Samsung is pushing deeper into health AI, and this time the focus is not a chatbot, an app assistant, or another generative AI feature.
It is the data already coming off your wrist.
Samsung Research America’s Digital Health Team has developed two AI foundation models built to learn from wearable biosignals, including heart activity, sleep patterns, and physical activity. The models, called xMAE and HiMAE, are designed to extract useful patterns from streams of physiological data without requiring enormous amounts of manually labelled health information.
That matters because wearables collect a lot of data. Making sense of it continuously, efficiently, and on relatively small hardware is the harder part.
Samsung Wants AI to Understand the Signals Behind Wearable Health Data
Samsung’s research fits into a broader Connected Care strategy the company discussed during its Health Forum at Galaxy Unpacked in July 2026.
The idea is fairly clear: health devices should move beyond simply showing measurements and start identifying patterns that could become useful health insights.
Foundation models are one possible route.
Instead of training a completely different model for every health function, Samsung is exploring models that first learn general characteristics from large amounts of biosignal data. Those learned representations can then be adapted for tasks such as classification, biomarker development, physiological prediction, or detecting unusual health patterns.
It is a similar philosophy to foundation models elsewhere in AI, but physiological data behaves very differently from text or images. Heartbeats, blood-flow signals, sleep cycles, and activity patterns all unfold across time.
That makes timing part of the data itself.
xMAE Tries to Connect PPG and ECG Data
The first model, xMAE, stands for Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning.
The important part is what it connects.
Most modern smartwatches can collect photoplethysmography, or PPG, continuously. PPG uses optical sensors to detect changes in blood flow. ECG readings work differently. They measure the heart’s electrical activity and can provide useful information about heart rhythm and heart-rate variability.
The catch is that wearable ECG measurements usually require the user to stop and actively take a reading.
PPG can keep running in the background.
Samsung trained xMAE to learn relationships between those two types of cardiac signals. During training, portions of ECG data are hidden and the model learns to reconstruct them using corresponding PPG information.
Samsung says the model was pretrained using roughly 9,400 hours of ECG and PPG data.
The potential is interesting: continuously collected PPG could become more useful for cardiovascular analysis without requiring users to perform an ECG measurement every time.
That does not turn a smartwatch into a hospital ECG machine. It does give AI more context from signals the watch is already collecting.
Samsung Says xMAE Beat Other Models in Most Evaluation Tasks
Samsung reports that xMAE outperformed unimodal biosignal models and existing multimodal approaches in 15 of 19 evaluation tasks.
Those evaluations included areas such as cardiovascular disease prediction, detecting abnormal test results, and identifying sleep stages. The company also says features learned by the model showed an ability to transfer across different devices, sensor placements, and data-collection environments.
That last part could become particularly important.
Wearable health data is messy. Different devices use different sensors. People wear those devices differently. Signal quality changes with movement, body position, hardware, and environment.
A health AI system that only performs well under one carefully controlled setup has limited usefulness.
HiMAE Looks at Health Data Across Different Timescales
Samsung’s second model, HiMAE, attacks another problem.
Not every meaningful health signal appears over the same amount of time.
A heartbeat changes in seconds. Sleep patterns develop across hours. Physical activity trends can stretch across much longer windows.
HiMAE, short for Hierarchical Masked Autoencoder, uses multiple encoders to study wearable time-series data across shorter and longer intervals.
The model can then emphasize the timescale that matters for the task it is performing.
Heart-rate analysis may need very short signal windows. Sleep prediction needs something entirely different.
Samsung says HiMAE can support classification, numerical prediction, and data generation using one pretrained model. Like xMAE, it learns by reconstructing masked portions of biosignal data rather than depending entirely on labelled examples.
The More Interesting Detail: It Could Run Directly on a Watch
Large AI models usually create an obvious hardware problem.
They need computing power.
Samsung says HiMAE achieved strong performance while using a smaller architecture than competing models and can produce results in under one millisecond on a smartwatch-class CPU.
That changes the conversation.
If meaningful health analysis can happen on the wearable itself, some AI features would not need to continuously send raw physiological data to cloud servers.
There are practical advantages: faster responses, less dependence on connectivity, lower server costs, and potentially better privacy depending on how Samsung ultimately deploys the technology.
That same shift toward local and on-device AI is becoming increasingly important across PCs, edge systems and other consumer hardware.
On-device AI is already becoming a major battleground in smartphones. Health wearables may be an even more natural place for it.
Health AI Is Quietly Becoming One of the Biggest Wearable Use Cases
Smartwatch companies spent years competing over displays, battery life, fitness tracking, and sensor counts.
AI creates a different kind of race.
Once several devices can measure similar biological signals, the competitive advantage increasingly comes from what software can infer from those measurements.
Samsung’s research suggests a future where a wearable does not simply record heart rate, sleep, or activity. A model continuously examines how those signals relate to one another and how they change over time.
The broader healthcare industry is moving in a similar direction, with AI increasingly being embedded directly into diagnostic and clinical systems rather than remaining a separate software experiment.
There is still a large gap between experimental AI research and clinically validated consumer health features. Samsung has presented the models as foundational research rather than announcing specific Galaxy Watch features powered by xMAE or HiMAE.
Still, the direction is hard to miss.
The next major upgrade to wearable health technology may not be another sensor.
It may be an AI model that understands the sensors already there.
