{"generated_at":"2026-10-16T14:01:47-05:00","count":92,"source":"https://chatppg.com/research","items":[{"slug":"aether","name":"Aether","tagline":"compute-optimal physiological waveform FM family (20M–2.1B)","summary":"Aether is a family of more than 100 physiological waveform foundation models, from 20M to 2.1B parameters, trained on up to 36.3M hours of waveforms to study scaling laws. The authors built eight clinical prediction tasks from MIMIC-III and tested the models with linear probes. The 720M model beat all existing baseline foundation models on all eight tasks, and the fitted scaling law predicted downstream error to within 0.5% MAE.","modalities":["ECG","PPG","ABP","RESP"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-10-06","url":"/research/models/aether","links":[{"href":"https://openreview.net/forum?id=hRTpFhuTuq","label":"Paper","ariaLabel":"Paper for Aether (opens in new tab)"},{"href":"https://arxiv.org/abs/2610.05649","label":"arXiv","ariaLabel":"arXiv for Aether (opens in new tab)"}]},{"slug":"ai-ppg-vascular-age","name":"AI-PPG age (PPG-VascularAge)","tagline":null,"summary":null,"modalities":["PPG"],"year":null,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-10-16","url":"/research/models/ai-ppg-vascular-age","links":[]},{"slug":"anymo","name":"AnyMo","tagline":null,"summary":"AnyMo models human motion from IMUs placed anywhere on the body. It simulates IMU signals over dense body-surface positions, pretrains a graph encoder on them, turns multi-position IMU data into full-body motion tokens and aligns those tokens 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.","modalities":["IMU","ACC","gyro"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-10-16","url":"/research/models/anymo","links":[{"href":"https://arxiv.org/abs/2605.22715","label":"Paper","ariaLabel":"Paper for AnyMo (opens in new tab)"},{"href":"https://github.com/Breezelled/AnyMo","label":"Code","ariaLabel":"Code for AnyMo (opens in new tab)"}]},{"slug":"anyppg","name":"AnyPPG","tagline":null,"summary":"ECG-guided CLIP-style PPG FM (PPG encoder aligned to frozen ECGFounder) over ~110k hours paired clinical PPG+ECG; inference is PPG-only.","modalities":["PPG","ECG"],"year":2026,"params":null,"open_weights":true,"license":"No license granted for the code or checkpoint","updated":"2026-10-12","url":"/research/models/anyppg","links":[{"href":"https://arxiv.org/abs/2511.01747","label":"Paper","ariaLabel":"Paper for AnyPPG (opens in new tab)"},{"href":"https://doi.org/10.1145/3770855.3818911","label":"DOI","ariaLabel":"DOI for AnyPPG (opens in new tab)"},{"href":"https://github.com/PKUDigitalHealth/AnyPPG","label":"Code","ariaLabel":"Code for AnyPPG (opens in new tab)"}]},{"slug":"apple-accel-kd","name":"Apple accelerometer FM via PPG→ACC distillation","tagline":null,"summary":"This Apple model is an accelerometer foundation model trained by distilling knowledge from a PPG encoder. It used 20 million minutes of unlabelled data from about 172K participants in the Apple Heart and Movement Study. Accelerometer embeddings could retrieve the matching PPG embedding with 99.2% top-1 accuracy, and the distilled encoders beat accelerometer-only encoders by 23% to 49% on health targets.","modalities":["ACC (student)","PPG (teacher)"],"year":2024,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/apple-accel-kd","links":[{"href":"https://arxiv.org/abs/2412.11276","label":"Paper","ariaLabel":"Paper for Apple accelerometer FM via PPG→ACC distillation (opens in new tab)"}]},{"slug":"apple-wbm","name":"Apple WBM (Wearable Behavioral Model)","tagline":null,"summary":null,"modalities":["behavioral_metrics (27 HealthKit variables)"],"year":2025,"params":null,"open_weights":false,"license":"No released weights or code","updated":"2026-09-26","url":"/research/models/apple-wbm","links":[{"href":"https://arxiv.org/abs/2507.00191","label":"Paper","ariaLabel":"Paper for Apple WBM (Wearable Behavioral Model) (opens in new tab)"}]},{"slug":"apple-wearable-biosignals","name":"Apple wearable biosignals FM","tagline":null,"summary":"Apple ICLR 2024 large-scale PPG/ECG FMs with participant-level contrastive pairs on ~141k AHMS participants — no downloadable checkpoint on official page.","modalities":["PPG","ECG"],"year":2024,"params":null,"open_weights":false,"license":"No released weights or code; no license","updated":"2026-09-26","url":"/research/models/apple-wearable-biosignals","links":[{"href":"https://arxiv.org/abs/2312.05409","label":"Paper","ariaLabel":"Paper for Apple wearable biosignals FM (opens in new tab)"}]},{"slug":"beatformer","name":"BeatFormer","tagline":"delay-aware beat-resolved ECG–PPG cardiovascular FM","summary":"Results (abstract): beats existing FMs on 9/15 physiological classification/regression tasks; PAT MAE 11.48 ms (24.6% below strongest baseline); −11.3% mean RMSE over 11 masked-reconstruction conditions; best median SNR gain in 13/15 denoising conditions.","modalities":["ECG","PPG"],"year":2027,"params":null,"open_weights":"promised","license":"Not stated","updated":"2026-10-05","url":"/research/models/beatformer","links":[{"href":"https://openreview.net/forum?id=fI7KCDBNtX","label":"Paper","ariaLabel":"Paper for BeatFormer (opens in new tab)"},{"href":"https://anonymous.4open.science/r/anonymous-model-release-6541/","label":"Code","ariaLabel":"Code for BeatFormer (opens in new tab)"}]},{"slug":"bio-pm","name":"Bio-PM","tagline":null,"summary":"Bio-PM is a self-supervised model for wrist accelerometer data that splits motion into movement segments based on the submovement theory of motor control. A Transformer encoder is pretrained to reconstruct masked segments, which focuses it on how movements are organized rather than on raw waveform shape. Pretrained on about 28,000 hours from 11,000 NHANES participants, it beat strong wearable SSL baselines on six subject-disjoint activity recognition benchmarks.","modalities":["ACC"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-10-01","url":"/research/models/bio-pm","links":[{"href":"https://arxiv.org/abs/2603.10961","label":"Paper","ariaLabel":"Paper for Bio-PM (opens in new tab)"}]},{"slug":"blpm","name":"BLPM (Brain Latent Predictive Model)","tagline":null,"summary":"BLPM (Brain Latent Predictive Model) is an EEG-language foundation model that treats EEG decoding tasks as predicting continuous semantic embeddings. Its CELP encoder learns by predicting latent targets, and a multi-query module aligns EEG representations with text in a shared space. It generalized consistently across several benchmarks.","modalities":["EEG + text embeddings"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/blpm","links":[{"href":"https://arxiv.org/abs/2608.11656","label":"Paper","ariaLabel":"Paper for BLPM (Brain Latent Predictive Model) (opens in new tab)"}]},{"slug":"cap","name":"CAP","tagline":null,"summary":"Patient-level PPG↔EHR contrastive foundation-style encoder (KDD 2026). Anchors long PPG windows to LLM-distilled clinical semantics from MC-MED ED data; reports strong AF/HR/RR/BP transfer vs SiamQuality / GPT-PPG / PaPaGei.","modalities":["PPG","EHR_text"],"year":2026,"params":"~130M","open_weights":false,"license":"Not stated","updated":"2026-09-28","url":"/research/models/cap","links":[{"href":"https://arxiv.org/abs/2606.15284","label":"Paper","ariaLabel":"Paper for CAP (opens in new tab)"},{"href":"https://doi.org/10.1145/3770855.3818881","label":"DOI","ariaLabel":"DOI for CAP (opens in new tab)"},{"href":"https://github.com/gody123gody/CAP","label":"Code","ariaLabel":"Code for CAP (opens in new tab)"}]},{"slug":"capfm","name":"CapFM","tagline":"Foundation Model for Capacitive Touch Sensing","summary":"Announced capacitive-touch foundation model.","modalities":["capacitive-touch"],"year":null,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-02","url":"/research/models/capfm","links":[]},{"slug":"cardiofm","name":"CardioFM","tagline":null,"summary":"CardioFM is a self-supervised foundation model that encodes ECG lead II and PPG together, using cross-modal attention and residual vector quantization. It reached an F1 of 0.86 for cardiovascular disease classification on PTB-XL and estimated QT interval with a mean error of 20.2 ms and pulse arrival time with 22.7 ms. Its embeddings also predict age, sex and BMI without diagnostic labels.","modalities":["ECG","PPG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-26","url":"/research/models/cardiofm","links":[{"href":"https://doi.org/10.21203/rs.3.rs-9652631/v1","label":"DOI","ariaLabel":"DOI