Samsung Develops AI Foundation Models for Smarter Health Tracking on Wearables

Samsung Research America is developing new AI foundation models that could make smartwatches more capable of understanding heart activity, sleep, physical movement, and other health-related biosignals.

The company’s Digital Health Team introduced two research models called xMAE and HiMAE, both designed to learn patterns from large amounts of wearable health data without requiring every piece of information to be manually labelled.

The research supports Samsung’s broader vision of preventive, personalised, and connected healthcare, where wearable devices could continuously analyse health signals and provide useful insights directly to users.

Samsung discussed this direction during its Health Forum at Galaxy Unpacked in July 2026.

AI That Learns Directly From Biosignals

Health foundation models work by learning underlying patterns from large datasets containing physiological signals.

Instead of training a separate AI model for every task, Samsung's approach allows one pretrained model to potentially support multiple applications, including health monitoring, biomarker development, sleep analysis, cardiovascular assessment, and predictive health features.

Sharanya Desai, Head of Digital Health Algorithms at Samsung Research America, said the research could help create health insights that are more efficient, accurate, and continuous.

Samsung is also focusing on developing models capable of running directly on wearable devices with limited processing power and sensors.

xMAE Connects PPG and ECG Heart Signals

The first model, xMAE, stands for Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning.

It focuses on understanding the relationship between two major cardiovascular signals: ECG and PPG.

An electrocardiogram or ECG measures the electrical activity of the heart and can provide information about heart rate, heart-rate variability, and abnormal heart rhythms.

However, smartwatch ECG readings usually require users to stop and actively perform a measurement.

Photoplethysmography or PPG, on the other hand, monitors changes in blood flow using optical sensors and can operate continuously in the background.

Samsung's xMAE model learns how these two signals relate to each other over time.

The company compares the relationship to seeing lightning before hearing thunder. The two events come from the same source but arrive at slightly different times.

Using this relationship, xMAE attempts to reconstruct portions of ECG information based on continuously collected PPG signals.

The goal is to allow wearable devices to extract more cardiovascular information without requiring users to repeatedly take manual ECG measurements.

Samsung trained xMAE using approximately 9,400 hours of ECG and PPG data.

According to Samsung, the model outperformed existing unimodal and multimodal biosignal models in 15 out of 19 evaluation tasks.

These evaluations included cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification.

Samsung also reported that the model's learned representations may work across different sensors, body locations, and data collection environments.

The research was accepted at the International Conference on Machine Learning, or ICML.

HiMAE Studies Health Data Across Different Time Periods

Samsung's second model, HiMAE, or Hierarchical Masked Autoencoder, takes a different approach.

Instead of focusing mainly on relationships between different biosignals, HiMAE looks at how wearable health information changes across different periods of time.

Some health information happens extremely quickly.

A heartbeat, for example, may need to be analysed within seconds.

Sleep patterns or physical activity trends, however, may require analysis across minutes or hours.

HiMAE uses multiple AI encoders to study short-term and long-term information separately.

This allows the model to determine which time scale is most important for a particular health-related task.

Samsung says the model can support several types of AI analysis using one pretrained system, including classification, numerical prediction, and data generation.

The company also reports that HiMAE can achieve strong performance while using a relatively small model.

Most notably, Samsung says the system can process information in less than one millisecond on a smartwatch-class CPU.

That opens the possibility of running advanced health AI directly on wearable devices rather than constantly sending health information to cloud servers.

HiMAE was accepted at the International Conference on Learning Representations, or ICLR.

Toward On-Device Health AI

Samsung's research points toward a future where wearable devices do more than simply record steps, heart rate, and sleep duration.

AI foundation models could allow smartwatches to understand relationships between multiple physiological signals, recognise longer-term health patterns, and potentially identify changes that deserve closer attention.

Running these models directly on a smartwatch could also reduce reliance on cloud processing while enabling faster and more continuous health analysis.

Subbu Venkatraman, Head of the Digital Health Research Lab at Samsung Research America, said the research demonstrates how health foundation models can capture both relationships between biosignals and the changing patterns that occur over time.

Samsung says it plans to continue advancing foundational health AI research and exploring how the technology can be translated into future healthcare and consumer wellness applications.

For Samsung, the smartwatch may no longer be just a device that collects health data.

With increasingly capable AI running directly on the wrist, it could become an intelligent system capable of continuously interpreting what those signals actually mean.