Samsung Research America has presented two foundation models designed to learn from the streams of physiological data collected by wearable devices. The research targets heart activity, sleep and movement, with the broader aim of producing health insights continuously on hardware with limited sensors and computing power.
The models are called xMAE and HiMAE. Both use self-supervised learning, which allows a system to discover useful patterns in large quantities of biosignal data without requiring every segment to be manually labelled. Samsung describes that approach as groundwork for future preventive and personalised health experiences, rather than announcing a finished diagnostic product.
That distinction is important. Smartwatches generate abundant measurements, but turning them into reliable health conclusions requires evidence across devices, populations and real-world conditions. Samsung has reported encouraging benchmark results and conference acceptances; it has not shown that the models can replace clinical assessment.
xMAE connects two views of cardiac activity
Samsung’s xMAE model focuses on the relationship between electrocardiogram and photoplethysmography signals. ECG measures the heart’s electrical activity and can support analysis of heart rate, variability and abnormal rhythms. On a wearable, however, an ECG reading normally requires the user to stop and initiate a measurement.
PPG instead uses an optical sensor to detect changes in blood flow. It can operate passively and continuously in a smartwatch. Because electrical activity and blood-flow changes arise from the same cardiac cycle but occur at different times, Samsung compares their relationship to seeing lightning before hearing thunder.
xMAE learns that timing by reconstructing masked parts of an ECG signal from PPG data. Its pretraining used roughly 9,400 hours of paired ECG and PPG recordings. The intended benefit is to extract cardiovascular features from continuously available PPG without requiring a separate manual ECG every time.
Samsung says xMAE outperformed unimodal models and existing multimodal methods in 15 of 19 evaluation tasks, including cardiovascular prediction, abnormal-result detection and sleep-stage classification. Those are company-reported research results. Performance in controlled datasets does not by itself establish accuracy for consumer guidance or clinical decisions.
HiMAE reads patterns at different speeds
Wearable signals carry meaning across several time scales. A short interval may reveal individual heartbeats, while longer windows can capture sleep cycles or patterns in physical activity. HiMAE uses multiple encoders to analyse short and long segments separately, then identifies which scale is useful for a particular task.
Like xMAE, it learns by reconstructing hidden portions of time-series data. Samsung says one pretrained HiMAE system can then support classification, numerical prediction and data generation. The company also reports strong performance from a smaller model than existing alternatives.
The most commercially relevant claim is speed. Samsung says HiMAE can produce results in less than one millisecond on a smartwatch-class CPU. If that result transfers from research tests to shipping devices, it could support analysis on the wrist rather than sending every signal to cloud servers.
On-device processing may reduce latency and dependence on connectivity. It may also limit some data transfers, although the source does not describe a product architecture or privacy policy for a future implementation. Efficiency is therefore demonstrated as a research objective, not yet a complete privacy guarantee.
Research performance still needs clinical context
Samsung presented the work as part of its Connected Care vision, which combines consumer technology with healthcare partnerships. The company says foundation models could support biosignal analysis, biomarker development and prediction of health issues from a common pretrained base.
That flexibility could make wearable development faster: instead of training an entirely separate model for each measurement, teams could adapt a shared representation of physiological patterns. It could also create a common point of failure if the pretraining data fails to represent particular users, sensor placements or health conditions.
Samsung says xMAE’s learned features showed potential across different sensors, body locations and collection environments. HiMAE’s compact design suggests that broader models need not remain confined to data centres. Neither statement confirms clinical validity, regulatory clearance or a release date.
The research therefore narrows a practical gap between abundant wearable data and useful interpretation. Its next test is not whether a model can reconstruct a hidden signal, but whether future products communicate uncertainty, remain accurate outside research datasets and help users without making a smartwatch appear more medically authoritative than the evidence supports.
