recognition foundation model pretrained first on physics-based simulated motion and then on curated real datasets. Across ten benchmarks it gave state-of-the-art generalization in zero-shot, few-shot and full-shot settings.","venue":"Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies","year":2026,"open_code":null,"updated":"2026-09-29","url":"/research/papers/wonderwall-virtual-to-real-imu-har-2026","links":[{"href":"https://doi.org/10.1145/3789688","label":"Paper","ariaLabel":"Paper for Wonderwall: A Virtual-to-Real Foundation Model for IMU-based HAR (opens in new tab)"}]},{"slug":"thapa-wrist-ppg-stress-wesad-interval-vs-morphology-2026","title":"Wrist PPG Stress Classification on WESAD: Interval Features Compared with Beat-Shape and Wavelet Features","summary":"This study compared interval features with beat-shape and wavelet features for stress detection from wrist PPG in WESAD (15 subjects, Empatica E4 at 64 Hz). Heart rate was the only feature that changed significantly (+24.4 bpm). In leave-one-subject-out tests, interval features alone reached AUC 0.911 to 0.933 and all features 0.934 to 0.965, with overlapping confidence intervals.","venue":"Zenodo","year":2026,"open_code":false,"updated":"2026-10-08","url":"/research/papers/thapa-wrist-ppg-stress-wesad-interval-vs-morphology-2026","links":[{"href":"https://doi.org/10.5281/zenodo.23206781","label":"Paper","ariaLabel":"Paper for Wrist PPG Stress Classification on WESAD: Interval Features Compared with Beat-Shape and Wavelet Features (opens in new tab)"}]},{"slug":"xmae-cross-modal-reconstruction-2026","title":"xMAE: Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning","summary":"xMAE pretrains on biosignals by reconstructing one signal from another, such as ECG and the delayed PPG pulse, so the model learns physiological timing. It beat unimodal and multimodal baselines on 15 of 19 tasks, from cardiovascular outcomes to sleep staging, across devices and body locations.","venue":"arXiv preprint","year":2026,"open_code":null,"updated":"2026-10-11","url":"/research/papers/xmae-cross-modal-reconstruction-2026","links":[{"href":"https://arxiv.org/abs/2605.00973v1","label":"Paper","ariaLabel":"Paper for xMAE: Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning (opens in new tab)"},{"href":"https://arxiv.org/abs/2605.00973","label":"arXiv","ariaLabel":"arXiv for xMAE: Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning (opens in new tab)"}]},{"slug":"udpad-wrist-ppg-imu-microgesture-2026","title":"μDPad: A Large-Scale Multimodal PPG and IMU Dataset for Wrist-Worn Microgesture Recognition","summary":"The authors built a wrist prototype: an Analog Devices MAX86177A PPG module with 8 photodiodes in a 2 cm circular layout on the dorsal wrist, using green + IR LEDs, plus a QMI8658 6-axis IMU, all sampled at 112 Hz. They recorded 66 participants (25 h; 9 h with hand video) performing thumb microgestures: 4 directional swipes, pinch close/open, side tap, and wrist rotation during a held pinch. The question is whether spatial PPG captures the tissue/tendon deformations under the watch.","venue":"ICMI 2026 (Proceedings of the 28th International Conference On Multimodal Interaction, Napoli, pp. 837–850)","year":2026,"open_code":false,"updated":"2026-10-07","url":"/research/papers/udpad-wrist-ppg-imu-microgesture-2026","links":[{"href":"https://doi.org/10.1145/3776574.3831197","label":"Paper","ariaLabel":"Paper for μDPad: A Large-Scale Multimodal PPG and IMU Dataset for Wrist-Worn Microgesture Recognition (opens in new tab)"}]},{"slug":"biosignal-fm-survey-barmpas-2025","title":"A Comprehensive Review of Biosignal Foundation Models","summary":"This survey by Na Lee, Konstantinos Barmpas and colleagues reviews foundation models for biosignals, including EEG, ECG, EMG, EOG and PPG. It explains what makes each signal hard to train large models on and organizes existing unimodal and multimodal work by data processing, feature extraction, architecture, pretraining approach and evaluation.","venue":"Preprint [survey]","year":2025,"open_code":null,"updated":null,"url":"/research/papers/biosignal-fm-survey-barmpas-2025","links":[{"href":"https://www.barmpas.com/publication/2025-10-05-survey_fm_biosignals","label":"Paper","ariaLabel":"Paper for A Comprehensive Review of Biosignal Foundation Models (opens in new tab)"}]},{"slug":"dscnet-ppg-glucose-sensors-2025","title":"A Deep Sparse Capsule Network for Non-Invasive Blood Glucose Level Estimation Using a PPG Sensor","summary":"This Sensors paper estimates blood glucose from PPG with a deep sparse capsule network (DSCNet), using data collected with a custom PPG sensor and Raspberry Pi module. Signals were cleaned with Savitzky-Golay and moving-average filters, and a version with self-attention was also tested. The model reported MAPE of 3.022, MARD of 10.81 and R² of 0.98.","venue":"Sensors (MDPI)","year":2025,"open_code":null,"updated":null,"url":"/research/papers/dscnet-ppg-glucose-sensors-2025","links":[{"href":"https://www.mdpi.com/1424-8220/25/6/1868","label":"Paper","ariaLabel":"Paper for A Deep Sparse Capsule Network for Non-Invasive Blood Glucose Level Estimation Using a PPG Sensor (opens in new tab)"}]},{"slug":"ai-ecg-afib-scoping-review-2025","title":"Artificial Intelligence Applied to Electrocardiograms Recorded in Sinus Rhythm for Detection and Prediction of Atrial Fibrillation: A Scoping Review","summary":"This scoping review covers 19 studies that used AI on ECGs recorded in normal sinus rhythm to find hidden atrial fibrillation or predict new AF. Models detected subclinical AF with AUROC 0.75 to 0.90 across ten studies and predicted long-term new AF with AUROC 0.69 to 0.85 across six. Eleven studies had robust external validation, and the authors note retrospective designs and limited calibration reporting as common weaknesses.","venue":"Medicina (MDPI) [survey]","year":2025,"open_code":null,"updated":null,"url":"/research/papers/ai-ecg-afib-scoping-review-2025","links":[{"href":"https://www.mdpi.com/1648-9144/62/1/199","label":"Paper","ariaLabel":"Paper for Artificial Intelligence Applied to Electrocardiograms Recorded in Sinus Rhythm for Detection and Prediction of Atrial Fibrillation: A Scoping Review (opens in new tab)"}]},{"slug":"apple-beyond-sensor-wbm-behavioral-fm-2025","title":"Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (WBM)","summary":"WBM is an Apple foundation model built on behavioral data from wearables, such as derived activity and health metrics, instead of raw sensor waveforms. It was trained on more than 2.5 billion hours of data from 162K people in the Apple Heart and Movement Study and evaluated on 57 health tasks. It did best on behavior-driven tasks like sleep prediction and improved further when combined with a PPG foundation model.","venue":"ICML 2025 (Apple Machine Learning Research; published July 2025)","year":2025,"open_code":null,"updated":null,"url":"/research/papers/apple-beyond-sensor-wbm-behavioral-fm-2025","links":[{"href":"https://machinelearning.apple.com/research/beyond-sensor","label":"Paper","ariaLabel":"Paper for Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (WBM) (opens in new tab)"}]},{"slug":"clef-ecg-clinical-contrastive-2026","title":"CLEF: Clinically-Guided Contrastive Learning for Electrocardiogram Foundation Models","summary":"Evaluated on 18 clinical classification/regression tasks across 7 held-out datasets, pretrained on 12-lead but tested on extracted lead-I (their proxy for single-lead wearable ECG). Code and pretrained weights released.","venue":"arXiv preprint (Nokia Bell Labs)","year":2025,"open_code":null,"updated":null,"url":"/research/papers/clef-ecg-clinical-contrastive-2026","links":[{"href":"https://arxiv.org/abs/2512.02180","label":"Paper","ariaLabel":"Paper for CLEF: Clinically-Guided Contrastive Learning for Electrocardiogram Foundation Models (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2512.02180","label":"DOI","ariaLabel":"DOI for CLEF: Clinically-Guided Contrastive Learning for Electrocardiogram Foundation Models (opens in new tab)"},{"href":"https://github.com/Nokia-Bell-Labs/ecg-foundation-model","label":"Code","ariaLabel":"Code for CLEF: Clinically-Guided Contrastive Learning for Electrocardiogram Foundation Models (opens in new tab)"}]},{"slug":"cnn-bilstm-attention-ppg-bp-2025","title":"Cuff-less blood pressure monitoring via PPG signals using a hybrid CNN-BiLSTM deep learning model with attention mechanism","summary":"This Scientific Reports paper estimates blood pressure from PPG with a hybrid network that combines CNN layers, bidirectional LSTM layers and attention. It was validated on 2,064 patients from the MIMIC-II database and reached mean absolute errors of 1.88 mmHg systolic and 1.34 mmHg diastolic in 5-fold cross-validation.","venue":"Scientific Reports","year":2025,"open_code":null,"updated":null,"url":"/research/papers/cnn-bilstm-attention-ppg-bp-2025","links":[{"href":"https://www.nature.com/articles/s41598-025-07087-2","label":"Paper","ariaLabel":"Paper for Cuff-less blood pressure monitoring via PPG signals using a hybrid CNN-BiLSTM deep learning model with attention mechanism (opens in new tab)"}]},{"slug":"deep-cnn-ppg-afib-bulut-2025","title":"Deep CNN-based detection of cardiac rhythm disorders using PPG signals from wearable devices","summary":"This PLOS ONE paper trains a 1D CNN to detect heart rhythm disorders from 30-second PPG segments, using ECG from the same University of Massachusetts Medical Center dataset as the reference. It reported 95.17% accuracy in separating normal sinus rhythm, atrial fibrillation and premature atrial contractions. In July 2026 the PLOS ONE editors issued an Expression of Concern about retracted references and authorship policy compliance, advising readers to interpret the article with caution.","venue":"PLOS ONE","year":2025,"open_code":null,"updated":null,"url":"/research/papers/deep-cnn-ppg-afib-bulut-2025","links":[{"href":"https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0314154","label":"Paper","ariaLabel":"Paper for Deep CNN-based detection of cardiac rhythm disorders using PPG signals from wearable devices (opens in new tab)"}]},{"slug":"mayo-ecg-ai-sleep-apnea-2025","title":"Deep neural network algorithm using the electrocardiogram for detection of obstructive sleep apnea","summary":"Mayo Clinic's cardiovascular AI group developed a deep neural network that detects obstructive sleep apnea (OSA) directly from 12-lead ECG results. The study analysed ECGs from 11,299 patients at Mayo Clinic who had undergone simultaneous sleep evaluations, of whom 7,000+ had confirmed OSA. The model achieves AUROC ~0.87, and importantly demonstrates superior detection specifically in women — a population that is historically underdiagnosed for OSA due to different symptom presentation.","venue":"JACC: Advances","year":2025,"open_code":null,"updated":null,"url":"/research/papers/mayo-ecg-ai-sleep-apnea-2025","links":[{"href":"https://www.mayoclinic.org/medical-professionals/cardiovascular-diseases/news/mayo-clinic-cardiovascular-medicine-artificial-intelligence-groups-researchers-develop-ai-ecg-to-detect-obstructive-sleep-apnea/mac-20594007","label":"Paper","ariaLabel":"Paper for Deep neural network algorithm using the electrocardiogram for detection of obstructive sleep apnea (opens in new tab)"}]},{"slug":"mcbp-net-beat-to-beat-bp-multimodal-cnn-lstm-jbhi-2025","title":"Dynamic Beat-to-Beat Measurements of Blood Pressure Using Multimodal Physiological Signals and a Hybrid CNN-LSTM Model","summary":"Hybrid CNN-LSTM: 2 convolution layers + 4 LSTM layers. Inputs: PPG, ECG, impedance plethysmography (IPG) and skin temperature (ST).","venue":"IEEE J. Biomed. Health Inform. 