[Edaily Reporter JAEMIN SONG ] SamsungElectronics is advancing its artificial intelligence (AI) technology to provide personalized health information by comprehensively analyzing vital signs—such as heart rate, sleep, and activity levels—measured by its smartwatches.

According to SamsungElectronics on the 14th, researchers at Samsung Research America (SRA) Digital Health have developed the “xMAE” and “HiMAE” health foundation models, which learn from wearable biosignals. The two studies were accepted by the International Conference on Machine Learning (ICML) and the International Conference on Learning Representations (ICLR), respectively—both world-renowned AI conferences.

Health foundation models are AI models that are pre-trained on large-scale medical and health data and then applied to various tasks, such as disease prediction, biosignal analysis, and sleep stage classification. Unlike the conventional approach of developing separate models for each specific use case, a single model can be fine-tuned for multiple health management functions.

xMAE learns the causal relationships and time differences between different biosignals. While electrocardiograms (ECGs) directly measure the heart’s electrical activity, allowing for precise analysis of arrhythmias and atrial fibrillation, they often require manual measurement by the user, which limits their use for continuous monitoring. In contrast, photoplethysmography (PPG) uses smartwatch sensors to routinely measure changes in blood flow.

xMAE learns the relationship between ECG and PPG signals, which are generated at a fixed time interval from the same cardiac activity. Based on this, it is designed to reconstruct missing ECG segments using PPG, which is easier to measure. The researchers explain that this allows for more precise analysis of cardiovascular health features based on PPG data obtained from wearables, without the need for separate ECG measurements.

The research team pre-trained xMAE using approximately 9,400 hours of ECG and PPG data. As a result, it demonstrated higher performance than single biosignals or existing training methods in 15 out of 19 evaluation tasks, including cardiovascular disease prediction, detection of abnormal test results, and sleep stage classification. The team also confirmed the potential to utilize the learned features even when the sensor device, measurement location, or environment changes.

HiMAE is a model that analyzes wearable data by dividing it into multiple time intervals. It learns changes that occur over short periods—such as heartbeats—and trends that accumulate over long periods—such as sleep and activity levels—in separate layers. By masking part of the data and then reconstructing it, the model was designed to identify key patterns in biosignals even when separate ground-truth information is lacking.

The research results showed that HiMAE is smaller in size than existing models while delivering high performance on key evaluation metrics. Computational efficiency was also improved to the point where results can be generated within 1 millisecond on a smartwatch’s central processing unit (CPU). The researchers explained that this demonstrates the potential of on-device health foundation models that analyze biosignals in real time on the device itself, without relying on cloud servers.

Sharanya Desai, a researcher on the SRA Digital Health Team who led the study, said, “We have laid the technological foundation for providing more efficient, precise, and continuous health insights,” adding, “We hope to contribute to the development of health AI models that operate even with limited sensors and computational resources.”

Subhu Venkatraman, a researcher on the SRA Digital Health team, said, “We have demonstrated the potential of health foundation models that learn the relationships and temporal structures among biosignals,” adding, “We will continue our health AI research to support a wide range of healthcare applications.”