Apple AHMS PPG Foundation Model

The Apple Watch's optical heart sensor records photoplethysmography (PPG) continuously and passively, producing an enormous volume of signal — and almost no labels. The medical annotations needed to train supervised models on this data are scarce, expensive, and biased toward people who already have a diagnosis.

This model, from Apple, addresses that gap by learning general-purpose representations of PPG 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 ECG model was trained the same way on the study's electrocardiogram recordings; the two share a recipe but not a signal, a cohort split, or a set of 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 wearable signal already encode meaningful information about participant demographics and health conditions, readable with simple probes on the frozen features.

#Key Features

• Participant-level contrastive objective: Positive pairs are drawn from the same participant rather than only from augmentations of a single segment, encouraging the model to capture stable, person-specific physiology instead of segment-level noise.
• Tailored augmentation and regularized loss: A stochastic augmentation module suited to periodic physiological waveforms and a regularized contrastive loss, optimized with momentum training, stabilize self-supervised learning on noisy real-world data.
• Frozen-feature transfer: Demographics such as age, body mass index, and sex are predicted from frozen embeddings, demonstrating transfer without fine-tuning the encoder.
• Continuous-signal scale: Because PPG is recorded passively rather than on demand, training draws on far more signal per participant than an equivalent ECG corpus, across roughly three years of longitudinal recordings.

#Technical Details

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 PPG stream of the ~141K-participant AHMS cohort, independently of the ECG model. 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.

#Applications

The intended use is health and wellness inference from wearable optical heart sensor data: estimating demographic and physiological attributes, screening for or stratifying health 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, remote monitoring, and population-scale cardiovascular research. PPG in particular supports continuous monitoring in a way on-demand ECG cannot, so it is the modality of the two better matched to always-on applications.

#Impact

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 PPG 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.