29(8):5438–5451","year":2025,"open_code":null,"updated":"2026-10-11","url":"/research/papers/mcbp-net-beat-to-beat-bp-multimodal-cnn-lstm-jbhi-2025","links":[{"href":"https://doi.org/10.1109/JBHI.2025.3548771","label":"Paper","ariaLabel":"Paper for Dynamic Beat-to-Beat Measurements of Blood Pressure Using Multimodal Physiological Signals and a Hybrid CNN-LSTM Model (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/40768436","label":"PubMed","ariaLabel":"PubMed for Dynamic Beat-to-Beat Measurements of Blood Pressure Using Multimodal Physiological Signals and a Hybrid CNN-LSTM Model (opens in new tab)"}]},{"slug":"galaxyppg-scientific-data-2025","title":"GalaxyPPG — PPG dataset in semi-naturalistic settings (Scientific Data 2025)","summary":"GalaxyPPG is a dataset from 24 participants with wrist PPG from a Samsung Galaxy Watch 5 and an Empatica E4, plus chest ECG from a Polar H10. It is meant for heart rate tracking during varied activity and HRV-based stress detection. The authors also released an open-source toolkit for collecting and analysing Galaxy Watch data.","venue":"Scientific Data","year":2025,"open_code":null,"updated":"2026-09-27","url":"/research/papers/galaxyppg-scientific-data-2025","links":[{"href":"https://doi.org/10.1038/s41597-025-05152-z","label":"Paper","ariaLabel":"Paper for GalaxyPPG — PPG dataset in semi-naturalistic settings (Scientific Data 2025) (opens in new tab)"},{"href":"https://github.com/Kaist-ICLab/GalaxyPPG-Supplementary-Code","label":"Code","ariaLabel":"Code for GalaxyPPG — PPG dataset in semi-naturalistic settings (Scientific Data 2025) (opens in new tab)"}]},{"slug":"generalist-vs-specialist-tsfm-ppg-kataria-2025","title":"Generalist vs Specialist Time Series Foundation Models: Investigating Potential Emergent Behaviors in Assessing Human Health Using PPG Signals","summary":"This paper compares generalist time-series foundation models, trained on data like weather and traffic (for example MOMENT), with specialist models trained on PPG. In full fine-tuning the specialist PPG model scored 27% higher on the authors' win score.","venue":"arXiv preprint","year":2025,"open_code":null,"updated":"2026-10-11","url":"/research/papers/generalist-vs-specialist-tsfm-ppg-kataria-2025","links":[{"href":"https://arxiv.org/abs/2510.14254","label":"Paper","ariaLabel":"Paper for Generalist vs Specialist Time Series Foundation Models: Investigating Potential Emergent Behaviors in Assessing Human Health Using PPG Signals (opens in new tab)"}]},{"slug":"generalizable-ppg-bp-benchmarking-2025","title":"Generalizable deep learning for photoplethysmography-based blood pressure estimation — A benchmarking study","summary":"This benchmarking study by Moulaeifard, Charlton and Strodthoff tests how well deep learning models that estimate blood pressure from PPG generalize to data unlike their training set. It was published open access in Machine Learning: Health (IOP Publishing) in September 2025.","venue":"Machine Learning: Health (IOP Publishing)","year":2025,"open_code":null,"updated":null,"url":"/research/papers/generalizable-ppg-bp-benchmarking-2025","links":[{"href":"https://iopscience.iop.org/article/10.1088/3049-477X/ae01a8/meta","label":"Paper","ariaLabel":"Paper for Generalizable deep learning for photoplethysmography-based blood pressure estimation — A benchmarking study (opens in new tab)"}]},{"slug":"gpt-ppg-gpt-foundation-model-ppg-physiolmeas-2025","title":"GPT-PPG: a GPT-based foundation model for photoplethysmography signals","summary":"GPT-PPG adapts the GPT architecture to continuous PPG signals, pretraining on more than 200 million 30-second samples and then fine-tuning for each task. It matched or beat earlier methods on several tasks, notably atrial fibrillation detection, and can denoise signals without extra fine-tuning.","venue":"Physiological Measurement","year":2025,"open_code":null,"updated":"2026-10-11","url":"/research/papers/gpt-ppg-gpt-foundation-model-ppg-physiolmeas-2025","links":[{"href":"https://doi.org/10.1088/1361-6579/add988","label":"Paper","ariaLabel":"Paper for GPT-PPG: a GPT-based foundation model for photoplethysmography signals (opens in new tab)"},{"href":"https://arxiv.org/abs/2503.08015","label":"arXiv","ariaLabel":"arXiv for GPT-PPG: a GPT-based foundation model for photoplethysmography signals (opens in new tab)"}]},{"slug":"lsm2-aim-incomplete-wearable","title":"LSM-2: Learning from Incomplete Wearable Sensor Data","summary":"LSM-2 is Google Research's second Large Sensor Model, trained with Adaptive and Inherited Masking (AIM) so it can learn directly from wearable data with gaps, without filling them in first. It was trained on 40 million hours of data from more than 60,000 people wearing Fitbit and Google Pixel devices between March and May 2024. The model handles classification, regression and generative tasks, such as predicting hypertension, from fragmented sensor data.","venue":"Google Research blog (2025-07-22); arXiv preprint (submitted 2025-06-05)","year":2025,"open_code":null,"updated":null,"url":"/research/papers/lsm2-aim-incomplete-wearable","links":[{"href":"https://research.google/blog/lsm-2-learning-from-incomplete-wearable-sensor-data/","label":"Paper","ariaLabel":"Paper for LSM-2: Learning from Incomplete Wearable Sensor Data (opens in new tab)"}]},{"slug":"protomm-pulse-motion-prototypes-2026","title":"ProtoMM: Shared Prototypes for a Multimodal Pulse-Motion Foundation Model","summary":"ProtoMM is a self-supervised method that uses a shared dictionary of prototypes to place different sensor modalities in one embedding space. The authors used it to build a pulse-motion foundation model that beat contrastive-only and earlier multimodal methods, and the prototypes help interpret what the model learned.","venue":"arXiv preprint","year":2025,"open_code":null,"updated":"2026-10-11","url":"/research/papers/protomm-pulse-motion-prototypes-2026","links":[{"href":"https://arxiv.org/abs/2510.09764","label":"Paper","ariaLabel":"Paper for ProtoMM: Shared Prototypes for a Multimodal Pulse-Motion Foundation Model (opens in new tab)"}]},{"slug":"pulse-ppg","title":"Pulse-PPG: An Open-Source Field-Trained PPG Foundation Model for Wearable Applications Across Lab and Field Settings","summary":"Pulse-PPG is an open-source PPG foundation model trained only on raw PPG from a 100-day field study with 120 participants. Earlier open PPG models were trained on clinical data, and field-trained ones were closed. Pulse-PPG generalized better than clinically trained models across clinical and mobile health tasks in both lab and field settings.","venue":"Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies","year":2025,"open_code":null,"updated":"2026-10-11","url":"/research/papers/pulse-ppg","links":[{"href":"https://arxiv.org/abs/2502.01108","label":"Paper","ariaLabel":"Paper for Pulse-PPG: An Open-Source Field-Trained PPG Foundation Model for Wearable Applications Across Lab and Field Settings (opens in new tab)"},{"href":"https://doi.org/10.1145/3749494","label":"DOI","ariaLabel":"DOI for Pulse-PPG: An Open-Source Field-Trained PPG Foundation Model for Wearable Applications Across Lab and Field Settings (opens in new tab)"}]},{"slug":"sensorlm-wearable-sensor-language-model","title":"SensorLM: Learning the Language of Wearable Sensors","summary":"SensorLM is a family of sensor-language foundation models that describe wearable data in natural language. A hierarchical captioning pipeline generates statistical, structural and semantic descriptions of the sensor data. The models show scaling behaviour, label efficiency, sensor captioning and zero-shot transfer to new tasks.","venue":"arXiv preprint","year":2025,"open_code":null,"updated":null,"url":"/research/papers/sensorlm-wearable-sensor-language-model","links":[{"href":"https://arxiv.org/abs/2506.09108","label":"Paper","ariaLabel":"Paper for SensorLM: Learning the Language of Wearable Sensors (opens in new tab)"}]},{"slug":"text2imu-bsn-2025","title":"Text2IMU: Advancing Human Activity Recognition by Text-Driven IMU Data Synthesis","summary":"Freiburg (Amft lab) introduces Text2IMU: text prompts → motion-synthesis model → multiple body surface morphologies → virtual wearable accelerometer + gyroscope at configurable placements, then train HAR solely on synthetic data (no measured target-domain IMU required). OpenReview abstract reports average HAR accuracy 79.2% on correctly synthesised activities across three public benchmarks — roughly double IMUGPT-class baselines in their comparison. Precursor to the OJEMB 2026 diversity study (text2imu-diversity-sources-ojemb-2026, best synthetic-only bal.","venue":"2025 IEEE 21st International Conference on Body Sensor Networks (BSN)","year":2025,"open_code":null,"updated":"2026-09-28","url":"/research/papers/text2imu-bsn-2025","links":[{"href":"https://ieeexplore.ieee.org/document/11337798","label":"Paper","ariaLabel":"Paper for Text2IMU: Advancing Human Activity Recognition by Text-Driven IMU Data Synthesis (opens in new tab)"},{"href":"https://doi.org/10.1109/BSN66969.2025.11337798","label":"DOI","ariaLabel":"DOI for Text2IMU: Advancing Human Activity Recognition by Text-Driven IMU Data Synthesis (opens in new tab)"}]},{"slug":"zara-training-free-motion-reasoning-llm-agents-acl-2026","title":"ZARA: Training-Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents","summary":"ZARA is an LLM agent framework that reasons about motion sensor data without any training, using domain knowledge and retrieval. On eight benchmarks it generalized well to unseen people and across datasets.","venue":"arXiv preprint","year":2025,"open_code":true,"updated":"2026-10-12","url":"/research/papers/zara-training-free-motion-reasoning-llm-agents-acl-2026","links":[{"href":"https://arxiv.org/abs/2508.04038","label":"arXiv","ariaLabel":"arXiv for ZARA: Training-Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents (opens in new tab)"}]},{"slug":"ppg-dl-scoping-review-nie-2024","title":"A Scoping Review of Deep Learning Methods for Photoplethysmography Data","summary":"This scoping review covers deep learning methods for PPG data. It finds that deep learning generally beats handcrafted-feature approaches and is more flexible, but notes ongoing problems: few large high-quality datasets, little real-world validation, and concerns about interpretability, scalability and efficiency.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/ppg-dl-scoping-review-nie-2024","links":[{"href":"https://arxiv.org/abs/2401.12783","label":"Paper","ariaLabel":"Paper for A Scoping Review of Deep Learning Methods for Photoplethysmography Data (opens in new tab)"}]},{"slug":"kebede-lca-accelerometry-scoping-review-plos-one-2024","title":"A systematic scoping review of latent class analysis applied to accelerometry-assessed physical activity and sedentary behavior","summary":"This systematic scoping review examined how latent class analysis (LCA) has been applied to accelerometry-assessed physical activity and sedentary behaviour in studies published through 2021. Of 2,555 papers found, 12 papers from 8 unique studies (sample sizes 217-7,931) met the inclusion criteria. All 8 studies used physical activity measures and 75% used sedentary behaviour measures, while the accelerometer variables fed into the LCA models varied widely between studies.","venue":"PLOS ONE","year":2024,"open_code":null,"updated":"2026-10-15","url":"/research/papers/kebede-lca-accelerometry-scoping-review-plos-one-2024","links":[{"href":"https://doi.org/10.1371/journal.pone.0283884","label":"Paper","ariaLabel":"Paper for A systematic scoping review of latent class analysis applied to accelerometry-assessed physical activity and sedentary behavior (opens in new tab)"}]},{"slug":"afppgecg-triggers-physiol-meas-2024","title":"AF triggers in week-long ECG+wrist PPG (relational strength) — Pluščiauskaitė 2024","summary":"This study looked at whether triggers detected in wearable signals line up with atrial fibrillation episodes. Thirty-five people with paroxysmal AF wore a chest ECG patch and a wrist PPG device for a week of normal life, and the patch data were used to detect physical exertion, stress, lying on the left side and sleep disturbance. Physical exertion was linked to the biggest rise in AF for the most patients, and PPG-detected AF tracked annotated AF with r = 0.62.","venue":"Physiological Measurement","year":2024,"open_code":null,"updated":"2026-09-27","url":"/research/papers/afppgecg-triggers-physiol-meas-2024","links":[{"href":"https://doi.org/10.1088/1361-6579/ad79b3","label":"Paper","ariaLabel":"Paper for AF triggers in week-long ECG+wrist PPG (relational strength) — Pluščiauskaitė 2024 (opens in new tab)"}]},{"slug":"mayo-ecg-ai-lvef-2024","title":"Artificial Intelligence–Enhanced Electrocardiography Identifies Patients With Normal Ejection Fraction at Risk of Worse Outcomes","summary":"This Mayo Clinic study followed 100,586 patients who had both an echocardiogram and an ECG, focusing on those whose AI-ECG score flagged low ejection fraction even though their measured EF was normal. These false-positive patients had an echocardiographic abnormality 97% of the time. Over a median of 2.7 years they had higher mortality than true negatives (age- and sex-adjusted HR 1.64), especially when their echo was also abnormal.","venue":"JACC: Advances","year":2024,"open_code":null,"updated":null,"url":"/research/papers/mayo-ecg-ai-lvef-2024","links":[{"href":"https://www.jacc.org/doi/10.1016/j.jacadv.2024.101179","label":"Paper","ariaLabel":"Paper for Artificial Intelligence–Enhanced Electrocardiography Identifies Patients With Normal Ejection Fraction at Risk of Worse Outcomes (opens in new tab)"}]},{"slug":"capture-24-dataset","title":"CAPTURE-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition","summary":"CAPTURE-24 is a free-living wrist accelerometer dataset from 151 participants, recorded over more than 24 hours each together with wearable cameras and sleep diaries. It has 3,883 hours of accelerometer data, 2,562 of them annotated, which the authors say is two to three orders of magnitude larger than earlier public datasets. It was published in Scientific Data to support activity recognition models that work outside the lab.","venue":"Scientific Data","year":2024,"open_code":null,"updated":"2026-09-27","url":"/research/papers/capture-24-dataset","links":[{"href":"https://www.nature.com/articles/s41597-024-03960-3","label":"Paper","ariaLabel":"Paper for CAPTURE-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition (opens in new tab)"}]},{"slug":"capture24","title":"CAPTURE-24: Free-Living Wrist Accelerometer