Loading…
Self-supervised foundation model for wearable electrocardiograms, trained with participant-level contrastive learning on 141,000 participants.
No providers recorded yet. Browse all providers
Electrocardiograms recorded on an Apple Watch are user-initiated: a participant places a finger on the crown and holds still for a reading. That makes the corpus sparser and more episodic than a continuously sampled signal, but also cleaner and more clinically legible — ECG is the modality cardiologists actually read. What it lacks, at consumer scale, is labels.
This model, from Apple, learns general-purpose representations of wearable ECG through self-supervised contrastive learning, so downstream health tasks can be solved from frozen embeddings rather than from large labeled datasets. Introduced by Salar Abbaspourazad, Oussama Elachqar, Andrew C. Miller, Saba Emrani, Udhyakumar Nallasamy, and Ian Shapiro, and published at ICLR 2024, it was trained on data from roughly 141,000 participants of the Apple Heart and Movement Study (AHMS), collected over approximately three years. A separate PPG model was trained the same way on the study's optical heart sensor stream; the two share a recipe and a cohort but are trained independently and produce separate weights.
The authors describe this as part of the first study to build foundation models from large-scale PPG and ECG data captured by consumer wearables, as opposed to clinical-grade equipment in controlled settings. The central finding is that representations learned purely from unlabeled recordings already encode meaningful information about participant demographics and health conditions, readable with simple probes on the frozen features.
The encoder is convolutional, trained with a SimCLR-style contrastive framework adapted for biosignals. The design choices that matter are participant-level positive pair selection — two segments from the same participant are treated as a positive pair — a stochastic augmentation pipeline suited to periodic physiological waveforms, and a regularized contrastive loss optimized with a momentum-based scheme for stable training at scale.
Training ran on the ECG stream of the ~141K-participant AHMS cohort, independently of the PPG model rather than as a joint or multimodal fit. Evaluation uses linear or lightweight probes on the frozen embeddings: the learned representations recover participant demographics (age, BMI, sex) and signal-derived attributes, and carry predictive information about health conditions. Exact parameter counts and per-task metrics are reported in the paper rather than summarized here.
The intended use is health inference from wearable electrocardiograms: estimating demographic and physiological attributes, screening for or stratifying cardiac conditions, and serving as a feature backbone for downstream clinical and research tasks where labeled wearable data is limited. Because transfer works from frozen embeddings, researchers can build task-specific classifiers or regressors with modest labeled datasets, which suits digital health studies and population-scale cardiovascular research. ECG's clinical interpretability makes it the better of the two modalities where a downstream finding has to be explained to a clinician, even though it is sampled far less often than PPG.
This work helped establish that the foundation-model recipe — large-scale self-supervised pretraining followed by lightweight transfer — extends to consumer wearable biosignals, and it has become a widely cited reference point for subsequent ECG representation-learning efforts. Its main limitation for the open research community is access: training used proprietary Apple Heart and Movement Study data, and neither the trained weights nor the training code have been released, so the results cannot be directly reproduced or the encoder reused outside Apple. The contribution is therefore primarily conceptual and methodological — a demonstration of feasibility and a blueprint — rather than a shared artifact that others can build on directly.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.