for CardioFM (opens in new tab)"}]},{"slug":"cardiostate-jepa","name":"CardioState-JEPA","tagline":null,"summary":"CardioState-JEPA is a cardiac foundation model that learns one shared representation across ECG, PPG and heart sounds (PCG). Each signal is mapped into a common token space and fed to 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.","modalities":["ECG","PPG","PCG"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-10-01","url":"/research/models/cardiostate-jepa","links":[{"href":"https://arxiv.org/abs/2608.12944v2","label":"Paper","ariaLabel":"Paper for CardioState-JEPA (opens in new tab)"},{"href":"https://arxiv.org/abs/2608.12944","label":"arXiv","ariaLabel":"arXiv for CardioState-JEPA (opens in new tab)"},{"href":"https://github.com/hamzashafiq28/CardioState-Jepa","label":"Code","ariaLabel":"Code for CardioState-JEPA (opens in new tab)"}]},{"slug":"clef","name":"CLEF","tagline":null,"summary":"CLEF is a foundation model for single-lead ECG trained with contrastive learning guided by patient risk scores derived from clinical metadata. It was pretrained on 12-lead ECGs from 161K patients in MIMIC-IV-ECG and tested on 18 classification and regression tasks across 7 held-out datasets. It improved average AUROC by at least 2.6% and cut regression error by at least 3.2% compared with self-supervised baselines, and weights are released in small, base and large sizes.","modalities":["ECG"],"year":2025,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/clef","links":[{"href":"https://arxiv.org/abs/2512.02180","label":"Paper","ariaLabel":"Paper for CLEF (opens in new tab)"},{"href":"https://github.com/Nokia-Bell-Labs/ecg-foundation-model","label":"Code","ariaLabel":"Code for CLEF (opens in new tab)"}]},{"slug":"cm-ssl","name":"CM-SSL (cross-modal token-vocabulary SSL for ECG/PPG)","tagline":null,"summary":"CM-SSL trains single-signal ECG or PPG encoders against a shared token vocabulary learned from synchronized ECG and PPG pairs. A joint tokenizer is pretrained first, then the unimodal encoder, then the model is fine-tuned. Pretraining uses PulseDB, and fine-tuning covers UCI BP, BIDMC and PPG-DaLiA for PPG and CSN, PhysioNet 2017 and PTB-XL for ECG.","modalities":["PPG","ECG"],"year":null,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/cm-ssl","links":[{"href":"https://github.com/ximosss/CM-SSL","label":"Code","ariaLabel":"Code for CM-SSL (cross-modal token-vocabulary SSL for ECG/PPG) (opens in new tab)"}]},{"slug":"cogadapt","name":"CogAdapt","tagline":null,"summary":"CogAdapt adapts a clinical ECG foundation model to estimate cognitive load from wearable ECG. Its ProFine strategy unfreezes encoder layers in stages to limit drift from the pretrained model. On the CLARE and CL-Drive datasets 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.","modalities":["ECG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-26","url":"/research/models/cogadapt","links":[{"href":"https://arxiv.org/abs/2605.22774","label":"Paper","ariaLabel":"Paper for CogAdapt (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2605.22774","label":"DOI","ariaLabel":"DOI for CogAdapt (opens in new tab)"}]},{"slug":"crosshar-har","name":"CrossHAR","tagline":null,"summary":"CrossHAR is an activity recognition model built to work on new datasets without seeing them during training. It augments sensor data and then uses hierarchical self-supervised pretraining to learn general representations. Across several real-world datasets it beat previous methods by 10.83% in accuracy on unseen targets.","modalities":["IMU","ACC","gyro"],"year":2024,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/crosshar-har","links":[{"href":"https://doi.org/10.1145/3659597","label":"Paper","ariaLabel":"Paper for CrossHAR (opens in new tab)"}]},{"slug":"csfm","name":"CSFM (Cardiac Signal FM)","tagline":null,"summary":null,"modalities":["ECG","PPG"],"year":2026,"params":null,"open_weights":null,"license":"Weights: Academic Access Agreement; code snapshot on Zenodo: CC BY 4.0","updated":"2026-09-26","url":"/research/models/csfm","links":[{"href":"https://doi.org/10.1038/s42256-026-01180-5","label":"Paper","ariaLabel":"Paper for CSFM (Cardiac Signal FM) (opens in new tab)"}]},{"slug":"dna-ppg","name":"DNA-PPG","tagline":null,"summary":"DNA-PPG is a PPG foundation model pretrained with a method the authors call Dual Neighborhood Alignment, which avoids pushing apart samples that look alike and keeps physiological states continuous. It was pretrained on 10.7 million PPG segments from more than 8,400 subjects. On six benchmarks it cut regression error by 18% and improved classification by 11.5% over prior models.","modalities":["PPG"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-27","url":"/research/models/dna-ppg","links":[{"href":"https://doi.org/10.24963/ijcai.2026/777","label":"DOI","ariaLabel":"DOI for DNA-PPG (opens in new tab)"},{"href":"https://github.com/yyzyyz001/DNA-PPG","label":"Code","ariaLabel":"Code for DNA-PPG (opens in new tab)"}]},{"slug":"ecg-fm","name":"ECG-FM","tagline":null,"summary":"Toronto/UHN open transformer ECG FM (contrastive + masked recon); public-data HF weights. Backbone referenced by CogAdapt.","modalities":["ECG"],"year":2025,"params":null,"open_weights":true,"license":"MIT","updated":"2026-09-26","url":"/research/models/ecg-fm","links":[{"href":"https://arxiv.org/abs/2408.05178","label":"Paper","ariaLabel":"Paper for ECG-FM (opens in new tab)"}]},{"slug":"ecg-gpt","name":"ECG-GPT","tagline":null,"summary":null,"modalities":["ECG (image)"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/ecg-gpt","links":[]},{"slug":"ecg-ppg-jepa","name":"ECG-PPG JEPA (Bridging ECG and PPG; unnamed in paper)","tagline":null,"summary":"This KDD 2026 model is a multimodal joint-embedding predictive architecture (JEPA) for PPG and ECG, built on LeJEPA, that makes its predictions in a shared latent space instead of reconstructing raw signals. Separate encoders for each signal are pushed toward an isotropic Gaussian embedding, and random modality dropout lets the model run on PPG alone after pretraining on paired data. It improved on baselines for heart rate, atrial fibrillation and blood pressure tasks and stayed robust on lower-quality signals.","modalities":["PPG","ECG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/ecg-ppg-jepa","links":[{"href":"https://doi.org/10.1145/3770855.3819000","label":"Paper","ariaLabel":"Paper for ECG-PPG JEPA (Bridging ECG and PPG; unnamed in paper) (opens in new tab)"}]},{"slug":"ecgfm-u-lite","name":"ECGFM-U-Lite (1D-MobileViT student distilled from a unified ECG FM)","tagline":null,"summary":"ECGFM-U-Lite is a compact 1D-MobileViT student model distilled from a unified ECG foundation model, for use where compute and memory are limited. It was built with a three-stage compression procedure. It reached an overall AUC of 0.862 versus 0.918 for the teacher, keeping about 94% of the teacher's performance on both rhythm and structure tasks.","modalities":["ECG"],"year":2027,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-12","url":"/research/models/ecgfm-u-lite","links":[{"href":"https://doi.org/10.1016/j.ijmedinf.2026.106758","label":"DOI","ariaLabel":"DOI for ECGFM-U-Lite (1D-MobileViT student distilled from a unified ECG FM) (opens in new tab)"}]},{"slug":"ecgfounder","name":"ECGFounder","tagline":null,"summary":"Large supervised ECG FM on HEEDB (~10.8M ECGs / 150 diagnoses); MIT code+weights.","modalities":["ECG"],"year":2025,"params":null,"open_weights":true,"license":"MIT (Hugging Face weights)","updated":"2026-10-01","url":"/research/models/ecgfounder","links":[{"href":"https://arxiv.org/abs/2410.04133","label":"Paper","ariaLabel":"Paper for ECGFounder (opens in new tab)"},{"href":"https://github.com/PKUDigitalHealth/ECGFounder","label":"Code","ariaLabel":"Code for ECGFounder (opens in new tab)"}]},{"slug":"edgehar","name":"EdgeHAR","tagline":null,"summary":"EdgeHAR is a compact sensor foundation model for activity recognition designed to run on wearable devices rather than in the cloud. Lightweight adaptation modules help it handle new users, devices, sampling rates and sensor placements within edge limits on compute, memory and latency.","modalities":["IMU","ACC","gyro"],"year":2026,"params":"Full 1.41M / Deploy 1.086M","open_weights":false,"license":"Not stated","updated":"2026-09-29","url":"/research/models/edgehar","links":[{"href":"https://arxiv.org/abs/2609.14498","label":"Paper","ariaLabel":"Paper