How does a smartwatch learn heart rhythms without an ECG sensor?

Two models, one wrist

The Digital Health Team at Samsung Research America introduced xMAE and HiMAE in papers accepted to ICML and ICLR. Both are health foundation models — AI systems pretrained on large amounts of unlabeled biosignal data so they can later be adapted to many downstream tasks without fresh labels each time. The pair targets the same raw streams: heart rate, heart rate variability, sleep stages, activity. They solve different problems.

One model learns the temporal link between two signals from the same heartbeat. The other learns to read those signals at multiple time scales on the device itself. Together they move analysis from the cloud onto the watch.

xMAE: borrowing the ECG rhythm from light

An electrocardiogram on a watch measures the heart's electrical activity directly. It is accurate but demands that the user stop, touch the bezel, and wait. Photoplethysmography uses green LEDs to detect blood-volume changes under the skin. It runs passively all day. The two signals share a physiological source yet arrive at the sensor with a fixed delay, like thunder after lightning. This pretraining framework reconstructs masked portions of the ECG waveform using the simultaneously recorded PPG stream. The researchers fed the model approximately 9,400 hours of paired ECG and PPG recordings. After pretraining, the system outperformed unimodal baselines and existing multimodal methods on 15 of 19 evaluation tasks. Those tasks included cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification. The learned features also transferred across different sensor hardware, body locations, and recording environments.

HiMAE: the model that fits on the watch

Health data reveals different patterns at different resolutions. A few seconds expose individual heartbeats. Minutes expose sleep cycles. Hours expose activity trends. HiMAE uses separate encoders for short and long windows, then fuses the representations. During self-supervised training it reconstructs masked segments, learning to recognise the structures that matter for each task. A single pretrained HiMAE model handles classification, numerical regression, and data generation. It is smaller than prior architectures and produces results in less than one millisecond on a smartwatch-class CPU. That speed makes it the first demonstrated on-device health foundation model. Raw signals stay on the wrist. No cloud round-trip is required.

What changes for the person wearing it

The Connected Care vision shown at Galaxy Unpacked July 2026 describes a shift from reactive treatment to preventive, personalised, and connected experiences. xMAE and HiMAE are the technical substrate for that shift. A watch that continuously infers ECG-grade features from PPG can flag atrial fibrillation risk without the user initiating a measurement. A watch that classifies sleep stages and predicts cardiovascular markers locally can deliver insights in real time while the battery lasts for days. The announcement does not say when these models will ship in a Galaxy Watch firmware update. It does not disclose how many hours of labelled data were needed for fine-tuning after pretraining. It does not report clinical trial outcomes or regulatory clearance status. What it does show is that the core research problems — cross-modal alignment, multi-scale representation, on-device inference — have working solutions that run on the silicon already strapped to a wrist.