Dataset","summary":"CAPTURE-24 is a large free-living activity dataset from 151 participants who wore a wrist accelerometer and a wearable camera and kept a sleep diary for over 24 hours. It contains 3,883 hours of accelerometer data, 2,562 of them annotated. It was made because lab datasets lack the mixed, nuanced movement of daily life.","venue":"Scientific Data","year":2024,"open_code":null,"updated":"2026-09-27","url":"/research/papers/capture24","links":[{"href":"https://doi.org/10.1038/s41597-024-03960-3","label":"DOI","ariaLabel":"DOI for CAPTURE-24: Free-Living Wrist Accelerometer Dataset (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/39414802","label":"PubMed","ariaLabel":"PubMed for CAPTURE-24: Free-Living Wrist Accelerometer Dataset (opens in new tab)"}]},{"slug":"crosshar","title":"CrossHAR: Generalizing Cross-dataset Human Activity Recognition via Hierarchical Self-Supervised Pretraining","summary":"CrossHAR (UbiComp/IMWUT 2024) addresses the cross-dataset generalization problem in wearable HAR: models trained on one dataset (specific device, placement, user population) frequently fail to transfer to unseen target datasets. CrossHAR introduces a hierarchical self-supervised pretraining strategy that explicitly learns domain-invariant features.","venue":"IMWUT/UbiComp 2024","year":2024,"open_code":null,"updated":null,"url":"/research/papers/crosshar","links":[{"href":"https://dl.acm.org/doi/10.1145/3659597","label":"Paper","ariaLabel":"Paper for CrossHAR: Generalizing Cross-dataset Human Activity Recognition via Hierarchical Self-Supervised Pretraining (opens in new tab)"}]},{"slug":"ecg-fm-open","title":"ECG-FM: An Open Electrocardiogram Foundation Model","summary":"ECG-FM is an open-weight foundation model for ECG analysis, studied on a dataset of 1.5 million ECGs. It was tested on predicting reduced left ventricular ejection fraction and ECG interpretation labels, and the authors released a benchmark task on MIMIC-IV-ECG. It clearly beat task-specific models when labelled data were limited, with AUROCs of 0.996 for atrial fibrillation and 0.929 for LVEF of 40% or less.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/ecg-fm-open","links":[{"href":"https://arxiv.org/abs/2408.05178","label":"Paper","ariaLabel":"Paper for ECG-FM: An Open Electrocardiogram Foundation Model (opens in new tab)"}]},{"slug":"ecgfounder","title":"ECGFounder: An ECG Foundation Model Built on 10+ Million Recordings","summary":"ECGFounder is a general-purpose ECG foundation model trained on more than 10 million recordings with real-world annotations from cardiology experts. On held-out validation it reached expert-level performance, with AUROC above 0.95 for 80 diagnoses.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/ecgfounder","links":[{"href":"https://arxiv.org/abs/2410.04133","label":"Paper","ariaLabel":"Paper for ECGFounder: An ECG Foundation Model Built on 10+ Million Recordings (opens in new tab)"}]},{"slug":"viswanath-five-million-nights-sleep-phenotypes-npjdm-2024","title":"Five million nights: temporal dynamics in human sleep phenotypes","summary":"Using five million nights of wearable sleep data, Viswanath and colleagues tested whether an individual's sleep phenotype changes over time and whether those changes carry information about acute illness such as flu, fever or COVID-19. They found 13 sleep phenotypes associated with sleep quality, with individuals transitioning between them over time. Transition patterns differed between people with and without chronic conditions and within people before versus during acute illness, and carried 2-10x more information than phenotype membership alone.","venue":"npj Digital Medicine","year":2024,"open_code":null,"updated":"2026-10-15","url":"/research/papers/viswanath-five-million-nights-sleep-phenotypes-npjdm-2024","links":[{"href":"https://doi.org/10.1038/s41746-024-01125-5","label":"Paper","ariaLabel":"Paper for Five million nights: temporal dynamics in human sleep phenotypes (opens in new tab)"}]},{"slug":"hubert-ecg","title":"HuBERT-ECG: A Self-Supervised Foundation Model for Broad and Scalable Cardiac Applications","summary":"HuBERT-ECG is a self-supervised ECG foundation model pretrained on 9.1 million 12-lead ECGs covering 164 cardiovascular conditions. It can be fine-tuned for new tasks by adding one output layer and reached AUROCs from 84.3% in low-data settings to 99% in large-scale ones. It predicted death within two years with AUROC 93.4%, and the models and code are released.","venue":"medRxiv 2024","year":2024,"open_code":null,"updated":"2026-10-05","url":"/research/papers/hubert-ecg","links":[{"href":"https://www.medrxiv.org/content/10.1101/2024.11.14.24317328v2","label":"Paper","ariaLabel":"Paper for HuBERT-ECG: A Self-Supervised Foundation Model for Broad and Scalable Cardiac Applications (opens in new tab)"}]},{"slug":"imugpt-2-language-cross-modality-har-2024","title":"IMUGPT 2.0: Language-Based Cross Modality Transfer for Sensor-Based Human Activity Recognition","summary":"Georgia Tech / Plötz lab practical upgrade of IMUGPT (Leng et al. ISWC 2023 best-paper honorable mention): LLM generates diverse activity text → text-to-motion (e.g. T2M-GPT) → virtual IMU (IMUTube / IMUSim backend) for HAR training.","venue":"Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT)","year":2024,"open_code":null,"updated":"2026-09-28","url":"/research/papers/imugpt-2-language-cross-modality-har-2024","links":[{"href":"https://arxiv.org/abs/2402.01049","label":"Paper","ariaLabel":"Paper for IMUGPT 2.0: Language-Based Cross Modality Transfer for Sensor-Based Human Activity Recognition (opens in new tab)"},{"href":"https://doi.org/10.1145/3678545","label":"DOI","ariaLabel":"DOI for IMUGPT 2.0: Language-Based Cross Modality Transfer for Sensor-Based Human Activity Recognition (opens in new tab)"},{"href":"https://github.com/ZikangLeng/IMUGPT","label":"Code","ariaLabel":"Code for IMUGPT 2.0: Language-Based Cross Modality Transfer for Sensor-Based Human Activity Recognition (opens in new tab)"}]},{"slug":"cnn-gru-ppg-glucose-2024","title":"Non-Invasive Glucose Level Monitoring from PPG using a Hybrid CNN-GRU Deep Learning Network","summary":"This paper estimates blood glucose from PPG with a hybrid network that uses a CNN for patterns within the signal and a GRU for patterns over time. On its test set it reported a mean absolute error of 2.96 mg/dL, RMSE of 3.94 mg/dL and R² of 0.97, with all points in Clarke Error Grid zone A.","venue":"arXiv","year":2024,"open_code":null,"updated":null,"url":"/research/papers/cnn-gru-ppg-glucose-2024","links":[{"href":"https://arxiv.org/abs/2411.11094","label":"Paper","ariaLabel":"Paper for Non-Invasive Glucose Level Monitoring from PPG using a Hybrid CNN-GRU Deep Learning Network (opens in new tab)"}]},{"slug":"normwear","title":"NormWear: Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals","summary":"NormWear is a foundation model for multivariate wearable sensing that works across different sensor setups. A channel-aware attention mechanism with a shared [CLS] token captures patterns within and between sensors, and it was pretrained on public PPG, ECG, EEG, GSR and IMU data. Across 11 public datasets and 18 applications in mental health, body state, vital signs and disease risk, it beat competitive baselines in zero-shot, partial-shot and full-shot settings.","venue":"IMWUT/UbiComp 2025","year":2024,"open_code":null,"updated":null,"url":"/research/papers/normwear","links":[{"href":"https://arxiv.org/abs/2412.09758","label":"Paper","ariaLabel":"Paper for NormWear: Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals (opens in new tab)"}]},{"slug":"nas-ppg-bp-wearable-burrello-2024","title":"Optimization and Deployment of Deep Neural Networks for PPG-based Blood Pressure Estimation Targeting Low-power Wearables","summary":"This paper optimizes deep networks for estimating blood pressure from PPG so they fit low-power wearables such as the GAP8 chip. Starting from top models on four public datasets, the optimized versions had up to 4.99% lower error or 73.36% smaller size at the same error. The most accurate one used just 0.37 mJ per estimate and reached a diastolic MAE of 8.08.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/nas-ppg-bp-wearable-burrello-2024","links":[{"href":"https://arxiv.org/abs/2409.07485","label":"Paper","ariaLabel":"Paper for Optimization and Deployment of Deep Neural Networks for PPG-based Blood Pressure Estimation Targeting Low-power Wearables (opens in new tab)"}]},{"slug":"papagei-ppg-fm","title":"PaPaGei: Open Foundation Models for Optical Physiological Signals","summary":"PaPaGei is an open foundation model for PPG, pretrained on more than 57,000 hours (20 million segments) of public PPG data. Its training uses knowledge of how PPG waveform shape varies between people. It improved classification by 6.3% and regression by 2.9% across at least 14 tasks.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":"2026-10-12","url":"/research/papers/papagei-ppg-fm","links":[{"href":"https://arxiv.org/abs/2410.20542","label":"Paper","ariaLabel":"Paper for PaPaGei: Open Foundation Models for Optical Physiological Signals (opens in new tab)"}]},{"slug":"ppg-cvd-risk-google-2024","title":"Predicting cardiovascular disease risk using photoplethysmography and deep learning","summary":"This Google study built a deep learning risk score that predicts 10-year major cardiovascular events from fingertip PPG plus age, sex and smoking status, aimed at low-resource settings. Trained on 141,509 UK Biobank participants and tested on 54,856 others, its C-statistic of 71.1% was non-inferior to an office-based WHO-style score that needs blood pressure, height and weight (70.9%). Adding its features to that score raised the C-statistic by 1.0%.","venue":"PLOS Global Public Health","year":2024,"open_code":null,"updated":null,"url":"/research/papers/ppg-cvd-risk-google-2024","links":[{"href":"https://journals.plos.org/globalpublichealth/article?id=10.1371/journal.pgph.0003204","label":"Paper","ariaLabel":"Paper for Predicting cardiovascular disease risk using photoplethysmography and deep learning (opens in new tab)"}]},{"slug":"pyppg","title":"pyPPG: A Wearable PPG Analysis Toolbox","summary":"pyPPG is an open-source Python toolbox by Goda, Charlton and Behar for analysing long-term finger PPG recordings. It loads .mat, .txt, .csv and .edf files, filters the signal, detects beats and fiducial points, and extracts standard pulse-wave biomarkers. Its algorithms were validated on freely available PPG databases, and it is described in Physiological Measurement.","venue":"Open-source toolbox (Physiological Measurement 2024)","year":2024,"open_code":null,"updated":null,"url":"/research/papers/pyppg","links":[{"href":"https://github.com/godamartonaron/GODA_pyPPG","label":"Paper","ariaLabel":"Paper for pyPPG: A Wearable PPG Analysis Toolbox (opens in new tab)"}]},{"slug":"relcon","title":"RelCon: Relative Contrastive Learning for a Motion Foundation Model","summary":"RelCon is a self-supervised method for training a motion foundation model from wearable accelerometers. It first learns a distance measure that captures motif similarity and rotation invariance, then trains the model on relative similarity across time and people. Trained on 1 billion segments from 87,376 participants, it set state-of-the-art results on tasks including activity recognition and gait metrics.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/relcon","links":[{"href":"https://doi.org/10.48550/arXiv.2411.18822","label":"DOI","ariaLabel":"DOI for RelCon: Relative Contrastive Learning for a Motion Foundation Model (opens in new tab)"},{"href":"https://arxiv.org/abs/2411.18822","label":"arXiv","ariaLabel":"arXiv for RelCon: Relative Contrastive Learning for a Motion Foundation Model (opens in new tab)"}]},{"slug":"relcon-motion-fm","title":"RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data","summary":"RelCon is a self-supervised relative contrastive method for training a motion foundation model on wrist accelerometer data, published at ICLR 2025. It first learns a distance measure that captures motif similarity and rotation invariance, then trains the model on relative similarity across time and between people. Trained on 1 billion segments from 87,376 participants, it set state-of-the-art results on tasks including activity recognition and gait metric regression.","venue":"ICLR 2025","year":2024,"open_code":null,"updated":null,"url":"/research/papers/relcon-motion-fm","links":[{"href":"https://arxiv.org/abs/2411.18822","label":"Paper","ariaLabel":"Paper for RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data (opens in new tab)"}]},{"slug":"lsm-scaling-wearable-fm","title":"Scaling Wearable Foundation Models (LSM)","summary":"LSM (Large Sensor Model) is a Google study of how wearable foundation models scale with compute, data and model size. It used up to 40 million hours of per-minute heart rate, HRV, electrodermal activity, accelerometer, skin temperature and altimeter data from more than 165,000 people. The authors report scaling laws for imputation, interpolation and extrapolation, and show sample-efficient learning for exercise and activity recognition.","venue":"ICLR 2025","year":2024,"open_code":null,"updated":null,"url":"/research/papers/lsm-scaling-wearable-fm","links":[{"href":"https://arxiv.org/abs/2410.13638","label":"Paper","ariaLabel":"Paper for Scaling Wearable Foundation Models (LSM) (opens in new tab)"}]},{"slug":"ssl-har-700k-ukbiobank","title":"Self-supervised learning for human activity recognition using 700,000 person-days of wearable data","summary":"This Oxford study used self-supervised learning on 700,000 person-days of unlabelled wrist accelerometer data from UK Biobank to build activity recognition models. The models beat strong baselines on eight benchmark datasets, with relative F1 gains of 2.5% to 130.9% (median 24.4%), and generalized across cohorts, living environments and devices. The pretrained models are open source.","venue":"npj Digital Medicine","year":2024,"open_code":null,"updated":null,"url":"/research/papers/ssl-har-700k-ukbiobank","links":[{"href":"https://www.nature.com/articles/s41746-024-01062-3","label":"Paper","ariaLabel":"Paper for Self-supervised learning for human activity recognition using 700,000 person-days of wearable data (opens in new tab)"}]},{"slug":"siamquality-ppg-fm","title":"SiamQuality: A ConvNet-Based Foundation Model for Imperfect Physiological Signals","summary":"SiamQuality is a convolutional foundation model for physiological signals built to cope with noisy, incomplete real-world data, demonstrated on PPG. The authors present it as a toolset for building foundation models on imperfect physiological data.