for EdgeHAR (opens in new tab)"}]},{"slug":"femwear","name":"FemWear","tagline":null,"summary":"FemWear repurposes a general wearable foundation model for women's health. It freezes the pretrained encoder and trains only 239,236 parameters (1.11% of a 21.54M-parameter encoder) through low-rank adapters and task heads covering menstrual, symptom, mood, sleep, autonomic, activity and pregnancy outcomes. It improved cycle-phase macro-F1 by 8.15% and cut errors for cramps and sleep problems by about 9%.","modalities":["ACC","PPG","skin_temp"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-26","url":"/research/models/femwear","links":[{"href":"https://arxiv.org/abs/2608.08244","label":"Paper","ariaLabel":"Paper for FemWear (opens in new tab)"}]},{"slug":"geenjaar-ppg-fm","name":"Geenjaar multimodal-supervision PPG FM","tagline":null,"summary":"This PPG foundation model is pretrained on ICU data and uses the accompanying ECG and respiratory signals to choose contrastive pairs, so it does not need curated or field-like PPG. Pretrained on three times fewer subjects than leading models, it improved results on 14 of 15 downstream tasks, including daily activity and heart rate prediction.","modalities":["PPG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-12","url":"/research/models/geenjaar-ppg-fm","links":[{"href":"https://arxiv.org/abs/2606.07365","label":"Paper","ariaLabel":"Paper for Geenjaar multimodal-supervision PPG FM (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2606.07365","label":"DOI","ariaLabel":"DOI for Geenjaar multimodal-supervision PPG FM (opens in new tab)"}]},{"slug":"glucofm","name":"GlucoFM","tagline":null,"summary":"GlucoFM is a lightweight foundation model for continuous glucose monitoring data. It aligns readings to a 24-hour grid and splits glucose into a slow trend and short-term deviations, and it was pretrained on 109,066 hours of CGM data from 477 subjects. Across four cohorts and seven phenotype tasks it gave the best linear-probe results, raising average PR-AUC by 4.1 points over the best CGM-specific model.","modalities":["CGM"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/glucofm","links":[{"href":"https://arxiv.org/abs/2605.30865","label":"Paper","ariaLabel":"Paper for GlucoFM (opens in new tab)"}]},{"slug":"glucoprism","name":"GlucoPRISM (protocol-supervised CGM encoder; Trait/State/Sensor bands)","tagline":null,"summary":"GlucoPRISM is a CGM encoder that splits each day's representation into person, day and sensor parts, so sensor differences do not carry over when the model moves to a new cohort. It learns this from synthetic sensor views calibrated on paired Dexcom and Libre data, without clinical labels. Dropping the sensor part raised mean cross-cohort AUROC from 68.75 to 73.57, against 69.3 for GlucoFM.","modalities":["CGM"],"year":2026,"params":null,"open_weights":true,"license":"MIT (anonymous review repo)","updated":"2026-10-12","url":"/research/models/glucoprism","links":[{"href":"https://arxiv.org/abs/2610.12528","label":"Paper","ariaLabel":"Paper for GlucoPRISM (protocol-supervised CGM encoder; Trait/State/Sensor bands) (opens in new tab)"}]},{"slug":"google-lsm","name":"Google LSM","tagline":null,"summary":"Google Large Sensor Model — MAE on ~40M hours / 165k users of minutely aggregated wearable features (not raw waveforms). Closed corpus.","modalities":["HR","ACC","skin_temp"],"year":2025,"params":null,"open_weights":false,"license":"No released weights or code","updated":"2026-09-26","url":"/research/models/google-lsm","links":[{"href":"https://arxiv.org/abs/2410.13638","label":"Paper","ariaLabel":"Paper for Google LSM (opens in new tab)"}]},{"slug":"google-lsm2","name":"Google LSM-2 / AIM","tagline":null,"summary":"LSM-2 adds Adaptive and Inherited Masking (AIM) for incomplete day-long minutely wearable grids; 25M ViT-1D; closed Google corpus.","modalities":["PPG-derived","ACC","EDA","altimeter","skin_temp"],"year":2025,"params":null,"open_weights":false,"license":"No released weights or code","updated":"2026-09-26","url":"/research/models/google-lsm2","links":[{"href":"https://arxiv.org/abs/2506.05321","label":"Paper","ariaLabel":"Paper for Google LSM-2 / AIM (opens in new tab)"}]},{"slug":"gpt-ppg","name":"GPT-PPG","tagline":null,"summary":"GPT-PPG adapts the GPT architecture to continuous PPG signals. It was pretrained on more than 200 million 30-second PPG samples and then fine-tuned for specific tasks. It matched or beat earlier methods on several tasks, notably atrial fibrillation detection, and can denoise signals without extra fine-tuning.","modalities":["PPG"],"year":2025,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/gpt-ppg","links":[{"href":"https://doi.org/10.1088/1361-6579/add988","label":"Paper","ariaLabel":"Paper for GPT-PPG (opens in new tab)"},{"href":"https://arxiv.org/abs/2503.08015","label":"arXiv","ariaLabel":"arXiv for GPT-PPG (opens in new tab)"}]},{"slug":"halo","name":"HALO","tagline":null,"summary":"Domain-specific IMU foundation model for open-vocabulary HAR across heterogeneous rates/channels/placements. Two-stage: heterogeneity-aware SSL encoder → synonym-aware soft contrastive alignment to frozen Sentence-BERT. ~35M trainable params; zero-shot = cosine retrieval over arbitrary label text (no per-dataset classifier).","modalities":["IMU","ACC","gyro","magnetometer"],"year":2026,"params":"~35M trainable (+ frozen MiniLM 22.7M shared)","open_weights":false,"license":"Not stated","updated":"2026-09-29","url":"/research/models/halo","links":[{"href":"https://arxiv.org/abs/2608.27233","label":"Paper","ariaLabel":"Paper for HALO (opens in new tab)"}]},{"slug":"harnet-ukb","name":"HARNet / OxWearables SSL","tagline":null,"summary":"HARNet is the set of self-supervised activity recognition models from Oxford's study of 700,000 person-days of UK Biobank wrist accelerometer data. They beat strong baselines on eight benchmarks with a median relative F1 gain of 24.4%. Pretrained models for 5, 10 and 30-second windows at 30 Hz can be loaded through torch.hub from the OxWearables ssl-wearables repository.","modalities":["ACC"],"year":2024,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-10-13","url":"/research/models/harnet-ukb","links":[{"href":"https://www.nature.com/articles/s41746-024-01062-3","label":"Paper","ariaLabel":"Paper for HARNet / OxWearables SSL (opens in new tab)"},{"href":"https://github.com/OxWearables/ssl-wearables","label":"Code","ariaLabel":"Code for HARNet / OxWearables SSL (opens in new tab)"}]},{"slug":"himae","name":"HiMAE","tagline":null,"summary":"HiMAE is a hierarchical masked autoencoder for wearable time series from Samsung researchers and collaborators, published at ICLR 2026. It uses a 1D convolutional encoder-decoder and is designed to find structure at different time resolutions. The repository includes a pretrained checkpoint and a linear-probe pipeline for detecting premature ventricular contractions in 10-second PPG segments, under a BSD license.","modalities":["PPG","wearable_time_series"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/himae","links":[{"href":"https://github.com/Simonlee711/HiMAE","label":"Code","ariaLabel":"Code for HiMAE (opens in new tab)"}]},{"slug":"hippt","name":"HiPPT","tagline":"wrist-PPG encoder aligned to text at record, cross-recording and generative scopes","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.","modalities":["PPG","text"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-14","url":"/research/models/hippt","links":[{"href":"https://arxiv.org/abs/2610.14489","label":"Paper","ariaLabel":"Paper for HiPPT (opens in new tab)"}]},{"slug":"hubert-ecg","name":"HuBERT-ECG","tagline":null,"summary":"Brescia HuBERT-style masked discrete-token ECG FM on 9.1M multi-country ECGs; open HF weights under CC BY-NC.","modalities":["ECG"],"year":2024,"params":null,"open_weights":true,"license":"CC BY-NC 4.0","updated":"2026-10-05","url":"/research/models/hubert-ecg","links":[{"href":"https://www.medrxiv.org/content/10.1101/2024.11.14.24317328v2","label":"Paper","ariaLabel":"Paper for HuBERT-ECG (opens in new tab)"},{"href":"https://doi.org/10.1101/2024.11.14.24317328","label":"DOI","ariaLabel":"DOI for HuBERT-ECG (opens in new tab)"},{"href":"https://github.com/Edoar-do/HuBERT-ECG","label":"Code","ariaLabel":"Code for HuBERT-ECG (opens in new tab)"}]},{"slug":"imu2clip","name":"IMU2CLIP","tagline":null,"summary":"IMU2CLIP aligns IMU motion recordings with video and text by projecting them into CLIP's joint