Loading…
Photoplethysmography foundation model pretrained by reconstructing masked ECG from PPG, learning cardiac timing structure for wearable health tasks.
No providers recorded yet. Browse all providers
Every heartbeat produces two signals a wearable can see. The electrocardiogram (ECG) records the electrical activation that starts the beat; the photoplethysmogram (PPG) records the pressure wave that reaches the wrist milliseconds later, shaped by the vasculature it travelled through. That delay is not noise — pulse arrival time tracks blood pressure and vascular dynamics. Most multimodal self-supervised methods discard it, treating ECG and PPG as interchangeable views to align contrastively or reconstruct jointly — an objective a model can satisfy by interpolating within each locally smooth waveform, never reasoning across them.
xMAE makes the delay the training signal. PPG is left fully visible while continuous blocks covering 80–90% of the paired ECG are masked, and the model must reconstruct the missing ECG from PPG alone through a directional cross-attention layer in which ECG tokens are queries and PPG tokens supply keys and values. Minimizing that error requires beat-level temporal alignment, so timing structure is driven into the PPG encoder rather than left to chance. After pretraining the ECG branch is discarded: inference needs only PPG, the signal a smartwatch already collects passively and continuously.
xMAE was developed by the Algorithm Team in the Digital Health Lab at Samsung Research America, with collaborators at The Pennsylvania State University and Samsung Electronics, and presented at ICML 2026. It enters a crowded field of PPG foundation models — PaPaGei, Pulse-PPG, the Apple AHMS PPG model — with a different thesis: physiological structure in the pretraining objective can substitute for scale.
Both branches are transformers. A modality-specific convolutional stem processes each visible waveform; its output is split into non-overlapping 40-sample patches and projected to 256-dimensional tokens with learnable positional embeddings. The PPG encoder is two layers deep, the ECG encoder one, and a single-layer bridge carries the cross-attention into a lightweight ECG decoder trained with mean squared error over masked positions only — 6.5M parameters in total. Pretraining used the waveform-matched subset of MIMIC-III: roughly 3.4 million synchronized 10-second ECG–PPG segments sampled at 100 Hz, about 9,400 hours, from around 2,400 intensive-care subjects.
Evaluation freezes the PPG encoder and linear-probes it over 6 studies, 19 tasks, 2,300 hours of PPG and 12,500 subjects, using 5-fold subject-disjoint cross-validation on cohorts spanning different devices, body locations, and laboratory versus free-living settings. xMAE is best on 15 of 19 tasks. Its widest margins over the strongest baseline: hypertension (68.8 vs 56.8 AUROC), ectopic-beat detection (87.8 vs 86.2) and A1C abnormality (65.1 vs 52.5). A direct probe supports the mechanism: the ECG–PPG delay measured from reconstructed signals has a median error of 21.5 ms, a 53.3% reduction against a multimodal masked autoencoder that reconstructs both modalities symmetrically.
The target setting is passive cardiovascular monitoring from consumer wearables, where PPG is abundant and ECG requires an active reading. The frozen embeddings support hypertension assessment in laboratory and free-living conditions, premature ventricular contraction detection, abnormal blood-lab screening (A1C, hemoglobin, platelets, sodium), sleep-stage classification, blood pressure regression and demographic inference. Label efficiency matters as much as accuracy here: 128 labeled PPG segments per patient already beat multimodal baselines on PVC detection, and roughly 2,048 match fully supervised training. The authors do not claim direct clinical decision-making capability.
xMAE's central result is that a physiologically motivated inductive bias can stand in for data volume — it matches or beats open-weight PPG models pretrained on an order of magnitude more signal, at a fraction of the parameters. The framework generalizes to any pair of temporally ordered biosignals observing different stages of one process — ECG and ballistocardiography, EEG and fNIRS, muscle activation and motion. Its limitations are concrete: pretraining requires paired ECG–PPG recordings, not available at wearable scale; the mean-squared-error objective captures dominant R-peak timing but not finer structure such as P–R intervals; and while the pretraining and signal-processing code is public, the pretrained weights and the Samsung evaluation cohorts are withheld under data-licensing and industrial policy, so the 19-task comparison cannot be reproduced independently.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.