From Biosignals to Health Insights: Samsung Research’s Work on Health Foundation Models

At Galaxy Unpacked in July 2026, Samsung outlined its Connected Care vision, emphasizing preventive and personalized health monitoring through AI-driven biosignal analysis from wearables like smartwatches. Researchers at Samsung Research America (SRA) developed xMAE and HiMAE, foundation models designed to interpret health data such as heart rate, sleep patterns, and physical activity with greater precision by learning temporal and physiological relationships in biosignals.

xMAE focuses on the relationship between ECG and PPG signals, which measure heart activity differently but originate from the same source. By reconstructing masked ECG data using continuous PPG measurements, xMAE enables accurate cardiovascular health analysis without requiring manual ECG readings. Trained on 9,400 hours of data, it outperformed existing models in 15 of 19 health-related tasks, including disease prediction and sleep-stage classification.

HiMAE addresses the challenge of analyzing health data across varying time scales, from rapid heartbeats to long-term sleep patterns. Using multiple encoders, it processes short and long segments separately to identify relevant health insights efficiently. The model achieves high performance with low computational demands, producing results in under one millisecond on a smartwatch CPU, demonstrating the feasibility of on-device health AI without cloud dependency.

Both models were accepted to ICML and ICLR, highlighting their significance in advancing health AI. Samsung’s work aims to deliver continuous, personalized health insights by leveraging self-supervised learning on unlabeled biosignal data, reducing reliance on labeled datasets and enabling broader applications across wearable devices and healthcare tasks.