embedding space. This lets it translate motion into text descriptions and matching video. Fine-tuned for tasks such as activity recognition, it improved downstream performance.","modalities":["IMU","ACC","gyro"],"year":2022,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-26","url":"/research/models/imu2clip","links":[{"href":"https://arxiv.org/abs/2210.14395","label":"Paper","ariaLabel":"Paper for IMU2CLIP (opens in new tab)"}]},{"slug":"imudino","name":"IMUDINO","tagline":"self-distilled foundation model for inertial sensing","summary":"DINO-style SSL adapted to inertial data via semantics-preserving motion transformations (simulated motion-pattern and sensor-setup variation); plain transformer with frequency-aware positional encoding supporting variable input length and sampling rate. Claims competitive results across HAR, health, motion capture and robot odometry (incl. cross-embodiment).","modalities":["ACC","GYRO"],"year":2027,"params":null,"open_weights":"promised","license":"Not stated","updated":"2026-10-05","url":"/research/models/imudino","links":[{"href":"https://openreview.net/forum?id=psN5UQkhct","label":"Paper","ariaLabel":"Paper for IMUDINO (opens in new tab)"}]},{"slug":"imugpt","name":"IMUGPT","tagline":null,"summary":"IMUGPT generates virtual IMU data from text, using large language models and text-to-motion synthesis, to train activity recognition models. This paper runs a large evaluation of the approach and adds a motion filter that drops irrelevant generated sequences and diversity metrics that show when to stop generating. The metrics cut generation effort by at least 50%.","modalities":["IMU","ACC","gyro"],"year":2024,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-28","url":"/research/models/imugpt","links":[{"href":"https://arxiv.org/abs/2402.01049","label":"Paper","ariaLabel":"Paper for IMUGPT (opens in new tab)"},{"href":"https://doi.org/10.1145/3678545","label":"DOI","ariaLabel":"DOI for IMUGPT (opens in new tab)"},{"href":"https://github.com/ZikangLeng/IMUGPT","label":"Code","ariaLabel":"Code for IMUGPT (opens in new tab)"}]},{"slug":"inertia-1","name":"Inertia-1","tagline":null,"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 sources around the world, the authors built a controlled framework to test how sensor type, device placement, sampling rate, window length, architecture, model size, pretraining objective and data scale affect the resulting models.","modalities":["ACC","IMU"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/inertia-1","links":[{"href":"https://arxiv.org/abs/2607.06617","label":"Paper","ariaLabel":"Paper for Inertia-1 (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2607.06617","label":"DOI","ariaLabel":"DOI for Inertia-1 (opens in new tab)"},{"href":"https://github.com/yang-ai-lab/Inertia-1","label":"Code","ariaLabel":"Code for Inertia-1 (opens in new tab)"}]},{"slug":"inrtlwrist","name":"InrtlWrist","tagline":"masked-autoencoder wrist IMU encoder (IMU-Bench reference model)","summary":"InrtlWrist is an open-weights masked-autoencoder encoder for six-channel wrist IMU data (accelerometer plus gyroscope), released by Northwestern University's SPICE Lab as the reference model of the IMU-Bench benchmark. It is a 12-layer vision-transformer-style encoder over 0.32 s patches with about 21 million parameters, pretrained on 197,094 unlabeled windows from eight public wrist datasets plus simulated wrist motion. IMU-Bench evaluates it on ten kinematic, semantic and clinical tasks, including arm and body pose, odometry, microgestures, Parkinson's disease and activity recognition. Code and weights are MIT licensed.","modalities":["ACC","GYRO"],"year":2026,"params":null,"open_weights":true,"license":"MIT (code and weights); pretraining datasets keep their own licences","updated":"2026-10-15","url":"/research/models/inrtlwrist","links":[{"href":"https://arxiv.org/abs/2610.19481","label":"Paper","ariaLabel":"Paper for InrtlWrist (opens in new tab)"},{"href":"https://github.com/SPICExLAB/IMU-Bench","label":"Code","ariaLabel":"Code for InrtlWrist (opens in new tab)"},{"href":"https://huggingface.co/YiquanLi/InrtlWrist","label":"HuggingFace","ariaLabel":"HuggingFace for InrtlWrist (opens in new tab)"}]},{"slug":"lapras","name":"Lapras (latent-reasoning post-training for time-series language models)","tagline":null,"summary":"Lapras is a post-training method that teaches time-series language models to reason in latent space instead of writing out a chain of thought. The model works through continuous thoughts in a joint time-series and language space and only produces text for the final answer, learning this by self-distillation from a teacher trained on written reasoning traces. It improved average F1 by up to 10.79% over explicit chain-of-thought while generating 23.9 times fewer tokens.","modalities":["ECG","ACC","sleep-signal","time-series","text"],"year":2026,"params":null,"open_weights":true,"license":"Code Apache-2.0; HF weights 'other' (Apache-2.0 + Llama 3.2)","updated":"2026-10-12","url":"/research/models/lapras","links":[{"href":"https://arxiv.org/abs/2610.11111","label":"Paper","ariaLabel":"Paper for Lapras (latent-reasoning post-training for time-series language models) (opens in new tab)"},{"href":"https://github.com/yuc0805/Lapras","label":"Code","ariaLabel":"Code for Lapras (latent-reasoning post-training for time-series language models) (opens in new tab)"},{"href":"https://huggingface.co/leochen085/Lapras","label":"HuggingFace","ariaLabel":"HuggingFace for Lapras (latent-reasoning post-training for time-series language models) (opens in new tab)"}]},{"slug":"lima-mamba","name":"LIMA-Mamba","tagline":"lightweight distribution-robust Mamba for PPG biometrics","summary":null,"modalities":["PPG"],"year":2027,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-08","url":"/research/models/lima-mamba","links":[{"href":"https://doi.org/10.1016/j.patcog.2026.114955","label":"DOI","ariaLabel":"DOI for LIMA-Mamba (opens in new tab)"}]},{"slug":"limu-bert","name":"LIMU-BERT","tagline":null,"summary":"LIMU-BERT is a self-supervised model for IMU data that adapts BERT-style masked training to the properties of inertial sensors. The models are lightweight enough for mobile devices, and features learned from unlabelled data let task models trained on few labels beat earlier approaches. It was evaluated on four open datasets across two typical IMU sensing applications.","modalities":["IMU","ACC","gyro"],"year":2021,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/limu-bert","links":[{"href":"https://dl.acm.org/doi/10.1145/3485730.3485937","label":"Paper","ariaLabel":"Paper for LIMU-BERT (opens in new tab)"},{"href":"https://doi.org/10.1145/3485730.3485937","label":"DOI","ariaLabel":"DOI for LIMU-BERT (opens in new tab)"}]},{"slug":"luminaecg","name":"LuminaECG","tagline":null,"summary":"LuminaECG treats ECG interpretation as a measurement-based visual reading task. A general 2B-parameter vision-language model is trained with low-rank supervised fine-tuning to link ECG measurement primitives to diagnostic reasoning, without changing its architecture.","modalities":["ECG (grid image) + text"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/luminaecg","links":[{"href":"https://arxiv.org/abs/2608.09053","label":"Paper","ariaLabel":"Paper for LuminaECG (opens in new tab)"}]},{"slug":"m2ae","name":"M2AE (Biosignal Fingerprinting)","tagline":null,"summary":"M2AE is a multimodal masked autoencoder trained on more than 3.4 million paired ECG and PPG signals, whose compact embeddings the authors call biosignal fingerprints. Used frozen across 7 tasks, from cross-modal reconstruction to disease classification, hypertension detection, mortality and demographics, it matched or beat specialist foundation models. It reached an AUROC of 0.974 for five-class cardiovascular disease classification and 0.877 for hypertension detection.","modalities":["PPG","ECG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-28","url":"/research/models/m2ae","links":[{"href":"https://arxiv.org/abs/2605.09579","label":"Paper","ariaLabel":"Paper for M2AE (Biosignal Fingerprinting) (opens in new tab)"}]},{"slug":"mira","name":"MIRA","tagline":null,"summary":"Microsoft Research medical irregular time-series FM (CT-RoPE + frequency MoE + Neural ODE); NeurIPS 2025; MIT code; Large weights on HF.","modalities":["ECG","EEG","PPG","clinical_vitals","medical_time_series"],"year":2025,"params":null,"open_weights":true,"license":"MIT (code)","updated":"2026-09-27","url":"/research/models/mira","links":[{"href":"https://arxiv.org/abs/2506.07584","label":"Paper","ariaLabel":"Paper for MIRA (opens in new tab)"},{"href":"https://github.com/microsoft/MIRA","label":"Code","ariaLabel":"Code for MIRA (opens in new tab)"}]},{"slug":"ms-ecg-fm","name":"MS-ECG-FM","tagline":"multi-source contrastive ECG foundation model (Apple)","summary":"A transformer ECG encoder with a deep shared per-lead convolutional tokenizer (0.5 s patches, 12 × 20 = 240 tokens for a 10 s 12-lead ECG), learnable lead embeddings and 1D sinusoidal time encoding. It is built to transfer to arbitrary lead subsets (Lead I-only is called out as the wearable case). Instead of aligning ECGs only to machine interpretation reports (MERL, D-BETA, MELP), it contrastively aligns each ECG to five clinical text sources: ECG machine reports, cardiologist ECG reads, echo reports, chest X-ray reports and discharge summaries.","modalities":["ECG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-07","url":"/research/models/ms-ecg-fm","links":[{"href":"https://arxiv.org/abs/2610.07662","label":"Paper","ariaLabel":"Paper for MS-ECG-FM (opens in new tab)"}]},{"slug":"msma-ppg","name":"MSMA-PPG","tagline":"channel-aligned expert routing for selected-wavelength PPG","summary":"Wavelength-specific projections + residual adapters on a shared optical temporal backbone; full-to-subset alignment links paired training views; Channel-Aligned MoE (availability-masked top-2 routing) over optical, motion and fused experts. Evaluated HR, BP, AF single-stream transfer. Reported (routing disabled, PPG-only, seven-model comparison): lowest participant-macro HR MAE on EarSet 5.020 bpm and PPG-DaLiA 9.246 bpm; lowest SBP MAE 15.456 mmHg and second-best AF AUPRC/AUROC 0.5177/0.8909.","modalities":["PPG-multiwavelength","ACC"],"year":2027,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-05","url":"/research/models/msma-ppg","links":[{"href":"https://openreview.net/forum?id=ewFflGoy46","label":"Paper","ariaLabel":"Paper for MSMA-PPG (opens in new tab)"}]},{"slug":"neurorvq","name":"NeuroRVQ (PPG tokenizer)","tagline":null,"summary":"Multi-scale residual-vector-quantization (RVQ) biosignal tokenizer from Imperial College / Cogitat. The PPG tokenizer checkpoint is now packaged in braindecode (HF repo created 2026-10-08), which makes it a loadable, open-weight (non-commercial) PPG codec.","modalities":["PPG","EEG","ECG"],"year":2025,"params":null,"open_weights":true,"license":"CC BY-NC 4.0","updated":"2026-10-08","url":"/research/models/neurorvq","links":[{"href":"https://arxiv.org/abs/2510.13068","label":"Paper","ariaLabel":"Paper for NeuroRVQ (PPG tokenizer) (opens in new tab)"},{"href":"https://github.com/KonstantinosBarmpas/NeuroRVQ","label":"Code","ariaLabel":"Code for NeuroRVQ (PPG tokenizer) (opens in new tab)"}]},{"slug":"normwear","name":"NormWear","tagline":null,"summary":"Channel-agnostic MAE for heterogeneous wearable physio channels across consumer devices; public-data pretrain. Successor world-model: normwear-2.","modalities":["PPG","ECG","EEG","GSR","IMU"],"year":2025,"params":null,"open_weights":true,"license":"Apache-2.0","updated":"2026-10-06","url":"/research/models/normwear","links":[{"href":"https://arxiv.org/abs/2412.09758","label":"Paper","ariaLabel":"Paper for NormWear (opens in new tab)"},{"href":"https://huggingface.co/mosaic-laboratory/normwear","label":"HuggingFace","ariaLabel":"HuggingFace for NormWear (opens in new tab)"}]},{"slug":"normwear-2","name":"NormWear-2","tagline":null,"summary":"NormWear-2 is a world model for physiological signals that puts multivariate signals and clinical interventions into one latent space and models how they evolve together over time. It combines pretrained knowledge with fast non-parametric adaptation of latent state transitions, so it can forecast signals over several time scales under different interventions.","modalities":["PPG","ECG","EEG","respiration","HR","glucose","BP","skin_temp","IMU_derived"],"year":2026,"params":"~0.2B","open_weights":true,"license":"Not stated","updated":"2026-10-06","url":"/research/models/normwear-2","links":[{"href":"https://arxiv.org/abs/2605.15465","label":"Paper","ariaLabel":"Paper for NormWear-2 (opens in new tab)"},{"href":"https://huggingface.co/mosaic-laboratory/normwear2","label":"HuggingFace","ariaLabel":"HuggingFace for NormWear-2 (opens in new tab)"}]},{"slug":"opensla","name":"OpenSLA","tagline":"Sensor-Language-Action model (UCLA, Yuzhe Yang lab)","summary":"Existing sensor models however largely stop at perception: they recognize states or predict outcomes, leaving actions modeled separately through task-specific and often closed label spaces. The authors introduce Sensor-Language-Action (SLA) modeling, a framework that connects multimodal sensor observations, natural language, and actions within a unified model. SLA uses language as a semantic interface between sensing and acting, allowing heterogeneous actions to be represented, predicted, and explained while remaining grounded in the underlying sensor evidence. Building on this framework, the authors present OpenSLA, a unified SLA model for hierarchical action prediction, state understanding, and action explanation.","modalities":["ECG","PPG","RESP","ABP","CO2","EEG","CGM","vitals","labs","text"],"year":2026,"params":null,"open_weights":false,"license":"MIT; Qwen3.5-2B base model terms apply to the LLM","updated":"2026-10-08","url":"/research/models/opensla","links":[{"href":"https://arxiv.org/abs/2610.08244","label":"Paper","ariaLabel":"Paper for OpenSLA (opens in new tab)"},{"href":"https://github.com/yang-ai-lab/OpenSLA","label":"Code","ariaLabel":"Code for OpenSLA (opens in new tab)"},{"href":"https://huggingface.co/yang-ai-lab/OpenSLA","label":"HuggingFace","ariaLabel":"HuggingFace for OpenSLA (opens in new tab)"}]},{"slug":"osf-sleep-fm","name":"OSF (Open Sleep Foundation Models)","tagline":null,"summary":"OSF is a family of open sleep foundation models built to study how pretraining choices and scale affect generalization across polysomnography devices and cohorts. The authors assembled 166,500 hours of sleep recordings from nine public sources and released SleepBench, a fully open benchmark.","modalities":["EEG","EOG","EMG","respiratory_effort","SpO2"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/osf-sleep-fm","links":[{"href":"https://arxiv.org/abs/2603.00190","label":"Paper","ariaLabel":"Paper for OSF (Open Sleep Foundation Models) (opens in new tab)"},{"href":"https://github.com/yang-ai-lab/OSF-Open-Sleep-FM","label":"Code","ariaLabel":"Code for OSF (Open Sleep Foundation Models) (opens in new tab)"}]},{"slug":"papagei-s","name":"PaPaGei-S","tagline":null,"summary":"Nokia Bell Labs open PPG FM (compact PaPaGei-S ~5.7M) with morphology-aware contrastive SSL.","modalities":["PPG"],"year":2025,"params":null,"open_weights":true,"license":"BSD-3-Clause (code)","updated":"2026-10-12","url":"/research/models/papagei-s","links":[{"href":"https://arxiv.org/abs/2410.20542","label":"Paper","ariaLabel":"Paper for PaPaGei-S (opens in new tab)"},{"href":"https://github.com/nokia-bell-labs/papagei-foundation-model","label":"Code","ariaLabel":"Code for PaPaGei-S (opens in new tab)"}]},{"slug":"pat-actigraphy-transformer","name":"PAT (Pretrained Actigraphy Transformer)","tagline":null,"summary":"PAT (Pretrained Actigraphy Transformer) is an open-source foundation model for wearable movement data in mental health research. It uses transformers with patch embeddings, pretrained as a masked autoencoder on week-long minute-level actigraphy from 21,538 NHANES participants. PAT beat non-foundation baselines on predicting benzodiazepine and SSRI use, depression and sleep problems, including a 55.6% gain over an LSTM for benzodiazepine use.","modalities":["ACC (actigraphy counts)"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/pat-actigraphy-transformer","links":[{"href":"https://github.com/njacobsonlab/Pretrained-Actigraphy-Transformer/","label":"Code","ariaLabel":"Code for PAT (Pretrained Actigraphy Transformer) (opens in new tab)"}]},{"slug":"physioomni","name":"PhysioOmni","tagline":null,"summary":"PhysioOmni is a foundation model for multimodal physiological signals that still works when some modalities are missing. It trains a decoupled tokenizer and pretrains with masked-signal objectives that separate shared and modality-specific features, then fine-tunes with prototype alignment so it adapts to whatever combination of