Samsung Unveils Health AI Models for Wearable Biosignals

Smartwatches can capture continuous heart-rate, sleep and activity signals, but turning noisy streams into useful health insights typically requires labelled data and computing power that wrist-worn devices lack. Samsung Research America, the U.S. research arm of Samsung Electronics, is applying self-supervised learning to build reusable representations from unlabeled biosignals. Compact foundation models could move wearables beyond basic tracking toward continuous, personalized and preventive health monitoring without sending every calculation to the cloud.

Samsung Research America presented two approaches in 2026. HiMAE, published at ICLR on Feb. 5, learns health patterns across multiple time resolutions and runs in under one millisecond on smartwatch-class CPUs. xMAE, listed by Samsung Research on May 1 for ICML, was pretrained on about 3.4 million paired 10-second ECG and PPG recordings, or roughly 9,400 hours from 2,400 subjects. Samsung said xMAE beat unimodal and multimodal baselines on 15 of 19 downstream tasks, spanning cardiovascular prediction, laboratory results and sleep staging.