Samsung's New Health AI Reads Wearables in Under a Millisecond

Samsung's New Health AI Reads Wearables in Under a Millisecond — No Cloud Needed

Samsung Research America introduced two health foundation models — xMAE and HiMAE — that learn from wearable biosignals. Accepted at ICML and ICLR, they push cardiovascular analysis from continuously measured smartwatch data and, in HiMAE's case, run entirely on-device in under a millisecond on a smartwatch-class CPU.

Key Takeaways

Samsung Research America introduced two health foundation models — xMAE and HiMAE — that learn from wearable biosignals.
Accepted at ICML and ICLR, they push cardiovascular analysis from continuously measured smartwatch data and, in HiMAE's case, run entirely on-device in under a millisecond on a smartwatch-class CPU.

AI synthesis4 sources7

Samsung Research America's Digital Health Team has introduced two 'health foundation models' designed to interpret the flood of biosignal data streaming from smartwatches and other wearables. The models — xMAE and HiMAE — were accepted at ICML and ICLR respectively, and the company says together they point toward a future of continuous, personalized, on-device health AI.

A health foundation model learns meaningful features from unlabeled biosignal data using self-supervised learning, then applies them to many downstream tasks: analyzing ECG and PPG waveforms, building biomarkers, and predicting health issues. The first model, xMAE (Physiology-Aware Masked Cross-Modal Reconstruction), learns the temporal relationship between two cardiac signals: ECG, which is accurate but requires an active measurement, and PPG, which can be captured passively and continuously by a smartwatch. By reconstructing masked ECG segments from PPG, xMAE extracts cardiovascular features from continuous PPG alone. Pretrained on about 9,400 hours of ECG and PPG data, it outperformed unimodal models and existing multimodal methods in 15 of 19 evaluation tasks, including cardiovascular disease prediction and sleep-stage classification.

The second model, HiMAE (Hierarchical Masked Autoencoder), learns health patterns across different time scales — short segments capture heartbeats, long segments capture sleep or activity. A single pretrained model handles classification, numerical prediction and data generation, and does it efficiently enough to run in less than one millisecond on a smartwatch-class CPU. That makes HiMAE, in Samsung's telling, the first demonstration of an on-device health foundation model: raw health signals analyzed in real time without any cloud connection.

The models are part of the 'Connected Care' vision Samsung laid out at its Galaxy Unpacked event, in which care moves from reactive treatment toward preventive, personalized health experiences. For now they are research — but the sub-millisecond latency and on-device architecture suggest the next Galaxy Watch generation may not need the cloud to understand your heart.