signals a dataset has.","modalities":["EEG","ECG","EOG","EMG"],"year":2025,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-27","url":"/research/models/physioomni","links":[{"href":"https://arxiv.org/abs/2504.19596","label":"Paper","ariaLabel":"Paper for PhysioOmni (opens in new tab)"},{"href":"https://github.com/935963004/PhysioOmni","label":"Code","ariaLabel":"Code for PhysioOmni (opens in new tab)"}]},{"slug":"ppg-lm","name":"PPG-LM","tagline":null,"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 in LLMs each miss part of the picture, and PPG-LM combines the two.","modalities":["PPG","EHR_text"],"year":2026,"params":"Waveform encoder ~12.7M (Appendix C.1); full family not given as a single count","open_weights":false,"license":"Not stated","updated":"2026-09-29","url":"/research/models/ppg-lm","links":[{"href":"https://arxiv.org/abs/2609.33516","label":"Paper","ariaLabel":"Paper for PPG-LM (opens in new tab)"}]},{"slug":"protomm","name":"ProtoMM (Pulse-Motion)","tagline":null,"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 also help interpret the learned features.","modalities":["PPG","ACC"],"year":2025,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/protomm","links":[{"href":"https://arxiv.org/abs/2510.09764","label":"Paper","ariaLabel":"Paper for ProtoMM (Pulse-Motion) (opens in new tab)"}]},{"slug":"pulse-ppg","name":"Pulse-PPG","tagline":null,"summary":"Open field-trained wrist PPG foundation model (ResNet-26, 28.5M) from uncurated MOODS smartwatch PPG.","modalities":["PPG"],"year":2025,"params":null,"open_weights":true,"license":"MIT (code and weights)","updated":"2026-10-11","url":"/research/models/pulse-ppg","links":[{"href":"https://arxiv.org/abs/2502.01108","label":"Paper","ariaLabel":"Paper for Pulse-PPG (opens in new tab)"},{"href":"https://doi.org/10.1145/3749494","label":"DOI","ariaLabel":"DOI for Pulse-PPG (opens in new tab)"},{"href":"https://github.com/maxxu05/pulseppg","label":"Code","ariaLabel":"Code for Pulse-PPG (opens in new tab)"}]},{"slug":"pulselm","name":"PulseLM","tagline":null,"summary":"Foundation dataset + PPG–text QA benchmark (not a new PPG encoder). Aggregates 16 public PPG sources into ~1.08M 10 s / 125 Hz segments and ~2.48M closed-ended QA pairs across 12 tasks; baselines couple pretrained PPG encoders (PaPaGei / Pulse-PPG) to LoRA-adapted LLMs.","modalities":["PPG","text"],"year":2026,"params":null,"open_weights":false,"license":"CC BY 4.0 (code repository)","updated":"2026-09-27","url":"/research/models/pulselm","links":[{"href":"https://arxiv.org/abs/2603.03331","label":"Paper","ariaLabel":"Paper for PulseLM (opens in new tab)"},{"href":"https://doi.org/10.48550/arXiv.2603.03331","label":"DOI","ariaLabel":"DOI for PulseLM (opens in new tab)"},{"href":"https://github.com/manhph2211/PulseLM","label":"Code","ariaLabel":"Code for PulseLM (opens in new tab)"},{"href":"https://huggingface.co/datasets/Manhph2211/PulseLM","label":"HuggingFace","ariaLabel":"HuggingFace for PulseLM (opens in new tab)"}]},{"slug":"pulseox-fm","name":"PulseOx-FM","tagline":"overnight pulse-oximetry foundation model (Human Phenotype Project)","summary":"PulseOx-FM is a self-supervised foundation model trained on 6,995,558 pulse oximetry segments from 42,282 overnight sleep recordings of 10,704 people in the Human Phenotype Project. It predicted chronological age better than existing feature extraction methods and held up in an external cohort. Its representations predicted 64 phenotypes across cardiometabolic and neuropsychiatric domains and flagged two-year hypertension risk in people with normal blood pressure.","modalities":["PPG (pulse oximetry)","SpO2"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-10-11","url":"/research/models/pulseox-fm","links":[{"href":"https://doi.org/10.64898/2026.07.01.26356992","label":"DOI","ariaLabel":"DOI for PulseOx-FM (opens in new tab)"}]},{"slug":"rae-ppg","name":"RAE-PPG","tagline":null,"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 step. Features learned early stay recoverable from the longer-input representations, and learning earlier stages first made later features easier to learn.","modalities":["PPG"],"year":2026,"params":"Exported encoder ~10.76M","open_weights":false,"license":"Not stated","updated":"2026-09-30","url":"/research/models/rae-ppg","links":[{"href":"https://arxiv.org/abs/2609.36794","label":"Paper","ariaLabel":"Paper for RAE-PPG (opens in new tab)"}]},{"slug":"relcon","name":"RelCon","tagline":null,"summary":"Apple/UIUC relative-contrastive wrist motion FM (1D ResNet-34 ~4M) on ~87k participants; MIT code.","modalities":["ACC"],"year":2025,"params":null,"open_weights":false,"license":"MIT (code)","updated":"2026-09-26","url":"/research/models/relcon","links":[{"href":"https://arxiv.org/abs/2411.18822","label":"Paper","ariaLabel":"Paper for RelCon (opens in new tab)"},{"href":"https://github.com/maxxu05/relcon","label":"Code","ariaLabel":"Code for RelCon (opens in new tab)"},{"href":"https://openreview.net/forum?id=k2uUeLCrQq","label":"OpenReview","ariaLabel":"OpenReview for RelCon (opens in new tab)"}]},{"slug":"rl-ppg","name":"RL-PPG (Real-Life PPG encoder; Emognition)","tagline":null,"summary":"RL-PPG tests self-supervised PPG representations for detecting intense emotions in daily life. The representations 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 under the same protocol. Using a person's own data during fine-tuning mattered more than population-level pretraining.","modalities":["PPG"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-10-11","url":"/research/models/rl-ppg","links":[{"href":"https://arxiv.org/abs/2608.14675","label":"arXiv","ariaLabel":"arXiv for RL-PPG (Real-Life PPG encoder; Emognition) (opens in new tab)"},{"href":"https://github.com/Emognition/RL-PPG","label":"Code","ariaLabel":"Code for RL-PPG (Real-Life PPG encoder; Emognition) (opens in new tab)"}]},{"slug":"samsung-mmr","name":"Samsung MMR","tagline":null,"summary":"Samsung's Masked Multiscale Reconstruction (MMR) is a self-supervised pretraining method for PPG that learns from several time-frequency scales, from fine waveform shape to overall rhythm. Trained on large-scale wearable PPG, it matched or beat open-source PPG foundation models, time-series foundation models and other self-supervised baselines on 17 of 19 health tasks.","modalities":["PPG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/samsung-mmr","links":[{"href":"https://arxiv.org/html/2601.12215v1","label":"Paper","ariaLabel":"Paper for Samsung MMR (opens in new tab)"},{"href":"https://arxiv.org/abs/2601.12215","label":"arXiv","ariaLabel":"arXiv for Samsung MMR (opens in new tab)"}]},{"slug":"sensorfm","name":"SensorFM (Google)","tagline":null,"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 covering cardiovascular, metabolic, sleep and mental health, lifestyle and demographics.","modalities":["PPG-derived","ACC","EDA","skin_temp","altimeter"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/sensorfm","links":[{"href":"https://arxiv.org/abs/2605.22759","label":"Paper","ariaLabel":"Paper for SensorFM (Google) (opens in new tab)"}]},{"slug":"sensori","name":"Sensori","tagline":null,"summary":"Sensori is a self-supervised foundation model that learns health representations from 24 hours of raw wrist accelerometer data. It was built and tested on four population cohorts from the UK, China and the United States, totalling 122,640 people and 683,617 person-days of free-living data. Added to standard clinical covariates, it improved disease classification for 52 of 102 conditions.","modalities":["ACC"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-27","url":"/research/models/sensori","links":[{"href":"https://arxiv.org/abs/2608.29494","label":"Paper","ariaLabel":"Paper for Sensori (opens in new tab)"},{"href":"https://github.com/OxWearables/Sensori","label":"Code","ariaLabel":"Code for Sensori (opens in new tab)"}]},{"slug":"sensorlm","name":"SensorLM","tagline":null,"summary":"Google sensor-language model aligning wearable streams with hierarchical natural-language captions; open code path under Google-Health consumer-health-research.","modalities":["PPG","ACC","skin_temp","HR","SpO2"],"year":2025,"params":null,"open_weights":false,"license":"No released weights","updated":"2026-09-26","url":"/research/models/sensorlm","links":[{"href":"https://arxiv.org/abs/2506.09108","label":"Paper","ariaLabel":"Paper for SensorLM (opens in new tab)"},{"href":"https://github.com/Google-Health/consumer-health-research/tree/main/sensorlm","label":"Code","ariaLabel":"Code