Models, datasets
& papers.

code release

egoPPG estimates heart rate from eye-tracking cameras in egocentric systems (ICCV 2025 paper lineage). Official code is at github.com/eth-siplab/egoPPG; no pretrained weights were published in the recorded release.

NMF multidimensional attention 3D-CNN for rPPG (NeurIPS 2024)

Bias-aware fine-tuning of PPG-GPT for heart-rate prediction across ICU, wrist, and smartphone PPG

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%.

feature-level optimal warping for domain-generalized rPPG

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.

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.

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.

Google Large Sensor Model — MAE on ~40M hours / 165k users of minutely aggregated wearable features (not raw waveforms). Closed corpus.

LSM-2 adds Adaptive and Inherited Masking (AIM) for incomplete day-long minutely wearable grids; 25M ViT-1D; closed Google corpus.

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.

Keras deep-learning models for 3-class activity intensity from wrist accelerometry at 12.5 Hz, developed on 30 healthy and 30 chronic-pain participants plus CAPTURE-24 (Doomen, Felius et al., Frontiers in Digital Health 2026). Of four DL candidates (CNN, LSTM, CNN-LSTM, larger CNN) the authors chose a CNN. The repo ships weights for all candidates plus an HMM.

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).

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.

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.

wrist-PPG encoder aligned to text at record, cross-recording and generative scopes

Brescia HuBERT-style masked discrete-token ECG FM on 9.1M multi-country ECGs; open HF weights under CC BY-NC.

weakly-supervised interpretability-enhancing ECG foundation model

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.

self-distilled foundation model for inertial sensing

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%.

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.

masked-autoencoder wrist IMU encoder (IMU-Bench reference model)

Hypertension 87%, atrial flutter 70%, ME/CFS 81%, sick sinus syndrome 87%. Beats MAE and PrimeNet baselines. Proprietary data.

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.

Average macro-F1 over a plain MAE baseline: +2.56 (broad multi-label), +4.85 (morphology), +12.80 (rhythm) across nine downstream tasks. No comparison with published ECG FMs in the abstract, so no board row.

lightweight distribution-robust Mamba for PPG biometrics

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.

Prospective sleep staging κ 0.845 ± 0.066 against expert consensus (inter-expert κ 0.850 ± 0.102). Three-class narcolepsy 84.2% accuracy (κ 0.761); cross-institution hypersomnia normal-vs-abnormal AUC 0.879; severe OSA AUC 0.874 without respiratory channels; depression symptoms AUC 0.801 on surrogate-labelled APPLES and 0.965 on an external clinically diagnosed cohort.

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.

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.

memory-efficient real-time rPPG via temporal-spatial state space duality

code release

Microsoft Research medical irregular time-series FM (CT-RoPE + frequency MoE + Neural ODE); NeurIPS 2025; MIT code; Large weights on HF.

multi-source contrastive ECG foundation model (Apple)

multi-spectral (RGB + NIR) state space model for in-vehicle rPPG

channel-aligned expert routing for selected-wavelength PPG

multi-scale spatio-temporal transformer for rPPG

code release

Held-out SeizeIT2 test (23 patients), frozen encoder + linear probe: AUPRC 0.3748 ± 0.0127, ROC-AUC 0.8025 ± 0.0071, balanced accuracy 0.7308 ± 0.0075. Random-init encoder AUPRC 0.3275; NeuroLM probe 0.2059 (medRxiv v2 Tables 3–4). Segment-level only.

code release

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.

Channel-agnostic MAE for heterogeneous wearable physio channels across consumer devices; public-data pretrain. Successor world-model: normwear-2.

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.

Sensor-Language-Action model (UCLA, Yuzhe Yang lab)

Time-Series Language Models (Stanford / ETH Zurich)

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.

physiology-aware lightweight contrastive network for unsupervised rPPG

Nokia Bell Labs open PPG FM (compact PaPaGei-S ~5.7M) with morphology-aware contrastive SSL.