Samsung is developing new AI foundation models designed to learn from the biosignals collected by wearable devices. The research focuses on data from smartwatches, including heart activity, sleep, and physical movement, with the goal of creating more continuous and personalised health insights.
The company presented its broader Connected Care vision during Galaxy Unpacked, describing a future in which consumer health technology can support preventive care through connected devices and healthcare partnerships.
Samsung Research America’s Digital Health Team has introduced two models as part of this work: xMAE and HiMAE. Both use self-supervised learning to identify patterns in health data without relying entirely on labelled datasets.
Building foundation models for health data
Health foundation models are designed to learn general representations from large amounts of physiological data before being adapted to individual tasks.
Samsung says this approach could allow a single pretrained model to support applications including biosignal analysis, biomarker development, health prediction, and other forms of physiological monitoring.
The two models take different approaches to understanding wearable data. xMAE focuses on relationships between different biosignals, while HiMAE examines physiological patterns across multiple time scales.
Samsung says both models have been accepted at major machine learning conferences, with xMAE accepted to the International Conference on Machine Learning and HiMAE accepted to the International Conference on Learning Representations.
xMAE connects PPG and ECG signals
One of the central challenges in wearable health monitoring is balancing continuous measurement with the limitations of sensors.
Electrocardiography, or ECG, directly measures the heart’s electrical activity and can provide information about heart rate, heart-rate variability, and abnormal rhythms. However, ECG measurements on consumer wearables typically require users to actively perform a reading.
Photoplethysmography, or PPG, works differently. It uses optical sensors to detect changes in blood volume and can operate continuously in the background on devices such as smartwatches.
Because both signals reflect cardiovascular activity, Samsung’s researchers sought to teach an AI model how information from one signal relates to the other.
xMAE, short for Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning, uses masked portions of ECG and PPG data during training. The model learns temporal relationships between the signals by attempting to reconstruct missing ECG information using PPG data.
Samsung trained the model using approximately 9,400 hours of ECG and PPG recordings.
Potential for continuous cardiovascular monitoring
The approach could eventually make it possible to derive additional cardiovascular insights from continuously collected PPG measurements, reducing the need for users to manually perform separate ECG recordings.
Samsung reports that xMAE outperformed unimodal biosignal models and existing multimodal approaches across 15 of 19 evaluation tasks.
Those evaluations included cardiovascular disease prediction, detection of abnormal test results, and sleep-stage classification.
The company also reports that the representations learned by xMAE could remain useful across different sensor devices, body locations, and data-collection environments. This could be important for consumer health applications, where the available sensors and quality of measurements can vary considerably between devices.
HiMAE looks at health patterns over different time periods
The second model, HiMAE, takes a different approach by focusing on the way physiological information changes across time.
Wearable data can contain useful information at very different intervals. A short window may capture individual heartbeats or rapid changes in heart rate, while longer periods can reveal patterns associated with sleep, exercise, or daily activity.
HiMAE, or Hierarchical Masked Autoencoder, uses multiple encoders to examine short- and long-term segments of wearable time-series data.
The model learns by reconstructing portions of data that have been masked during training. This allows it to identify physiological patterns without requiring every training example to have a human-provided label.
Samsung says the resulting model can be adapted to tasks including classification, numerical prediction, and data generation.
Bringing AI processing onto smartwatches
One of the more notable aspects of Samsung’s research is its focus on running health AI directly on wearable hardware.
Large AI models can require significant computing resources, making cloud processing an attractive option for many applications. However, relying on cloud infrastructure can introduce latency, connectivity requirements, and additional considerations around sensitive health data.
Samsung says HiMAE achieved strong performance while using a smaller model than existing approaches. The company also reports that it can produce results in less than one millisecond on a smartwatch-class central processing unit.
If such performance translates into practical consumer applications, wearable devices could perform increasingly sophisticated health analysis locally rather than continuously sending physiological data to remote servers.
Toward more personalised wearable health insights
Samsung’s research illustrates a broader shift in health AI toward models that can learn directly from continuous physiological signals.
Rather than designing a separate model for every individual health task, foundation models can first learn general patterns from large datasets and then be adapted for specific applications.
For wearable technology, that could eventually support more continuous analysis of cardiovascular activity, sleep, exercise, and other health-related signals.
The research remains part of Samsung’s broader effort to develop the underlying technology for connected health experiences. Turning these models into reliable consumer health features will require further validation, particularly where AI-generated insights could influence decisions about a person’s health.
Still, the work demonstrates how advances in foundation models are extending beyond conventional text and image applications. By learning the complex temporal relationships within physiological data, AI systems could become increasingly capable of interpreting the signals collected by everyday wearable devices.