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":"2026-10-11","url":"/research/papers/siamquality-ppg-fm","links":[{"href":"https://arxiv.org/abs/2404.17667","label":"Paper","ariaLabel":"Paper for SiamQuality: A ConvNet-Based Foundation Model for Imperfect Physiological Signals (opens in new tab)"}]},{"slug":"sleepfm","title":"SleepFM: Multi-modal Representation Learning for Sleep Across Brain Activity, ECG and Respiratory Signals","summary":"SleepFM is a multimodal sleep foundation model trained on more than 100,000 hours of polysomnography from over 14,000 people, covering brain, heart and breathing signals. A leave-one-out contrastive approach beat standard pairwise contrastive learning. Simple models on its embeddings beat an end-to-end CNN on sleep staging (macro AUROC 0.88 vs 0.72).","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/sleepfm","links":[{"href":"https://arxiv.org/abs/2405.17766","label":"Paper","ariaLabel":"Paper for SleepFM: Multi-modal Representation Learning for Sleep Across Brain Activity, ECG and Respiratory Signals (opens in new tab)"}]},{"slug":"squwa-signal-quality-aware-af-dnn-2024","title":"SQUWA: Signal Quality Aware DNN Architecture for Enhanced Accuracy in Atrial Fibrillation Detection from Noisy PPG Signals","summary":"Atrial fibrillation (AF), a common cardiac arrhythmia, significantly increases the risk of stroke, heart disease, and mortality. Photoplethysmography (PPG) offers a promising solution for continuous AF monitoring, due to its cost efficiency and integration into wearable devices. Nonetheless, PPG signals are susceptible to corruption from motion artifacts and other factors often encountered in ambulatory settings.","venue":"arXiv preprint (eess.SP, cs.AI, cs.LG) · 15 pages; 9 figures; 2024 Conference on Health, Inference, and Learning (CHIL)","year":2024,"open_code":null,"updated":"2026-10-11","url":"/research/papers/squwa-signal-quality-aware-af-dnn-2024","links":[{"href":"https://arxiv.org/abs/2404.15353","label":"Paper","ariaLabel":"Paper for SQUWA: Signal Quality Aware DNN Architecture for Enhanced Accuracy in Atrial Fibrillation Detection from Noisy PPG Signals (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2404.15353","label":"DOI","ariaLabel":"DOI for SQUWA: Signal Quality Aware DNN Architecture for Enhanced Accuracy in Atrial Fibrillation Detection from Noisy PPG Signals (opens in new tab)"}]},{"slug":"synhar-ieee-access-2024","title":"SynHAR: Augmenting Human Activity Recognition With Synthetic Inertial Sensor Data Generated From Human Surface Models","summary":"Amft-lab SynHAR synthesises IMU time series from combined personalised biomechanical dynamics models and human surface models, then uses that synthetic ACC/gyro to augment HAR training (not synthetic-only foundation pretrain). Two scenarios: (1) fuse synthetic + measured IMU for direct HAR training; (2) pretrain on synthetic+measured then transfer-learn on public benchmarks. Low-data configs (fewer participants / IMUs) are stress-tested.","venue":"IEEE Access","year":2024,"open_code":null,"updated":"2026-09-28","url":"/research/papers/synhar-ieee-access-2024","links":[{"href":"https://ieeexplore.ieee.org/document/10786206","label":"Paper","ariaLabel":"Paper for SynHAR: Augmenting Human Activity Recognition With Synthetic Inertial Sensor Data Generated From Human Surface Models (opens in new tab)"},{"href":"https://doi.org/10.1109/ACCESS.2024.3513477","label":"DOI","ariaLabel":"DOI for SynHAR: Augmenting Human Activity Recognition With Synthetic Inertial Sensor Data Generated From Human Surface Models (opens in new tab)"}]},{"slug":"unimts","title":"UniMTS: Unified Pre-training for Motion Time Series","summary":"UniMTS is a unified pretraining method for motion time series from phones and smartwatches that generalizes across device location, orientation and activity. On 18 benchmarks it beat the best baselines by 340% zero-shot, 16.3% few-shot and 9.2% full-shot.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/unimts","links":[{"href":"https://arxiv.org/abs/2410.19818","label":"Paper","ariaLabel":"Paper for UniMTS: Unified Pre-training for Motion Time Series (opens in new tab)"}]},{"slug":"apple-accel-fm-kd","title":"Wearable Accelerometer Foundation Models for Health via Knowledge Distillation","summary":"This Apple paper shows that an accelerometer foundation model can predict many health targets when it learns from a PPG model through knowledge distillation. Training used 20 million minutes of unlabelled data from about 172K people in the Apple Heart and Movement Study. Accelerometer embeddings retrieved the matching PPG embedding with 99.2% top-1 accuracy, and distilled encoders beat accelerometer-only ones by 23% to 49%.","venue":"arXiv preprint","year":2024,"open_code":null,"updated":null,"url":"/research/papers/apple-accel-fm-kd","links":[{"href":"https://arxiv.org/abs/2412.11276","label":"Paper","ariaLabel":"Paper for Wearable Accelerometer Foundation Models for Health via Knowledge Distillation (opens in new tab)"}]},{"slug":"smartphone-ppg-hydration-ramadan-arxiv-2024","title":"You can monitor your hydration level using your smartphone camera","summary":"This preprint proposes checking hydration with an ordinary smartphone: the user records a short fingertip video, and a PPG signal extracted from it is classified into hydration levels. The authors built a multi-session dataset from 25 people fasting during Ramadan 2023 and trained machine learning, deep learning and transformer models for a hydrated / not hydrated task and a four-level task (fully hydrated to extremely dehydrated). They report accuracies of 95% to 99%, and explain the deep model's decisions with SHAP. The cohort is small and the dataset is not public.","venue":"arXiv preprint (arXiv:2402.07467)","year":2024,"open_code":null,"updated":"2026-10-13","url":"/research/papers/smartphone-ppg-hydration-ramadan-arxiv-2024","links":[]},{"slug":"apple-wearable-biosignals-fm","title":"Large-scale Training of Foundation Models for Wearable Biosignals","summary":"This Apple paper trains foundation models on large amounts of wearable PPG and ECG. It starts from the problem that labelled medical data are scarce and small, which limits new digital biomarkers, and uses large unlabelled wearable data to get around it.","venue":"arXiv preprint","year":2023,"open_code":null,"updated":"2026-09-11","url":"/research/papers/apple-wearable-biosignals-fm","links":[{"href":"https://arxiv.org/abs/2312.05409","label":"Paper","ariaLabel":"Paper for Large-scale Training of Foundation Models for Wearable Biosignals (opens in new tab)"}]},{"slug":"samsung-multispectral-smartwatch-skin-hydration-sci-rep-2023","title":"Multispectral sensor fusion in SmartWatch for skin hydration and body sweat loss","summary":"This Scientific Reports paper describes a smartwatch PPG prototype that adds two infrared light sources, at 970 nm and 1450 nm, to the usual visible-light sensor so it can track skin water content and sweat loss continuously. The authors modelled how light at each wavelength travels through dry and wet skin using Monte Carlo simulations, and checked those simulations against lab spectrometer and ultrasound measurements from four volunteers. They then analysed a user study of 19 adults (aged 21 to 52) who completed 103 indoor 5 km treadmill runs wearing the prototype, using machine learning (LightGBM) on the multi-wavelength PPG signals to track hydration changes. The work is a sensing-technology and modelling pipeline rather than a validated clinical hydration measure.","venue":"Scientific Reports","year":2023,"open_code":null,"updated":"2026-10-13","url":"/research/papers/samsung-multispectral-smartwatch-skin-hydration-sci-rep-2023","links":[{"href":"https://doi.org/10.1038/s41598-023-40339-7","label":"DOI","ariaLabel":"DOI for Multispectral sensor fusion in SmartWatch for skin hydration and body sweat loss (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/37591885","label":"PubMed","ariaLabel":"PubMed for Multispectral sensor fusion in SmartWatch for skin hydration and body sweat loss (opens in new tab)"}]},{"slug":"pulsedb-dataset-2023","title":"PulseDB: A large, cleaned dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood pressure estimation methods","summary":"PulseDB is a cleaned dataset for benchmarking cuffless blood pressure estimation, built from the MIMIC-III matched waveform subset and VitalDB. It contains 5,245,454 ten-second segments of ECG, PPG and arterial blood pressure from 5,361 subjects, plus demographics and the positions of ECG and PPG landmark points. The authors also used it to compare calibration-based and calibration-free testing.","venue":"Frontiers in Digital Health [dataset]","year":2023,"open_code":null,"updated":null,"url":"/research/papers/pulsedb-dataset-2023","links":[{"href":"https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2022.1090854/full","label":"Paper","ariaLabel":"Paper for PulseDB: A large, cleaned dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood pressure estimation methods (opens in new tab)"}]},{"slug":"windred-sleep-regularity-mortality-ukb-sleep-2024","title":"Sleep regularity is a stronger predictor of mortality risk than sleep duration: A prospective cohort study","summary":"Windred and colleagues calculated Sleep Regularity Index scores from more than 10 million hours of accelerometer data in 60,977 UK Biobank participants (median SRI 81.0) and compared sleep regularity with sleep duration as predictors of mortality. Across the top four SRI quintiles versus the least regular, higher regularity was associated with 20-48% lower all-cause, 16-39% lower cancer and 22-57% lower cardiometabolic mortality. Sleep regularity was a stronger predictor of all-cause mortality than sleep duration.","venue":"SLEEP","year":2023,"open_code":null,"updated":"2026-10-15","url":"/research/papers/windred-sleep-regularity-mortality-ukb-sleep-2024","links":[{"href":"https://doi.org/10.1093/sleep/zsad253","label":"Paper","ariaLabel":"Paper for Sleep regularity is a stronger predictor of mortality risk than sleep duration: A prospective cohort study (opens in new tab)"}]},{"slug":"spectrophon-dbm-wearable-hydration-plos-one-2022","title":"An accurate wearable hydration sensor: real-world evaluation of practical use (SpectroPhon DBM)","summary":"This study tested a wearable hydration sensor that combines PPG and galvanic sensors in a field trial with 240 participants, half men and half women. Device readings went to a phone over Bluetooth every 20 seconds and were compared with change in body mass on a medical scale. Agreement was close, with an average normalized error of about 2%, and fewer than 5% of values fell outside the Bland-Altman limits.","venue":"PLOS ONE","year":2022,"open_code":null,"updated":"2026-10-13","url":"/research/papers/spectrophon-dbm-wearable-hydration-plos-one-2022","links":[{"href":"https://doi.org/10.1371/journal.pone.0272646","label":"DOI","ariaLabel":"DOI for An accurate wearable hydration sensor: real-world evaluation of practical use (SpectroPhon DBM) (opens in new tab)"}]},{"slug":"pd-breathing-parkinsons-nocturnal","title":"Artificial Intelligence-Enabled Detection and Assessment of Parkinson's Disease using Nocturnal Breathing Signals","summary":"This Nature Medicine study built an AI model that detects Parkinson's disease and tracks its progression from breathing signals recorded during sleep. Evaluated on 7,671 people from several US hospitals and public datasets, it detected the disease with AUC 0.90 on held-out