How does a smartwatch learn to read your heart without an ECG sensor?

Two models, one wrist

Researchers at Samsung Research America introduced xMAE and HiMAE in July 2026. Both 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 the models can be applied to a wide range of downstream tasks including biosignal analysis and biomarker development. They can also support health issue prediction. The work was presented at the Health Forum during Galaxy Unpacked July 2026. xMAE was accepted to the International Conference on Machine Learning. HiMAE was accepted to the International Conference on Learning Representations.

Borrowing ECG accuracy from PPG convenience

An electrocardiogram on a wearable directly measures the electrical activity of the heart. It is useful for measuring heart rate and heart rate variability, identifying abnormal heart rhythms and risks for conditions such as atrial fibrillation. ECG is highly accurate but typically requires users to pause and take an active measurement. Photoplethysmography indirectly measures cardiac function by detecting changes in blood flow. It can be passively and continuously measured through sensors in wearable devices such as smartwatches. The two signals originate from the same cardiac activity but they appear with a certain time difference much like thunder is heard after lightning is seen.

xMAE is a biosignal pretraining framework designed to learn the temporal relationship between the two signals by reconstructing masked portions of the ECG signal using the more easily and continuously measured PPG signal. As a result cardiovascular health-related features can be analyzed more precisely using PPG without requiring separate manual ECG measurements. The researchers pretrained xMAE using approximately 9,400 hours of ECG and PPG data. The model outperformed unimodal biosignal models and existing multimodal learning methods in 15 of 19 evaluation tasks including cardiovascular disease prediction and abnormal test-result detection. It also improved sleep-stage classification. The study also confirmed the potential to use the learned features across different sensor devices and body locations. The features also transferred across data-gathering environments.

Fitting a health model on a watch chip

Health data measured from wearable devices can reveal different information depending on the time scale at which it is analyzed. Short time segments can reveal rapidly changing signals such as heartbeats. Longer time segments can uncover patterns that accumulate over time such as sleep or physical activity. This is similar to examining different information by zooming in and out of an image.

HiMAE is a self-supervised learning model that learns representation from wearable data across multiple time scales. It uses multiple encoders to analyze short and long segments separately allowing the model to identify information required for each health task such as heart rate analysis or sleep prediction at the appropriate time scale. During training the model reconstructs masked parts of the data. This enables it to learn key patterns in biosignals even when labeled data is limited. A single pretrained HiMAE model handles classification and numerical prediction. It can also generate data. The model achieved high performance while being smaller than existing models. It also improved computational efficiency to the point where it can produce results in less than one millisecond on a smartwatch-class central processing unit. Most importantly HiMAE demonstrates the potential of on-device health foundation models for the first time analyzing raw health signals in real time without relying on cloud servers.

What changes for the person wearing the watch

Both models aim to deliver precise continuous personalized health insights and to generalize across many health tasks from a single pretrained model. The Connected Care vision shared at Galaxy Unpacked July 2026 describes a future where care moves from reactive treatment toward preventive personalized and connected experiences supported by trusted health innovation and partnerships across the healthcare ecosystem. Galaxy Watch9 and Samsung Health are the consumer surfaces where this research is expected to appear. The announcement does not say when specific features will roll out or which watch generations will support on-device inference. It also does not disclose how many hours of labeled data were used for fine-tuning after pretraining or whether the 9,400-hour dataset represents a single study cohort or aggregated sources.