AI is rapidly transforming how wearable devices interpret the biosignals they collect — heart rate, sleep patterns, activity levels — turning raw data into actionable health insights. At the Galaxy Unpacked Health Forum in July 2026, Samsung outlined its Connected Care vision: a shift from reactive treatment toward preventive, personalized, and connected health experiences powered by trusted innovation and healthcare partnerships.
Central to this vision are health foundation models — AI systems trained on massive unlabeled biosignal datasets that can then be applied to diverse downstream tasks. Samsung Research America’s Digital Health Team has developed two such models: xMAE and HiMAE, both accepted at top-tier ML conferences (ICML and ICLR respectively).
What Are Health Foundation Models?
A health foundation model uses self-supervised learning to extract meaningful features from unlabeled biosignal data. After pretraining on large-scale health data, these models can be adapted for biosignal analysis, novel biomarker discovery, and health state prediction — all from a single pretrained backbone.
xMAE: Continuous Cardiac Insight from Wearables
An electrocardiogram (ECG) on wearables directly measures the heart’s electrical activity — highly accurate but typically requiring active, stationary measurements. Photoplethysmography (PPG) indirectly measures cardiac function through blood flow changes and can be passively, continuously captured.
These two signals stem from the same cardiac activity but appear with a time delay — like hearing thunder after seeing lightning. xMAE (Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning) learns the temporal relationship between them by reconstructing masked ECG segments using the continuously available PPG signal.
Pretrained on ~9,400 hours of ECG and PPG data, xMAE outperformed unimodal and existing multimodal baselines on 15 of 19 benchmark tasks — including cardiovascular disease prediction, abnormal test detection, and sleep-stage classification. It also generalized across different sensor types, body locations, and data collection environments.
HiMAE: Learning the Body at Every Time Scale — On Device
Health data from wearables reveals different patterns at different time scales. Short segments capture rapid signals like heartbeats; longer segments reveal accumulated patterns like sleep cycles or activity trends. HiMAE (Hierarchical Masked Autoencoder) uses multiple encoders to analyze short and long segments separately, identifying the right time scale for each health task — heart rate analysis vs. sleep prediction.
During training, HiMAE reconstructs masked portions of the data, learning key biosignal patterns even with limited labeled data. A single pretrained model handles classification, numerical prediction, and representation learning — all while being smaller than existing models.
Most critically: HiMAE delivers results in under one millisecond on a smartwatch-class CPU. This demonstrates, for the first time, the feasibility of on-device health foundation models — analyzing raw biosignals in real time without cloud dependency.
Toward Connected Care
Together, xMAE and HiMAE advance AI models that accurately capture the physiological relationships and temporal structures unique to biosignals. Both aim to deliver real-time, reliable, personalized health insights that generalize across many tasks from a single pretrained model — the foundation for Samsung’s vision of preventive, connected digital health.