data and 0.85 on external data, and its severity estimates matched the MDS-UPDRS rating scale (R = 0.94). Breathing can be captured without contact at home from radio waves reflected off the body.","venue":"Nature Medicine 2022","year":2022,"open_code":null,"updated":null,"url":"/research/papers/pd-breathing-parkinsons-nocturnal","links":[{"href":"https://www.nature.com/articles/s41591-022-01932-x","label":"Paper","ariaLabel":"Paper for Artificial Intelligence-Enabled Detection and Assessment of Parkinson's Disease using Nocturnal Breathing Signals (opens in new tab)"}]},{"slug":"charlton-ppg-vascular-age-vascagenet-ajp-2022","title":"Assessing hemodynamics from the photoplethysmogram to gain insights into vascular age: a review from VascAgeNet","summary":"The photoplethysmogram (PPG) signal is widely measured by clinical and consumer devices, and it is emerging as a potential tool for assessing vascular age. The shape and timing of the PPG pulse wave are both influenced by normal vascular aging, changes in arterial stiffness and blood pressure, and atherosclerosis. Evidence is then presented on the performance, repeatability and reproducibility, and clinical utility of PPG-derived parameters of vascular age.","venue":"Am J Physiol Heart Circ Physiol 2022","year":2022,"open_code":null,"updated":"2026-10-15","url":"/research/papers/charlton-ppg-vascular-age-vascagenet-ajp-2022","links":[{"href":"https://doi.org/10.1152/ajpheart.00392.2021","label":"Paper","ariaLabel":"Paper for Assessing hemodynamics from the photoplethysmogram to gain insights into vascular age: a review from VascAgeNet (opens in new tab)"}]},{"slug":"imu2clip","title":"IMU2CLIP: Multimodal Contrastive Learning for IMU Motion Sensors from Egocentric Videos and Text","summary":"IMU2CLIP aligns IMU motion recordings with egocentric video and text by projecting them into CLIP's joint embedding space. This lets it translate motion into text and matching video, and fine-tuning it improved downstream tasks such as activity recognition.","venue":"arXiv preprint","year":2022,"open_code":null,"updated":null,"url":"/research/papers/imu2clip","links":[{"href":"https://arxiv.org/abs/2210.14395","label":"Paper","ariaLabel":"Paper for IMU2CLIP: Multimodal Contrastive Learning for IMU Motion Sensors from Egocentric Videos and Text (opens in new tab)"}]},{"slug":"rf-gait-parkinsons-monitoring","title":"Monitoring Gait at Home with Radio Waves in Parkinson's Disease: a Marker of Severity, Progression, and Medication Response","summary":"This Science Translational Medicine study tracked gait speed at home with a radio device that senses radio waves reflected off people's bodies. Fifty people with and without Parkinson's disease were monitored for up to a year, giving more than 200,000 gait speed measurements. Home gait speed correlated strongly with MDS-UPDRS scores, tracked progression more sensitively than the rating scale and captured responses to medication.","venue":"Science Translational Medicine 2022","year":2022,"open_code":null,"updated":null,"url":"/research/papers/rf-gait-parkinsons-monitoring","links":[{"href":"https://www.science.org/doi/10.1126/scitranslmed.adc9669","label":"Paper","ariaLabel":"Paper for Monitoring Gait at Home with Radio Waves in Parkinson's Disease: a Marker of Severity, Progression, and Medication Response (opens in new tab)"}]},{"slug":"simper-ssl-periodic-signals","title":"SimPer: Simple Self-Supervised Learning of Periodic Targets","summary":"SimPer is a contrastive self-supervised method for learning periodic signals. It uses custom augmentations, similarity measures and a generalized contrastive loss, and beat other self-supervised methods on real-world tasks in behaviour analysis, environmental sensing and healthcare, with better data efficiency and robustness.","venue":"arXiv preprint","year":2022,"open_code":null,"updated":null,"url":"/research/papers/simper-ssl-periodic-signals","links":[{"href":"https://arxiv.org/abs/2210.03115","label":"Paper","ariaLabel":"Paper for SimPer: Simple Self-Supervised Learning of Periodic Targets (opens in new tab)"}]},{"slug":"katori-ukb-16-sleep-phenotypes-pnas-2022","title":"The 103,200-arm acceleration dataset in the UK Biobank revealed a landscape of human sleep phenotypes","summary":"Katori and colleagues optimised ACCEL, their sleep/wake classification algorithm for wristband accelerometers, and applied it to roughly 100,000 arm-acceleration recordings in the UK Biobank. UMAP dimension reduction followed by DBSCAN clustering identified 16 sleep phenotypes. These included phenotypes related to social jet lag and chronotype, as well as seven different insomnia-like phenotypes.","venue":"Proceedings of the National Academy of Sciences","year":2022,"open_code":null,"updated":"2026-10-15","url":"/research/papers/katori-ukb-16-sleep-phenotypes-pnas-2022","links":[{"href":"https://doi.org/10.1073/pnas.2116729119","label":"Paper","ariaLabel":"Paper for The 103,200-arm acceleration dataset in the UK Biobank revealed a landscape of human sleep phenotypes (opens in new tab)"}]},{"slug":"delving-deep-imbalanced-regression-dir","title":"Delving into Deep Imbalanced Regression","summary":"This paper addresses regression problems where some target values have far fewer examples than others. The authors propose smoothing both the label and feature distributions so that nearby target values inform each other, which calibrates learned features for continuous labels.","venue":"arXiv preprint","year":2021,"open_code":null,"updated":null,"url":"/research/papers/delving-deep-imbalanced-regression-dir","links":[{"href":"https://arxiv.org/abs/2102.09554","label":"Paper","ariaLabel":"Paper for Delving into Deep Imbalanced Regression (opens in new tab)"}]},{"slug":"bent-smartwatch-interstitial-glucose-biomarkers-npj-2021","title":"Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches","summary":"Empatica E4 wristband + Dexcom G6 CGM, 16 participants, 8–10 days each, >25,000 interstitial glucose readings. E4 streams: PPG 64 Hz (HR at 1 Hz, IBI-based HRV), EDA and skin temperature 4 Hz, accelerometer 32 Hz. Food logs.","venue":"npj Digital Medicine 4:89","year":2021,"open_code":true,"updated":"2026-10-11","url":"/research/papers/bent-smartwatch-interstitial-glucose-biomarkers-npj-2021","links":[{"href":"https://doi.org/10.1038/s41746-021-00465-w","label":"Paper","ariaLabel":"Paper for Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches (opens in new tab)"}]},{"slug":"hilo-csem-obpm","title":"Hilo (formerly Aktiia) / CSEM oBPM — cuffless wrist BP from green PPG pulse-wave analysis","summary":"Hilo, formerly Aktiia, makes a wrist band that estimates blood pressure from optical PPG using pulse-wave analysis. In a Scientific Reports validation with 91 patients over four weeks, after calibration on day one, mean systolic differences from auscultation ranged from -3.02 to 0.46 mmHg across sitting, lying, lap and standing positions. The authors found it accurate across common positions, though precision varied between them.","venue":"Scientific Reports 2021 / Blood Pressure Monitoring 2021 / J Hypertens 2023 (independent) / FDA 510(k) 2025","year":2021,"open_code":null,"updated":null,"url":"/research/papers/hilo-csem-obpm","links":[{"href":"https://doi.org/10.1038/s41598-021-99294-w","label":"Paper","ariaLabel":"Paper for Hilo (formerly Aktiia) / CSEM oBPM — cuffless wrist BP from green PPG pulse-wave analysis (opens in new tab)"}]},{"slug":"limu-bert","title":"LIMU-BERT: Unleashing the Potential of Unlabeled Data for IMU Sensing Applications","summary":"LIMU-BERT applies BERT-style self-supervised training to unlabelled IMU data to learn general features for mobile sensing. The authors adapted the method to the characteristics of IMU sensors and kept the models light enough for mobile devices. Tested on four open datasets, models built on LIMU-BERT features and trained with few labels clearly beat earlier approaches in two typical IMU sensing applications.","venue":"SenSys 2021","year":2021,"open_code":null,"updated":null,"url":"/research/papers/limu-bert","links":[{"href":"https://dl.acm.org/doi/10.1145/3485730.3485937","label":"Paper","ariaLabel":"Paper for LIMU-BERT: Unleashing the Potential of Unlabeled Data for IMU Sensing Applications (opens in new tab)"}]},{"slug":"chowdhury-ppg-demographics-ml-bp-sensors-2020","title":"Estimating Blood Pressure from the Photoplethysmogram Signal and Demographic Features Using Machine Learning Techniques","summary":"PPG-BP (Guilin) public set (ppgbp-guilin): 657 finger PPG signals from 219 subjects, 1 kHz, 2.1 s each. After signal-quality screening: 222 signals from 126 subjects.","venue":"Sensors 20(11):3127","year":2020,"open_code":false,"updated":"2026-10-11","url":"/research/papers/chowdhury-ppg-demographics-ml-bp-sensors-2020","links":[{"href":"https://doi.org/10.3390/s20113127","label":"Paper","ariaLabel":"Paper for Estimating Blood Pressure from the Photoplethysmogram Signal and Demographic Features Using Machine Learning Techniques (opens in new tab)"}]},{"slug":"natarajan-fitbit-hrv-8-million-lancet-dh-2020","title":"Heart rate variability with photoplethysmography in 8 million individuals: a cross-sectional study","summary":"Wrist-PPG interbeat intervals from Fitbit users (random 24 h); RMSSD, SDRR, HF/LF, Poincaré in 5-min windows, stratified by age, sex, BMI, steps. Population norms for an autonomic archetype.","venue":"Lancet Digital Health 2020","year":2020,"open_code":null,"updated":"2026-10-15","url":"/research/papers/natarajan-fitbit-hrv-8-million-lancet-dh-2020","links":[]},{"slug":"colvonen-wearable-racial-bias-sleep-2020","title":"Limiting racial disparities and bias for wearable devices in health science research","summary":"Optical (green-light) sensors may perform worse on darker skin; calls for diverse validation. Fairness pitfall for any PPG-based archetype.","venue":"SLEEP 2020","year":2020,"open_code":null,"updated":"2026-10-15","url":"/research/papers/colvonen-wearable-racial-bias-sleep-2020","links":[{"href":"https://doi.org/10.1093/sleep/zsaa159","label":"Paper","ariaLabel":"Paper for Limiting racial disparities and bias for wearable devices in health science research (opens in new tab)"}]},{"slug":"stachl-personality-from-smartphone-pnas-2020","title":"Predicting personality from patterns of behavior collected with smartphones","summary":"Stachl and colleagues tested how well Big Five personality dimensions, at both broad domain and narrow facet levels, can be predicted from smartphone-sensed behaviour in a large sample. Six classes of behaviour were used: communication and social behaviour, music consumption, app usage, mobility, overall phone activity, and day- and night-time activity. Communication and social behaviour were the most predictive overall, and the authors highlight both the benefits and the privacy risks of large-scale smartphone data collection.","venue":"Proceedings of the National Academy of Sciences","year":2020,"open_code":null,"updated":"2026-10-15","url":"/research/papers/stachl-personality-from-smartphone-pnas-2020","links":[{"href":"https://doi.org/10.1073/pnas.1920484117","label":"Paper","ariaLabel":"Paper for Predicting personality from patterns of behavior collected with smartphones (opens in new tab)"}]},{"slug":"jones-chronotype-gwas-697828-nat-commun-2019","title":"Genome-wide association analyses of chronotype in 697,828 individuals provides insights into circadian rhythms","summary":"Using genome-wide data from 697,828 UK Biobank and 23andMe participants, Jones and colleagues increased the number of genetic loci associated with being a morning person from 24 to 351. In 85,760 people with activity-monitor sleep timing, the 5% carrying the most morningness alleles slept on average 25 minutes earlier than the 5% carrying the fewest. Mendelian randomisation suggested that being a morning person is causally associated with better mental health but does not affect BMI or type 2 diabetes risk.","venue":"Nature Communications","year":2019,"open_code":null,"updated":"2026-10-15","url":"/research/papers/jones-chronotype-gwas-697828-nat-commun-2019","links":[{"href":"https://doi.org/10.1038/s41467-018-08259-7","label":"Paper","ariaLabel":"Paper for Genome-wide association analyses of chronotype in 697,828 individuals provides insights into circadian rhythms (opens in new tab)"}]},{"slug":"mild-dehydration-eda-prv-simband-nutrients-2020","title":"Mild dehydration identification using ML on autonomic responses to cognitive stress","summary":"This study tested whether mild dehydration can be detected from how the autonomic nervous system responds to cognitive stress. The authors tried every combination of nine autonomic indices with several classifiers, using leave-one-subject-out validation. Accuracy reached 91.2% using indices from rest only or test only, and 86.8% using the change from rest to test.","venue":"Nutrients","year":2019,"open_code":null,"updated":"2026-10-13","url":"/research/papers/mild-dehydration-eda-prv-simband-nutrients-2020","links":[{"href":"https://doi.org/10.3390/nu12010042","label":"DOI","ariaLabel":"DOI for Mild dehydration identification using ML on