for SensorLM (opens in new tab)"}]},{"slug":"siamquality","name":"SiamQuality","tagline":null,"summary":null,"modalities":["PPG"],"year":2024,"params":null,"open_weights":false,"license":"No license granted for the code","updated":"2026-10-11","url":"/research/models/siamquality","links":[{"href":"https://arxiv.org/abs/2404.17667","label":"Paper","ariaLabel":"Paper for SiamQuality (opens in new tab)"}]},{"slug":"sigma-ppg","name":"SIGMA-PPG","tagline":null,"summary":"SIGMA-PPG is a generative PPG foundation model that uses a reinforcement-learning teacher to choose difficult masks based on statistical priors, plus vector quantization so that similar waveforms map to the same codes. Pretrained on more than 120,000 hours of PPG, it beat five state-of-the-art baselines on average across 12 downstream tasks. Code is on GitHub and the checkpoint is on Hugging Face.","modalities":["PPG"],"year":null,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/sigma-ppg","links":[{"href":"https://github.com/ZonghengGuo/SigmaPPG","label":"Code","ariaLabel":"Code for SIGMA-PPG (opens in new tab)"}]},{"slug":"sleepfm-v1","name":"SleepFM (v1)","tagline":null,"summary":null,"modalities":["ECG","EEG","EMG","respiratory"],"year":2024,"params":null,"open_weights":null,"license":"MIT (code)","updated":"2026-09-26","url":"/research/models/sleepfm-v1","links":[{"href":"https://arxiv.org/abs/2405.17766","label":"Paper","ariaLabel":"Paper for SleepFM (v1) (opens in new tab)"}]},{"slug":"sleepfm-2","name":"SleepFM-2","tagline":null,"summary":"SleepFM-2 is a sleep foundation model built on 282,511 polysomnography recordings from 26 cohorts, 235,865 of them used for pretraining. It improves on SleepFM for disease prediction and sleep scoring, adds arousal, limb movement and respiratory event detection, and transfers to wearables. Combined with age, sex and BMI, it met a prespecified criterion for 188 later EHR phenotypes in two held-out cohorts.","modalities":["EEG","EOG","EMG","ECG","respiration","PPG","ACC"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-26","url":"/research/models/sleepfm-2","links":[{"href":"https://arxiv.org/abs/2609.06849","label":"Paper","ariaLabel":"Paper for SleepFM-2 (opens in new tab)"}]},{"slug":"sleeplm","name":"SleepLM","tagline":null,"summary":"SleepLM is a family of sleep-language foundation models that connect sleep signals with natural language. The authors built a sleep-text dataset of more than 100K hours from over 10,000 people with a multilevel captioning pipeline, and trained with contrastive alignment, caption generation and signal reconstruction. It beat previous methods on zero-shot and few-shot tasks, retrieval and sleep captioning.","modalities":["EEG","EOG","EMG","respiratory_effort","SpO2 + text"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/sleeplm","links":[{"href":"https://arxiv.org/abs/2602.23605","label":"Paper","ariaLabel":"Paper for SleepLM (opens in new tab)"},{"href":"https://github.com/yang-ai-lab/SleepLM","label":"Code","ariaLabel":"Code for SleepLM (opens in new tab)"}]},{"slug":"slip","name":"SLIP (Sensor Language-Informed Pretraining; any-channel sensor-language FM)","tagline":null,"summary":"SLIP is an open-source framework for language-aligned sensor representations that work across different sensor setups. It reuses a pretrained decoder-only language model through cross-attention and uses a flexible patch embedder, so it handles new sensor configurations, sampling rates and lengths without retraining. Across 11 datasets it reached 77.14% average linear-probe accuracy, 3.56 points above SensorLM.","modalities":["IMU","PPG","ECG","EEG","time-series","text"],"year":2026,"params":null,"open_weights":true,"license":"MIT (code + HF weights)","updated":"2026-10-12","url":"/research/models/slip","links":[{"href":"https://arxiv.org/abs/2603.11950","label":"Paper","ariaLabel":"Paper for SLIP (Sensor Language-Informed Pretraining; any-channel sensor-language FM) (opens in new tab)"},{"href":"https://github.com/yuc0805/SLIP","label":"Code","ariaLabel":"Code for SLIP (Sensor Language-Informed Pretraining; any-channel sensor-language FM) (opens in new tab)"},{"href":"https://huggingface.co/leochen085/SLIP","label":"HuggingFace","ariaLabel":"HuggingFace for SLIP (Sensor Language-Informed Pretraining; any-channel sensor-language FM) (opens in new tab)"}]},{"slug":"soter","name":"SOTER","tagline":null,"summary":"SOTER is a generative foundation model for wearable physiological time series. It combines a backbone that models links between channels, a mixture-of-experts layer that routes by fixed spectral bands, and a neural controlled differential equation decoder that predicts and imputes at any timestamp. It had the best zero-shot forecasting error on most of six datasets and the lowest imputation error on all six at 75% missing data.","modalities":["PPG","ECG","EDA","BVP","EEG","wearable_physiological_multichannel"],"year":2026,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-27","url":"/research/models/soter","links":[{"href":"https://arxiv.org/abs/2609.16804","label":"Paper","ariaLabel":"Paper for SOTER (opens in new tab)"},{"href":"https://github.com/SII-fkchen/SOTER","label":"Code","ariaLabel":"Code for SOTER (opens in new tab)"}]},{"slug":"spotr","name":"SPOTR","tagline":"single-token compress-reconstruct SSL for EEG/iEEG/ECG/PPG (IJCAI-ECAI 2026)","summary":"SPOTR is a pretraining method for physiological signals that compresses each recording into a single global token and reconstructs from it, with a compaction module to save compute. Pretrained on 20 datasets of EEG, iEEG, ECG and PPG, it beat the strongest baseline in linear probing by 18.49%, 21.71%, 17.86% and 4.64% AUC respectively. It also ran with about 78% lower latency and 52% less GPU memory than a general time-series foundation model.","modalities":["EEG","iEEG","ECG","PPG"],"year":2026,"params":null,"open_weights":null,"license":"MIT","updated":"2026-10-11","url":"/research/models/spotr","links":[{"href":"https://arxiv.org/abs/2606.21973","label":"Paper","ariaLabel":"Paper for SPOTR (opens in new tab)"},{"href":"https://github.com/5GYYYYY/SPOTR","label":"Code","ariaLabel":"Code for SPOTR (opens in new tab)"}]},{"slug":"stepfm","name":"StepFM","tagline":null,"summary":"StepFM is a foundation model built only on step-counter data for broad health prediction. Its pretraining captures temporal and behavioural patterns from large-scale step sequences, and it transfers to more than 20 health risk prediction tasks across different devices, regions and diseases.","modalities":["step_count (IMU-derived)"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-09-26","url":"/research/models/stepfm","links":[{"href":"https://arxiv.org/abs/2607.06954","label":"Paper","ariaLabel":"Paper for StepFM (opens in new tab)"}]},{"slug":"synhar","name":"SynHAR","tagline":null,"summary":"SynHAR creates synthetic IMU data from personalized biomechanical models and human body surface models to improve activity recognition. It was used both mixed with real data and for pretraining before transfer to public benchmarks, with three HAR models. Synthetic data raised macro-F1 by 8% on average when real data was limited, and pretraining with synthetic plus real data beat real data alone by 5.2% on average.","modalities":["IMU","ACC","gyro"],"year":2024,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-28","url":"/research/models/synhar","links":[{"href":"https://doi.org/10.1109/ACCESS.2024.3513477","label":"DOI","ariaLabel":"DOI for SynHAR (opens in new tab)"}]},{"slug":"tex-mex","name":"TEX-MEX","tagline":"similarity-guided temporal contrastive SSL; 560K-param PPG foundation model","summary":"TEX-MEX is a self-supervised contrastive objective for time series that unifies structural similarity (windows with similar morphology) and temporal similarity (windows close together in time) into a single relationship. Within each batch the encoder identifies structurally similar and dissimilar windows, then pulls each window toward the temporally shifted versions of similar windows and pushes it away from those of dissimilar ones. Across PPG, EEG, ECG, accelerometry and continuous glucose monitoring it broadly matches or beats contrastive, reconstructive, JEPA-based and modality-specific baselines, and the authors train a 560K-parameter PPG foundation model that is roughly 50x smaller than the next-best unimodally pretrained PPG foundation model in