Samsung Develops AI Health Models to Analyse Wearable Biosignal Data

Samsung is advancing the use of artificial intelligence in wearable health technology with two new health foundation models designed to understand biosignals collected from devices such as smartwatches. The research focuses on physiological data linked to heart activity, sleep and physical movement, with the broader goal of enabling more continuous, personalised and preventive health insights.

Samsung Research America’s Digital Health Team presented the research as part of the company’s broader Connected Care vision during the Health Forum at Galaxy Unpacked in July 2026. The company says its future healthcare strategy will combine AI, consumer devices and healthcare partnerships to deliver more connected health experiences.

The two AI models, known as xMAE and HiMAE, are designed to learn from large volumes of wearable biosignal data. Samsung believes such foundation models could eventually support applications including health prediction, biomarker discovery and advanced biosignal analysis.

Samsung’s Health AI Foundation Models

Health foundation models use artificial intelligence techniques, particularly self-supervised learning, to discover useful patterns in large datasets without requiring every piece of data to be manually labelled.

This approach is important for wearable health technology because smartwatches continuously generate large amounts of physiological information. Instead of training separate AI systems for every individual task, a foundation model can learn general patterns from biosignals and then be adapted for different health applications.

Samsung’s research focuses on two different aspects of wearable data.

xMAE, or Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning, is designed to understand relationships between different physiological signals and their timing.

HiMAE, or Hierarchical Masked Autoencoder, focuses on identifying health patterns across different time scales in wearable time-series data.

According to Samsung, xMAE has been accepted at the International Conference on Machine Learning, while HiMAE has been accepted at the International Conference on Learning Representations.

Sharanya Desai, Head of Digital Health Algorithms at Samsung Research America, said the research provides technical foundations for producing health insights that are efficient, precise and continuous. Samsung also intends to develop models that can work with different biosignals while operating directly on devices with limited sensors and computing resources.

xMAE Connects PPG and ECG Signals

One of the key areas explored by xMAE is the relationship between photoplethysmography (PPG) and electrocardiography (ECG).

ECG measures the electrical activity of the heart and can provide information such as heart rate and heart-rate variability. ECG measurements can also be useful when identifying irregular heart rhythms and cardiovascular risks.

PPG works differently. It detects changes in blood flow using optical sensors and can be continuously collected by wearable devices such as smartwatches.

Although ECG and PPG measure physiological activity through different mechanisms, they are connected by the underlying activity of the cardiovascular system. Samsung’s xMAE model is designed to learn the timing relationship between the two signals.

The model uses masked signal reconstruction, meaning portions of one signal can be hidden while the AI attempts to reconstruct them using information from another signal. In its research, xMAE learned to reconstruct missing ECG information using PPG data.

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

The potential advantage is significant for wearable health monitoring. PPG can generally be collected passively and continuously through a smartwatch, whereas ECG measurements commonly require a user to initiate a specific measurement. An AI model capable of understanding the relationship between the two could potentially help extract additional cardiovascular information from continuously collected wearable data.

However, these research findings should not be interpreted as a replacement for clinical ECG testing or professional medical diagnosis.

xMAE Shows Strong Results Across Multiple Tasks

Samsung reports that xMAE performed better than unimodal biosignal models and existing multimodal learning approaches in 15 of 19 evaluation tasks.

The evaluated tasks included cardiovascular disease prediction, detection of abnormal test results and sleep-stage classification.

Another important finding was the model’s ability to transfer learned representations across different sensor devices, body locations and data-collection environments. This could be valuable because wearable health data is not always collected in identical conditions.

Different smartwatch designs, sensor placements and user behaviours can affect physiological measurements. A model capable of learning broader biosignal relationships could potentially be more adaptable across devices and real-world scenarios.

Subbu Venkatraman, Head of the Digital Health Research Lab at Samsung Research America, highlighted the dynamic nature of biosignals and said the research demonstrates the potential of health foundation models to capture relationships between different signals as well as their underlying temporal patterns.

HiMAE Studies Wearable Health Data Over Different Time Periods

While xMAE focuses on relationships between different biosignals, HiMAE takes a different approach by studying wearable data at multiple time scales.

Physiological information can change rapidly or develop gradually. For example, individual heartbeats represent short-term activity, while sleep patterns and physical activity may become meaningful only when observed over longer periods.

HiMAE uses multiple encoders to analyse short and long segments of wearable time-series data. This hierarchical structure allows the model to identify patterns that occur at different time intervals.

The AI system learns by reconstructing masked sections of wearable data. This self-supervised approach means the model can learn from large quantities of health information without requiring extensive manual labelling.

After pretraining, Samsung says HiMAE can be applied to several types of tasks, including classification, numerical prediction and data generation.

This flexibility could make a single pretrained model useful for different wearable health applications rather than requiring a separate AI system for every individual task.

AI Designed to Run Directly on Smartwatches

One of the most notable aspects of Samsung’s research is its focus on efficient AI processing.

Samsung reports that HiMAE achieved strong performance while using a smaller model than existing approaches. The company also says the model can generate results in less than one millisecond on a smartwatch-class central processing unit.

If such performance can be translated into practical consumer products, it could support more health analysis directly on wearable devices.

On-device AI can offer several potential advantages. Processing data locally can reduce dependence on continuous cloud connectivity, potentially improve response times and limit the amount of raw physiological information that needs to be transmitted to remote servers.

For wearable devices, efficiency is particularly important because smartwatches have limited processing power, battery capacity and sensor resources compared with desktop computers or cloud-based systems.

The Future of AI-Powered Wearable Healthcare

Samsung’s research reflects a wider shift toward AI-powered wearable health monitoring. Smartwatches are increasingly capable of collecting continuous information about heart activity, sleep, exercise and other physiological signals.

Foundation models could help turn this raw data into more useful patterns by learning relationships that may be difficult to identify using conventional algorithms.

The long-term objective is not simply to collect more health data, but to interpret it more efficiently and provide users with meaningful information. Potential applications could include continuous health monitoring, early risk identification, personalised wellness recommendations and improved understanding of physiological changes.

At the same time, healthcare AI must meet high standards for accuracy, privacy, reliability and clinical validation. Research models demonstrating strong performance do not automatically mean that they are ready for medical diagnosis or treatment.

Samsung’s xMAE and HiMAE research nevertheless represents an important development in AI for wearable healthcare. By combining self-supervised learning, multimodal biosignal analysis and efficient on-device processing, the company is exploring how future smartwatches could become more intelligent health-monitoring platforms.

As Samsung continues developing its Connected Care strategy, health foundation models may become an important technology for delivering more continuous, personalised and device-based health insights in the years ahead.