autonomic responses to cognitive stress (opens in new tab)"}]},{"slug":"santisteban-mctq-vs-actigraphy-sleep-disord-2018","title":"Association between the Munich Chronotype Questionnaire and Wrist Actigraphy","summary":"This study compared the chronotype midpoint of sleep on free days from the self-reported Munich Chronotype Questionnaire (MCTQ) with the same measure from wrist actigraphy. In 115 participants aged 18-34 monitored for 4-6 nights, the corrected midpoint was 4:56 by MCTQ and 4:51 by actigraphy, with no significant difference and a strong correlation (r = 0.73). The 95% limits of agreement ran from about -1:37 to +2:14 hours.","venue":"Sleep Disorders","year":2018,"open_code":null,"updated":"2026-10-15","url":"/research/papers/santisteban-mctq-vs-actigraphy-sleep-disord-2018","links":[{"href":"https://doi.org/10.1155/2018/5646848","label":"Paper","ariaLabel":"Paper for Association between the Munich Chronotype Questionnaire and Wrist Actigraphy (opens in new tab)"}]},{"slug":"lyall-relative-amplitude-mood-ukb-lancet-psych-2018","title":"Association of disrupted circadian rhythmicity with mood disorders, subjective wellbeing, and cognitive function: a cross-sectional study of 91 105 participants from the UK Biobank","summary":"Accelerometer-derived relative amplitude (RA) in 91,105 UK Biobank participants; low RA associated with lifetime mood disorder, lower wellbeing and cognition (cross-sectional, adjusted). Anchor for rest-activity-rhythm archetypes.","venue":"Lancet Psychiatry 2018","year":2018,"open_code":null,"updated":"2026-10-15","url":"/research/papers/lyall-relative-amplitude-mood-ukb-lancet-psych-2018","links":[]},{"slug":"huang-hmm-circadian-rhythmicity-jrsi-2018","title":"Hidden Markov models for monitoring circadian rhythmicity in telemetric activity data","summary":"Wearable computing devices allow collection of densely sampled real-time information on movement enabling researchers and medical experts to obtain objective and non-obtrusive records of actual activity of a subject in the real world over many days. Their interest here is motivated by the use of activity data for evaluating and monitoring the circadian rhythmicity of subjects for research in chronobiology and chronotherapeutic healthcare. In order to translate the information from such high-volume data arising the authors propose the use of a Markov modelling approach which (i) naturally captures the notable square wave form observed in activity data along with heterogeneous ultradian variances over the circadian cycle of human activity, (ii) thresholds activity into different states in a probabilistic way while respecting time dependence and (iii) gives rise to circadian rhythm parameter estimates, based on probabilities of transitions between rest and activity, that are interpretable and of interest to circadian research.","venue":"J R Soc Interface 2018","year":2018,"open_code":null,"updated":"2026-10-15","url":"/research/papers/huang-hmm-circadian-rhythmicity-jrsi-2018","links":[{"href":"https://doi.org/10.1098/rsif.2017.0885","label":"Paper","ariaLabel":"Paper for Hidden Markov models for monitoring circadian rhythmicity in telemetric activity data (opens in new tab)"}]},{"slug":"willetts-ukb-sleep-activity-phenotypes-sci-rep-2018","title":"Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK Biobank participants","summary":"Willetts and colleagues developed a machine-learning model combining balanced random forests with hidden Markov models to detect activity and sleep modes from simple wrist-worn activity monitors. The model classified free-living behaviours with 87% accuracy across 159,504 recorded minutes from 132 adults. Applied to 96,220 UK Biobank participants, it showed, for example, that men spend more time in low- and high-intensity behaviours while women spend more time in mixed behaviours.","venue":"Scientific Reports","year":2018,"open_code":null,"updated":"2026-10-15","url":"/research/papers/willetts-ukb-sleep-activity-phenotypes-sci-rep-2018","links":[{"href":"https://doi.org/10.1038/s41598-018-26174-1","label":"Paper","ariaLabel":"Paper for Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK Biobank participants (opens in new tab)"}]},{"slug":"azarbarzin-hypoxic-burden-cvd-mortality-ehj-2019","title":"The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the Osteoporotic Fractures in Men Study and the Sleep Heart Health Study","summary":"This study tested whether the 'hypoxic burden' of sleep apnoea, the respiratory event-associated area under the oxygen desaturation curve, predicts mortality better than the apnoea-hypopnoea index (AHI). In the MrOS cohort (2,743 men) and the Sleep Heart Health Study (5,111 adults), hypoxic burden strongly predicted cardiovascular mortality while the AHI did not. MrOS participants in the top two hypoxic-burden quintiles had hazard ratios of 1.81 and 2.73, and the top SHHS quintile had a hazard ratio of 1.96.","venue":"European Heart Journal","year":2018,"open_code":null,"updated":"2026-10-15","url":"/research/papers/azarbarzin-hypoxic-burden-cvd-mortality-ehj-2019","links":[{"href":"https://doi.org/10.1093/eurheartj/ehy624","label":"Paper","ariaLabel":"Paper for The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the Osteoporotic Fractures in Men Study and the Sleep Heart Health Study (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/30376054","label":"PubMed","ariaLabel":"PubMed for The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the Osteoporotic Fractures in Men Study and the Sleep Heart Health Study (opens in new tab)"}]},{"slug":"evenson-lca-activity-mortality-nhanes-ajpm-2017","title":"Accelerometry-Assessed Latent Class Patterns of Physical Activity and Sedentary Behavior With Mortality","summary":"Evenson and colleagues applied latent class analysis to one week of hip-worn accelerometry from 4,510 NHANES 2003-2006 participants aged 40 and over, with mortality follow-up through 2011. Over an average of 6.6 years (513 deaths), more-active classes had lower all-cause mortality than the least active class. The class spending the most of the day in sedentary bouts had about twice the mortality risk of the class with the fewest (adjusted HR 2.10, 95% CI 1.11-3.97).","venue":"American Journal of Preventive Medicine","year":2017,"open_code":null,"updated":"2026-10-15","url":"/research/papers/evenson-lca-activity-mortality-nhanes-ajpm-2017","links":[{"href":"https://doi.org/10.1016/j.amepre.2016.10.033","label":"Paper","ariaLabel":"Paper for Accelerometry-Assessed Latent Class Patterns of Physical Activity and Sedentary Behavior With Mortality (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/28109457","label":"PubMed","ariaLabel":"PubMed for Accelerometry-Assessed Latent Class Patterns of Physical Activity and Sedentary Behavior With Mortality (opens in new tab)"}]},{"slug":"shaffer-ginsberg-hrv-metrics-norms-2017","title":"An Overview of Heart Rate Variability Metrics and Norms","summary":"Healthy biological systems exhibit complex patterns of variability that can be described by mathematical chaos. Heart rate variability (HRV) consists of changes in the time intervals between consecutive heartbeats called interbeat intervals (IBIs). They stress the importance of measurement context, including recording period length, subject age, and sex, on baseline HRV values.","venue":"Front Public Health 2017","year":2017,"open_code":null,"updated":"2026-10-15","url":"/research/papers/shaffer-ginsberg-hrv-metrics-norms-2017","links":[{"href":"https://doi.org/10.3389/fpubh.2017.00258","label":"Paper","ariaLabel":"Paper for An Overview of Heart Rate Variability Metrics and Norms (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/29034226","label":"PubMed","ariaLabel":"PubMed for An Overview of Heart Rate Variability Metrics and Norms (opens in new tab)"}]},{"slug":"phillips-sleep-regularity-index-sci-rep-2017","title":"Irregular sleep/wake patterns are associated with poorer academic performance and delayed circadian and sleep/wake timing","summary":"Phillips and colleagues followed 61 undergraduates for 30 days with sleep diaries and introduced the Sleep Regularity Index (SRI), a metric of day-to-day consistency in sleep/wake timing. Compared with regular sleepers, irregular sleepers had dim-light melatonin onset about 2.5 hours later, a later-peaking sleep propensity rhythm and lower-amplitude light exposure measured by wrist-worn photometry. SRI correlated positively with academic performance (r = 0.37), and circadian modelling suggested the timing differences were driven mainly by light exposure patterns.","venue":"Scientific Reports","year":2017,"open_code":null,"updated":"2026-10-15","url":"/research/papers/phillips-sleep-regularity-index-sci-rep-2017","links":[{"href":"https://doi.org/10.1038/s41598-017-03171-4","label":"Paper","ariaLabel":"Paper for Irregular sleep/wake patterns are associated with poorer academic performance and delayed circadian and sleep/wake timing (opens in new tab)"}]},{"slug":"doherty-ukb-wrist-accelerometer-pa-plos-one-2017","title":"Large Scale Population Assessment of Physical Activity Using Wrist Worn Accelerometers: The UK Biobank Study","summary":"Doherty and colleagues describe how seven-day, 100 Hz wrist accelerometry was collected and processed for more than 100,000 UK Biobank participants, including calibration, gravity and noise removal, and non-wear detection. Of 103,712 datasets received, 96,600 participants (93.3%) provided valid physical activity data, with a median wear time of 6.9 days. Overall activity (vector magnitude) was 7.5% lower per decade of age and varied strongly by time of day, while weekday/weekend and seasonal differences were small.","venue":"PLOS ONE","year":2017,"open_code":null,"updated":"2026-10-15","url":"/research/papers/doherty-ukb-wrist-accelerometer-pa-plos-one-2017","links":[{"href":"https://doi.org/10.1371/journal.pone.0169649","label":"Paper","ariaLabel":"Paper for Large Scale Population Assessment of Physical Activity Using Wrist Worn Accelerometers: The UK Biobank Study (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/28146576","label":"PubMed","ariaLabel":"PubMed for Large Scale Population Assessment of Physical Activity Using Wrist Worn Accelerometers: The UK Biobank Study (opens in new tab)"}]},{"slug":"onnela-rauch-digital-phenotyping-npp-2016","title":"Harnessing Smartphone-Based Digital Phenotyping to Enhance Behavioral and Mental Health","summary":"Defines digital phenotyping (moment-by-moment quantification of the individual-level human phenotype in situ using personal digital devices). Conceptual basis Sahha cites.","venue":"Neuropsychopharmacology 2016","year":2016,"open_code":null,"updated":"2026-10-15","url":"/research/papers/onnela-rauch-digital-phenotyping-npp-2016","links":[{"href":"https://doi.org/10.1038/npp.2016.7","label":"Paper","ariaLabel":"Paper for Harnessing Smartphone-Based Digital Phenotyping to Enhance Behavioral and Mental Health (opens in new tab)"}]},{"slug":"tudor-locke-how-many-steps-adults-ijbnpa-2011","title":"How many steps/day are enough? For adults","summary":"This review updates the evidence on how many steps per day are enough for adults and translates physical activity guidelines into step-based recommendations. Healthy adults typically take 4,000-18,000 steps/day, and about 100 steps/minute is a reasonable floor for moderate-intensity walking. Step translations of recommended moderate-to-vigorous activity that include habitual activity equate to roughly 7,100-11,000 steps/day.","venue":"International Journal of Behavioral Nutrition and Physical Activity","year":2011,"open_code":null,"updated":"2026-10-15","url":"/research/papers/tudor-locke-how-many-steps-adults-ijbnpa-2011","links":[{"href":"https://doi.org/10.1186/1479-5868-8-79","label":"Paper","ariaLabel":"Paper for How many steps/day are enough? For adults (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/21798015","label":"PubMed","ariaLabel":"PubMed for How many steps/day are enough? For adults (opens in new tab)"}]},{"slug":"chida-steptoe-stress-reactivity-meta-hypertension-2010","title":"Greater cardiovascular responses to laboratory mental stress are associated with poor subsequent cardiovascular risk status: a meta-analysis of prospective evidence","summary":"Chida and Steptoe meta-analysed prospective cohort studies linking cardiovascular responses to laboratory mental stress with later cardiovascular risk, covering 169 reactivity associations (36 articles) and 30 recovery associations (5 articles). Greater reactivity to and poorer recovery from stress were both associated with worse future cardiovascular status (r = 0.091 and r = 0.096). Incident hypertension was most consistently predicted by greater reactivity, and increased carotid intima-media thickness by poor recovery.","venue":"Hypertension","year":2010,"open_code":null,"updated":"2026-10-15","url":"/research/papers/chida-steptoe-stress-reactivity-meta-hypertension-2010","links":[{"href":"https://doi.org/10.1161/HYPERTENSIONAHA.109.146621","label":"Paper","ariaLabel":"Paper for Greater cardiovascular responses to laboratory mental stress are associated with poor