their evaluation.","modalities":["PPG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-15","url":"/research/models/tex-mex","links":[{"href":"https://arxiv.org/abs/2610.19557","label":"Paper","ariaLabel":"Paper for TEX-MEX (opens in new tab)"}]},{"slug":"text2imu","name":"Text2IMU","tagline":null,"summary":"Text2IMU generates synthetic accelerometer and gyroscope data from text descriptions of activities. A motion synthesis model turns prompts into 3D movement, which body surface models of different shapes perform while virtual IMUs record. Without measured data from the target domain, it reached 79.2% average activity recognition accuracy across three public datasets for correctly generated activities, about double the baseline synthetic data.","modalities":["IMU","ACC","gyro"],"year":2025,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-28","url":"/research/models/text2imu","links":[{"href":"https://doi.org/10.1109/BSN66969.2025.11337798","label":"DOI","ariaLabel":"DOI for Text2IMU (opens in new tab)"}]},{"slug":"tf-mambanet","name":"TF-MambaNet","tagline":"bidirectional-Mamba PPG foundation model (ICASSP 2026)","summary":"Self-supervised PPG FM used as a frozen feature extractor. Across 15 tasks on 8 public datasets it is best on 9/15 against Chronos, PaPaGei and NormWear (authors' re-runs). Strongest on HR (DaLiA MAE 7.27), BP (PPG-BP SBP 12.89) and AF (MIMIC 0.83 acc).","modalities":["PPG"],"year":2026,"params":"≈1.5M params (vs PaPaGei 5.7M, NormWear 135.3M, Chronos 200M)","open_weights":false,"license":"Not stated","updated":"2026-10-08","url":"/research/models/tf-mambanet","links":[{"href":"https://doi.org/10.1109/ICASSP55912.2026.11461042","label":"DOI","ariaLabel":"DOI for TF-MambaNet (opens in new tab)"}]},{"slug":"unimts","name":"UniMTS","tagline":null,"summary":"UniMTS is a pretraining method for motion time series from phones and smartwatches that generalizes across device position, orientation and activity type. Tested on 18 motion classification benchmarks, it beat the best baselines by 340% in zero-shot, 16.3% in few-shot and 9.2% in full-shot settings.","modalities":["IMU","ACC","gyro"],"year":2024,"params":null,"open_weights":true,"license":"Not stated","updated":"2026-09-26","url":"/research/models/unimts","links":[{"href":"https://arxiv.org/abs/2410.19818","label":"Paper","ariaLabel":"Paper for UniMTS (opens in new tab)"},{"href":"https://github.com/xiyuanzh/UniMTS","label":"Code","ariaLabel":"Code for UniMTS (opens in new tab)"}]},{"slug":"varmae","name":"VarMAE","tagline":"variable-input masked autoencoder for wrist-worn accelerometry","summary":"Two views per signal (long/high-rate and short/low-rate via crop + downsample); masked reconstruction within each view plus a cross-resolution objective (long-view visible tokens predict short-view masked tokens). Matches or beats controlled FM baselines on eight HAR datasets (OpenWatch, WatchHand, MotionSense, MobiAct, HARTH, RealWorld, PAAWS-SimFL, Capture-24) with one encoder. Sizes: Small 0.7M params; scaled 3.82M; Medium/Large variants in the paper.","modalities":["ACC"],"year":2027,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-14","url":"/research/models/varmae","links":[{"href":"https://openreview.net/forum?id=otzhV2A17M","label":"Paper","ariaLabel":"Paper for VarMAE (opens in new tab)"},{"href":"https://arxiv.org/abs/2610.17383","label":"arXiv","ariaLabel":"arXiv for VarMAE (opens in new tab)"}]},{"slug":"vitalagent","name":"VitalAgent","tagline":"tool-augmented ECG/PPG monitoring agent (Univ. of Melbourne, Ting Dang group)","summary":"Wearable devices enable continuous monitoring of physiological signals such as ECG and PPG, but existing mHealth systems are largely limited to task-specific prediction pipelines or reactive question answering over static summaries. The authors propose VitalAgent, a tool-augmented agentic framework for ECG/PPG-based mHealth that supports both reactive question answering and proactive monitoring. The authors further introduce VitalBench, a longitudinal physiological monitoring benchmark dataset comprising 1,862 QA pairs for reactive question answering and 90.2 hours of continuous ECG/PPG recordings for proactive monitoring, covering cardiac, physical activity, and stress-related tasks.","modalities":["ECG","PPG","text"],"year":2026,"params":null,"open_weights":false,"license":"MIT (code); CC BY-NC-SA 4.0 (VitalBench data)","updated":"2026-10-10","url":"/research/models/vitalagent","links":[{"href":"https://arxiv.org/abs/2605.29483","label":"Paper","ariaLabel":"Paper for VitalAgent (opens in new tab)"},{"href":"https://github.com/DKSEZD/VitalAgent","label":"Code","ariaLabel":"Code for VitalAgent (opens in new tab)"}]},{"slug":"wavesfm","name":"WavesFM","tagline":"hierarchical longitudinal FM for raw wearable PPG + accelerometer waveforms","summary":"WavesFM is a foundation model for longitudinal physiological data that uses two stages of self-supervised learning. It aims to keep both fine waveform detail and long-term behavioural patterns over weeks of recordings, which earlier methods usually traded off.","modalities":["PPG","accelerometer"],"year":2026,"params":null,"open_weights":null,"license":"Not stated","updated":"2026-10-11","url":"/research/models/wavesfm","links":[{"href":"https://arxiv.org/abs/2605.09173","label":"Paper","ariaLabel":"Paper for WavesFM (opens in new tab)"}]},{"slug":"whubert","name":"wHuBERT","tagline":"HuBERT-style Conformer for 25 Hz wearable PPG + accelerometry","summary":"Iterative HuBERT masked prediction (cluster pseudo-labels → second iteration on intermediate-layer clusters). Paper proposes a cheap probe-based layer-selection protocol at regular checkpoints; best sleep-staging layer sits just below the top, rarely the final layer; 7 label-free stats (incl. effective rank/RankMe) unreliable for selection.","modalities":["PPG","ACC"],"year":2027,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-05","url":"/research/models/whubert","links":[{"href":"https://openreview.net/forum?id=u3iC4BjfaC","label":"Paper","ariaLabel":"Paper for wHuBERT (opens in new tab)"}]},{"slug":"wonderwall","name":"Wonderwall","tagline":null,"summary":"Wonderwall is an activity recognition foundation model trained first on physics-based simulated motion and then on curated real datasets. Tested on ten activity recognition benchmarks, it gave state-of-the-art generalization in zero-shot, few-shot and full-shot settings.","modalities":["IMU","ACC","gyro"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-09-29","url":"/research/models/wonderwall","links":[{"href":"https://doi.org/10.1145/3789688","label":"Paper","ariaLabel":"Paper for Wonderwall (opens in new tab)"}]},{"slug":"xecg","name":"xECG","tagline":"bidirectional xLSTM ECG foundation model (BenchECG)","summary":"ECG FM pretrained with SimDINOv2 on CODE + Chapman/Ningbo + INCART. It ranks best overall on BenchECG (published score 0.838) and is the strongest of five ECG FMs on DeepBeat wrist-PPG AF transfer (FT AUROC 0.883).","modalities":["ECG"],"year":2026,"params":"57.0M params; patch 25 at 100 Hz; embedding 1024","open_weights":true,"license":"MIT","updated":"2026-10-16","url":"/research/models/xecg","links":[{"href":"https://arxiv.org/abs/2509.10151","label":"Paper","ariaLabel":"Paper for xECG (opens in new tab)"},{"href":"https://doi.org/10.1038/s41746-026-03196-y","label":"DOI","ariaLabel":"DOI for xECG (opens in new tab)"},{"href":"https://github.com/dlaskalab/bench-xecg","label":"Code","ariaLabel":"Code for xECG (opens in new tab)"},{"href":"https://huggingface.co/riccardolunelli/xECG_base_model_v1","label":"HuggingFace","ariaLabel":"HuggingFace for xECG (opens in new tab)"}]},{"slug":"xmae","name":"xMAE","tagline":null,"summary":"xMAE is a biosignal pretraining method that reconstructs one signal from another, for example ECG and the delayed PPG pulse it produces, so the model learns physiological timing. Its representations beat unimodal and multimodal baselines on 15 of 19 tasks, including cardiovascular outcomes, abnormal lab tests, sleep staging and demographics, across different devices and body locations.","modalities":["ECG","PPG"],"year":2026,"params":null,"open_weights":false,"license":"Not stated","updated":"2026-10-11","url":"/research/models/xmae","links":[{"href":"https://arxiv.org/abs/2605.00973","label":"Paper","ariaLabel":"Paper for xMAE (opens in new tab)"},{"href":"https://github.com/hzhou3/xMAE","label":"Code","ariaLabel":"Code for xMAE (opens in new tab)"}]}]}