subsequent cardiovascular risk status: a meta-analysis of prospective evidence (opens in new tab)"}]},{"slug":"sarabia-wrist-temperature-circadian-marker-physiol-behav-2008","title":"Circadian rhythm of wrist temperature in normal-living subjects. A candidate of new index of the circadian system","summary":"Sarabia and colleagues evaluated whether wrist skin temperature could serve as a marker rhythm of the human circadian system. In 99 university students wearing a wireless iButton sensor on the inner side of a sport wristband, wrist temperature was compared with sleep-wake diaries and oral temperature. The wrist temperature rhythm showed an inverse phase relationship with oral temperature and was phase-advanced by about 60 minutes.","venue":"Physiology & Behavior","year":2008,"open_code":null,"updated":"2026-10-15","url":"/research/papers/sarabia-wrist-temperature-circadian-marker-physiol-behav-2008","links":[{"href":"https://doi.org/10.1016/j.physbeh.2008.08.005","label":"Paper","ariaLabel":"Paper for Circadian rhythm of wrist temperature in normal-living subjects. A candidate of new index of the circadian system (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/18761026","label":"PubMed","ariaLabel":"PubMed for Circadian rhythm of wrist temperature in normal-living subjects. A candidate of new index of the circadian system (opens in new tab)"}]},{"slug":"roenneberg-mctq-life-between-clocks-jbr-2003","title":"Life between clocks: daily temporal patterns of human chronotypes","summary":"Munich ChronoType Questionnaire (MCTQ): chronotype as mid-sleep on free days (later corrected for sleep debt, MSFsc). Directly computable from wearable sleep timing on free days.","venue":"J Biol Rhythms 2003","year":2003,"open_code":null,"updated":"2026-10-15","url":"/research/papers/roenneberg-mctq-life-between-clocks-jbr-2003","links":[{"href":"https://doi.org/10.1177/0748730402239679","label":"Paper","ariaLabel":"Paper for Life between clocks: daily temporal patterns of human chronotypes (opens in new tab)"}]},{"slug":"van-someren-nonparametric-rest-activity-is-iv-1999","title":"Bright light therapy: improved sensitivity to its effects on rest-activity rhythms in Alzheimer patients by application of nonparametric methods","summary":"It was demonstrated that the light-induced improvement in coupling of the rest-activity rhythm to the environmental zeitgeber of bright light is better detected using nonparametric procedures. Guidelines for analyses of actigraphic data are given to improve the sensitivity to treatment effects in future studies.","venue":"Chronobiology International 1999","year":1999,"open_code":null,"updated":"2026-10-15","url":"/research/papers/van-someren-nonparametric-rest-activity-is-iv-1999","links":[{"href":"https://doi.org/10.3109/07420529908998724","label":"Paper","ariaLabel":"Paper for Bright light therapy: improved sensitivity to its effects on rest-activity rhythms in Alzheimer patients by application of nonparametric methods (opens in new tab)"},{"href":"https://pubmed.ncbi.nlm.nih.gov/10442243","label":"PubMed","ariaLabel":"PubMed for Bright light therapy: improved sensitivity to its effects on rest-activity rhythms in Alzheimer patients by application of nonparametric methods (opens in new tab)"}]},{"slug":"horne-ostberg-meq-1976","title":"A self-assessment questionnaire to determine morningness-eveningness in human circadian rhythms","summary":"Morningness–Eveningness Questionnaire (MEQ); preference-based chronotype; the instrument behind Ultrahuman's AutoMEQ.","venue":"Int J Chronobiol 1976","year":1976,"open_code":null,"updated":"2026-10-15","url":"/research/papers/horne-ostberg-meq-1976","links":[]},{"slug":"ppg-fm-decodability-audit-clinical-claims-iclr2027-sub","title":"A Decodability-Grounded Audit of Clinical Prediction Claims in PPG Foundation Models","summary":"Audits frozen representations of four encoders on 37,784 MC-MED encounters via concept decodability, covariate-adjusted and incremental prediction, and post-erasure predictability with encounter-matched ECG comparisons. AnyPPG decodes pulse-derived concepts well, but for potassium (6,348 matched encounters) its correlation drops 0.21 → 0.05 after age and renal adjustment, while ECGFounder retains 0.32 → 0.27 (paired adjusted diff +0.22, 95% CI [0.19, 0.26]). Creatinine also attenuates.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/ppg-fm-decodability-audit-clinical-claims-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=WGmQvuKFU7","label":"Paper","ariaLabel":"Paper for A Decodability-Grounded Audit of Clinical Prediction Claims in PPG Foundation Models (opens in new tab)"}]},{"slug":"physio-hnet-physiology-aware-hierarchical-waveform-fm-iclr2027-sub","title":"A Physiology-Aware Hierarchical Foundation Model for Long-Context Waveform Representation Learning (Physio-HNet)","summary":"Causal hierarchical waveform FM replacing fixed patches with physiology-aware dynamic chunking (landmark / cardiac-cycle guided sub-beat and beat-level units), with hierarchical compression and recurrent state carrying context across windows. After SSL pretraining, routine vital signs (HR, RR, SpO₂, BP) supervise higher-level representations along each recording (\"clinically anchored continued pretraining\"). 1.2 M parameters; SSL encoder comparable to existing waveform FMs; continued pretraining gives highest R² and lowest MAE in 11 of 12 BP comparisons across two external clinical cohorts and two probing protocols (all six operating-room comparisons).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/physio-hnet-physiology-aware-hierarchical-waveform-fm-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=y4KXe5JOfn","label":"Paper","ariaLabel":"Paper for A Physiology-Aware Hierarchical Foundation Model for Long-Context Waveform Representation Learning (Physio-HNet) (opens in new tab)"}]},{"slug":"beatflow-ecg-rectified-flow-ppg-imu-ecg-reconstruction-iclr2027-sub","title":"BeatFlow-ECG: Rectified Flow for ECG Reconstruction from Indirect Wearable Signals","summary":"Conditional rectified-flow model reconstructing single-channel ECG from synchronized PPG + inertial signals (conv encoder–decoder, transformer bottleneck, explicit flow-time conditioning). Motion enters via IMU-derived conditioning features, motion-dependent loss weighting and an easy-to-hard curriculum. Leave-one-subject-out on PPG-DaLiA and WESAD: best among deterministic, adversarial and diffusion baselines on all waveform and beat-timing metrics — Pearson r 0.983 / 0.986, R-peak F1 0.946 / 0.955 (DaLiA / WESAD); vs Conditional DDPM-1D, L1 error 0.085→0.062 (DaLiA) and 0.074→0.055 (WESAD).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/beatflow-ecg-rectified-flow-ppg-imu-ecg-reconstruction-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=nTHoR4dSMe","label":"Paper","ariaLabel":"Paper for BeatFlow-ECG: Rectified Flow for ECG Reconstruction from Indirect Wearable Signals (opens in new tab)"}]},{"slug":"beatformer-delay-aware-ecg-ppg-fm-iclr2027-sub","title":"BEATFORMER: A Delay-Aware Cardiovascular Foundation Model for Beat-Resolved ECG–PPG Representation Learning","summary":"Organizes synchronized ECG and PPG as sequences of paired heartbeats anchored on shared R peaks. A delay-aware fusion module estimates a posterior over beat-specific ECG→PPG delay and gates bidirectional cross-modal attention by its confidence; an orthogonal beat–phase Transformer models within-beat morphology and across-beat rhythm with physical-time and rhythm-aware positional biases. Pretrained on 84.76M cardiac cycles; wins 9/15 downstream tasks vs existing FMs, PAT MAE 11.48 ms (−24.6%), −11.3% reconstruction RMSE, best denoising SNR gain in 13/15 conditions.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":"https://anonymous.4open.science/r/anonymous-model-release-6541/ (code; weights promised post-acceptance)","updated":"2026-10-05","url":"/research/papers/beatformer-delay-aware-ecg-ppg-fm-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=fI7KCDBNtX","label":"Paper","ariaLabel":"Paper for BEATFORMER: A Delay-Aware Cardiovascular Foundation Model for Beat-Resolved ECG–PPG Representation Learning (opens in new tab)"}]},{"slug":"cardiofm-joint-ecg-ppg-2026","title":"CardioFM: A Multimodal Foundation Model for Joint ECG and PPG Representation Learning","summary":"CardioFM is a self-supervised foundation model that encodes ECG lead II and PPG together using bidirectional cross-modal attention and adaptive residual vector quantization. It was pretrained on more than 500,000 hours from about 63,000 patients in ICU, surgical, ambulatory and consumer-wearable settings. It reached F1 0.86 for disease classification on PTB-XL, estimated QT interval within 20.2 ms and pulse arrival time within 22.7 ms, and beat ECG-FM, PaPaGei and TimesFM features for ICU false-alarm reduction.","venue":null,"year":null,"open_code":null,"updated":"2026-09-11","url":"/research/papers/cardiofm-joint-ecg-ppg-2026","links":[{"href":"https://pubmed.ncbi.nlm.nih.gov/42183386/","label":"Paper","ariaLabel":"Paper for CardioFM: A Multimodal Foundation Model for Joint ECG and PPG Representation Learning (opens in new tab)"}]},{"slug":"wearable-world-models-depth-beats-breadth-iclr2027-sub","title":"Depth Beats Breadth: Scaling Wearable World Models of Human Physiology","summary":"Action-conditioned wearable world model forecasting physiological trajectories from personal history and specified future behaviour. Corpus: 5.4M users, ~2B user-days. Scaling study over population size, training context, model capacity and inference context: longer personal histories beat more users, much of it from longer inference context; larger models benefit more from long context.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/wearable-world-models-depth-beats-breadth-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=a1BIT7rD0i","label":"Paper","ariaLabel":"Paper for Depth Beats Breadth: Scaling Wearable World Models of Human Physiology (opens in new tab)"}]},{"slug":"dpikan-physics-informed-kan-cuffless-bp-ring-iclr2027-sub","title":"DPIKAN: Decomposed Physics-Informed KANs for Physiological Signal Representation Learning","summary":"Compact physiology-informed framework: (1) multi-scale waveform modeling — decomposes signals into fast- and slow-varying components handled by a dual-order KAN; (2) cardiovascular-descriptor guidance — a physics-informed KAN integrates descriptors and regularizes how predictions respond to them. Evaluated for BP on four datasets spanning PPG, bioimpedance (BioZ) and mmWave. Introduces Ring-HGCPT, a smart-ring PPG dataset recorded during controlled cardiovascular perturbations; leave-one-subject-out there: mean error 1.08 / 0.09 mmHg (SBP/DBP), error SD 5.41 / 4.81 mmHg.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/dpikan-physics-informed-kan-cuffless-bp-ring-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=vCHNhjXVxg","label":"Paper","ariaLabel":"Paper for DPIKAN: Decomposed Physics-Informed KANs for Physiological Signal Representation Learning (opens in new tab)"},{"href":"https://github.com/Zest86/DPIKAN","label":"Code","ariaLabel":"Code for DPIKAN: Decomposed Physics-Informed KANs for Physiological Signal Representation Learning (opens in new tab)"}]},{"slug":"har-foundation-models-survey-outlook","title":"Foundation Models Defining a New Era in Human Activity Recognition: A Survey and Outlook","summary":"This ACM IMWUT survey reviews foundation models for sensor-based human activity recognition and organizes them with a lifecycle-based taxonomy. A companion GitHub repository keeps an updated index of models, datasets and resources, with contributors from DFKI, RPTU, Georgia Tech and NWPU.","venue":null,"year":null,"open_code":null,"updated":"2026-09-11","url":"/research/papers/har-foundation-models-survey-outlook","links":[{"href":"https://github.com/zhaxidele/Foundation-Models-Defining-A-New-Era-In-Human-Activity-Recognition","label":"Paper","ariaLabel":"Paper for Foundation Models Defining a New Era in Human Activity Recognition: A Survey and Outlook (opens in new tab)"}]},{"slug":"imudino-self-distilled-inertial-fm-iclr2027-sub","title":"IMUDINO: Self-distilled Foundation Model for Inertial Sensing","summary":"Adapts DINO self-distillation to IMU via motion transformations that simulate variations in motion pattern and sensor setup; frequency-aware positional encoding lets one plain transformer ingest any window length / sampling rate. Positioned as the first inertial FM competitive with SOTA supervised methods across HAR, health, motion capture and robot odometry; global and dense features transfer across users, setups, tasks and embodiments. \"Code and checkpoints will be released\" (abstract).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/imudino-self-distilled-inertial-fm-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=psN5UQkhct","label":"Paper","ariaLabel":"Paper for IMUDINO: Self-distilled Foundation Model for Inertial Sensing (opens in new tab)"}]},{"slug":"qraft-reliability-conditioned-masked-modeling-ecg-ppg-iclr2027-sub","title":"Learning What to Trust: Reliability-Conditioned Masked Modeling for Physiological Time Series (QRAFT)","summary":"Self-supervised masked learner where fine-grained per-patch reliability decides which patches stay visible and which become reconstruction targets; a shared temporal encoder lets reliable evidence in one source support recovery in the other; temporal + spectral + contrastive objective. Introduces difficulty-matched random masking as a control. At the same 38.5% mask ratio: PPG reconstruction RMSE −23.2% vs random masking and −10.0% vs the difficulty-matched control.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/qraft-reliability-conditioned-masked-modeling-ecg-ppg-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=u5PBzl1YZJ","label":"Paper","ariaLabel":"Paper for Learning What to Trust: Reliability-Conditioned Masked Modeling for Physiological Time Series (QRAFT) (opens in new tab)"}]},{"slug":"native-calibration-multimodal-waveform-fm-iclr2027-sub","title":"Long-Horizon Physiological Analysis via Native Calibration of a Multimodal Waveform Foundation Model","summary":"Introduces 'native calibration': estimate a current lab value as the patient's earlier measured value plus the change read from current vs past waveforms embedded by a frozen FM. Also introduces a new (unnamed) multimodal FM — transformer pretrained without labels to relate ECG and PPG and to match a patient's PPG across time. Tasks: hemoglobin change between ED visits, potassium within ICU stays, first-draw ED labs, intra-operative haemodynamics, and 24 standard PPG benchmark tasks.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-12","url":"/research/papers/native-calibration-multimodal-waveform-fm-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=UGsYiRYoBL","label":"Paper","ariaLabel":"Paper for Long-Horizon Physiological Analysis via Native Calibration of a Multimodal Waveform Foundation Model (opens in new tab)"}]},{"slug":"msma-ppg-selected-wavelength-moe-iclr2027-sub","title":"MSMA-PPG: Channel-Aligned Expert Routing for Selected-Wavelength PPG Prediction","summary":"Learns from synchronized multi-wavelength PPG and predicts from a single selected PPG stream + acceleration. Shared optical backbone with wavelength-specific projections/adapters, full-to-subset alignment, and a Channel-Aligned Mixture of Experts with availability-masked top-2 routing. With routing disabled: lowest participant-macro HR MAE on EarSet (5.020 bpm) and PPG-DaLiA (9.246 bpm), lowest SBP MAE (15.456 mmHg), second-best AF AUPRC/AUROC (0.5177/0.8909).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/msma-ppg-selected-wavelength-moe-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=ewFflGoy46","label":"Paper","ariaLabel":"Paper for MSMA-PPG: Channel-Aligned Expert Routing for Selected-Wavelength PPG Prediction (opens in new tab)"}]},{"slug":"people-over-segments-ppg-pretraining-scaling-law-iclr2027-sub","title":"People over Segments: A Data-Axis Scaling Law for Person-Identity Contrastive PPG Pretraining","summary":"On a real-world PPG cohort of 114k people and >200 controlled pretraining runs, splits a fixed segment budget into K people × M segments/person for person-identity contrastive SSL. More distinct people is the stronger axis; the people axis reduces its error share ~2× faster per decade than segments; more people also improve demographic prediction. Fitted law yields a budget-adaptive acquisition rule (recruit more people at small budgets, shift to more segments/person as budget grows), validated by training the recommended points.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/people-over-segments-ppg-pretraining-scaling-law-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=bRFnSHJjbJ","label":"Paper","ariaLabel":"Paper for People over Segments: A Data-Axis Scaling Law for Person-Identity Contrastive PPG Pretraining (opens in new tab)"}]},{"slug":"physiodyt-directed-dynamic-transport-ecg-ppg-iclr2027-sub","title":"PhysioDyT: Physiologically Directed Dynamic Transport for ECG–PPG Representation Learning","summary":"Cross-modal predictive SSL that predicts sample-dependent distributions over positive ECG→PPG temporal offsets, preserves ordering via soft monotone transport, and sends unreliable/unmatched events to a dustbin state; combined with bidirectional latent prediction. Evaluated on a strictly patient- and pretraining-disjoint wearable cohort incl. PPG-only coronary heart disease prediction across pretraining seeds; mechanism probes (direction reversal, cross-patient mismatch, temporal shifts, transport ablations).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/physiodyt-directed-dynamic-transport-ecg-ppg-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=ZE56HeqJxT","label":"Paper","ariaLabel":"Paper for PhysioDyT: Physiologically Directed Dynamic Transport for ECG–PPG Representation Learning (opens in new tab)"}]},{"slug":"physioskeptic-sqi-grounded-ecg-ppg-rhythm-reasoning-iclr2027-sub","title":"PhysioSkeptic: Skeptical Reasoning Grounded in Signal Quality for Multimodal Rhythm Analysis","summary":"PhysioPatch encoder aligns waveform patches to cardiac landmarks and emits heartbeat-level tokens plus a Patch Report (rhythm cues, confidence, signal quality), so an LLM can run conflict checks and quality-conditioned belief revision instead of reasoning over lossy waveform-to-text summaries. On MIMIC-III-Ext-PPG: 0.886 Macro-F1, +6.2 pp over generic Debate and +4.1 pp over an encoder-matched control; lowest observed ECE among GPT-5.2 methods; anchor-trap retention 56.6%→16.8%. Removing SQI anchoring gives the largest drop among reasoning controls.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/physioskeptic-sqi-grounded-ecg-ppg-rhythm-reasoning-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=Ytw4pWrGGW","label":"Paper","ariaLabel":"Paper for PhysioSkeptic: Skeptical Reasoning Grounded in Signal Quality for Multimodal Rhythm Analysis (opens in new tab)"}]},{"slug":"physppg-physics-informed-contrastive-ppg-iclr2027-sub","title":"PhysPPG: Physics-Informed Contrastive Learning for Photoplethysmography Representation","summary":"Temporal encoder maps PPG to latent trajectories shaped jointly by a contrastive objective and a generalized Euler–Lagrange residual (learnable kinetic/potential energies + non-conservative forces) — a low-order dynamical inductive bias without physiological labels. Frozen-encoder evaluations show benefits for BP and HR estimation on WildPPG (substantial motion / environmental variation). The learned residual also discriminates PPG vs ECG (compatibility diagnostic).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/physppg-physics-informed-contrastive-ppg-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=AqkAEGn47Z","label":"Paper","ariaLabel":"Paper for PhysPPG: Physics-Informed Contrastive Learning for Photoplethysmography Representation (opens in new tab)"}]},{"slug":"pulse2view-same-sensor-cross-view-ppg-pretraining-iclr2027-sub","title":"Pulse2View: Same-Sensor Cross-View Pre-Training for Patient-Level Multi-Disease Classification from Wearable PPG","summary":"Aligns two views of each 30 s PPG recording with symmetric InfoNCE: a Ricker-CWT three-channel image (PPG + 1st/2nd differences) per 10 s segment, and a Conv1D-BiGRU temporal view; both encoders fused downstream, inference PPG-only. Pretrained on 99,701 unlabeled 30 s windows; evaluated on the private hospital-derived Clinical Wearable PPG Cohort (CWPC; 14,811 patients, 100,657 windows): highest three-seed mean on all four primary patient-level metrics — CHD AUROC 0.7859 ± 0.0044, Macro AUROC 0.7150 ± 0.0012 (+3.2 / +1.4 pts over ImageNet-style ConvNeXtV2-Tiny init; +2.6 / +0.8 over MOMENT-small). Gains hold at 25% / 50% label budgets.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/pulse2view-same-sensor-cross-view-ppg-pretraining-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=r6dS5BYw3y","label":"Paper","ariaLabel":"Paper for Pulse2View: Same-Sensor Cross-View Pre-Training for Patient-Level Multi-Disease Classification from Wearable PPG (opens in new tab)"}]},{"slug":"pulsebound-future-beat-forecasting-ppg-ssl-iclr2027-sub","title":"PulseBound: Future-Beat State Forecasting Under an Explicit Information Boundary","summary":"Points out that causal attention doesn't guarantee one-sided information access when normalization, nonlocal transforms or companion views see withheld samples. PulseBound enforces a stored-window cutoff (prefix-only normalization, suffix replacement before derived views, aligned masking) so the forecast context is provably invariant to the stored suffix, then predicts nine rhythm/morphology descriptors for up to four future beats (optional ECG-derived PAT supervision in training only). On held-out MIMIC / VitalDB groups: 28.06% / 22.22% lower nine-state MAE vs last-beat persistence; in a seven-model comparison across 13 downstream tasks, best mean on nine frozen linear-probe and seven full-fine-tune tasks.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/pulsebound-future-beat-forecasting-ppg-ssl-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=agXgfyZW3G","label":"Paper","ariaLabel":"Paper for PulseBound: Future-Beat State Forecasting Under an Explicit Information Boundary (opens in new tab)"}]},{"slug":"pulselm-ppg-text-qa-2026","title":"pulselm-ppg-text-qa-2026","summary":"Singapore Management University / Cambridge team aggregates 16 public PPG datasets (clinical, lab, and in-the-wild — fingertip, ear, and wrist) into a single harmonized 10-second/125Hz format, then converts existing labels (HR, HRV, BP category, SQI, stress/affect, sleep-disordered breathing, AF, arrhythmia, SpO2, RR) into ~2.5 million natural-language question-answer pairs across 12 task categories. This is a dataset-and-benchmark paper, not a new pretraining method — no novel encoder or SSL objective is proposed. Baseline results use existing multimodal PPG-aware LLMs (PPG encoder + LoRA fine-tuned LLaMA/Qwen) on the QA format.","venue":null,"year":null,"open_code":null,"updated":null,"url":"/research/papers/pulselm-ppg-text-qa-2026","links":[]},{"slug":"ramp-reliability-adaptive-masking-multimodal-physio-pretraining-iclr2027-sub","title":"RAMP: Reliability Adaptive Masking for Pretraining Noisy Multimodal Physiological Signals","summary":"Profiles each ECG and PPG patch by quality, corruption type and spectral degradation; a compact Reliability Policy Network jointly selects masking threshold, temporal granularity and cross-modal fusion weights via differentiable relaxations, reconstructing degraded regions from cleaner temporal / cross-modal context. On MIMIC-III and VitalDB, pretraining RMSE −16% vs the strongest fixed-policy baseline (largest gains under asymmetric modality corruption). Pretrained encoder: 93.9 / 93.9 / 93.2% mean ACC / macro-F1 / ROC-AUC across six downstream benchmarks spanning ECG, PPG, EEG and accelerometry (reproduced baselines, unified protocol).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-06","url":"/research/papers/ramp-reliability-adaptive-masking-multimodal-physio-pretraining-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=iPvVOfpJOk","label":"Paper","ariaLabel":"Paper for RAMP: Reliability Adaptive Masking for Pretraining Noisy Multimodal Physiological Signals (opens in new tab)"}]},{"slug":"spaft-ppg-fm-structured-pruning-finetuning-iclr2027-sub","title":"Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning (SPAFT)","summary":"Finds that even fully fine-tuned PPG FMs do not consistently beat general time-series models, and that pretrained PPG FMs show inherent sparsity / redundancy. SPAFT constrains adaptive fine-tuning to a compact parameter subspace chosen by parameter importance and structural redundancy (prune-then-finetune). Across four PPG FMs, five benchmark datasets and ten downstream tasks: win rates 82.8% (regression) and 87.5% (classification).","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/spaft-ppg-fm-structured-pruning-finetuning-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=e7PAlA4pjo","label":"Paper","ariaLabel":"Paper for Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning (SPAFT) (opens in new tab)"}]},{"slug":"whubert-layer-selection-wearable-biosignals-iclr2027-sub","title":"wHuBERT: Layer Selection for Iterative Masked-Prediction Pre-training on Low-Resolution Wearable Biosignals","summary":"Conformer encoder pretrained with HuBERT-style iterative masked prediction on 25 Hz PPG + accelerometry. Shows the choice of clustering layer for iteration 2 matters and is expensive to validate; proposes per-layer probes at checkpoints; label-free representation statistics do not reliably pick the layer. On two public sleep corpora the teacher and a distilled student both exceed a consumer-device sleep-staging baseline.","venue":"ICLR 2027 submission (under review)","year":null,"open_code":false,"updated":"2026-10-05","url":"/research/papers/whubert-layer-selection-wearable-biosignals-iclr2027-sub","links":[{"href":"https://openreview.net/forum?id=u3iC4BjfaC","label":"Paper","ariaLabel":"Paper for wHuBERT: Layer Selection for Iterative Masked-Prediction Pre-training on Low-Resolution Wearable Biosignals (opens in new tab)"}]}]}