Samsung Research America’s Digital Health Team disclosed two new AI foundation models built to interpret wearable biosignal data on-device in research published in mid-2026, with coverage from Samsung’s newsroom and AI trade outlets in August confirming the models, named xMAE and HiMAE, process heart activity, sleep and physical activity data locally on a smartwatch rather than sending it to the cloud for analysis. Samsung has not announced a consumer feature, release date, or a specific Galaxy Watch model that will ship with either technology, describing both as research projects for now.
What xMAE and HiMAE actually do
xMAE, formally “Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning,” is designed to learn temporal relationships between different biosignal streams — for instance, how heart rate variability patterns relate to sleep stage transitions — rather than analyzing each signal in isolation. HiMAE, or “Hierarchical Masked Autoencoder,” is built to recognize health patterns across multiple time scales in wearable time-series data, from second-to-second heartbeat variation up to weeks-long trends. Samsung reports xMAE was accepted to the International Conference on Machine Learning and HiMAE to the International Conference on Learning Representations, two of the most competitive peer-reviewed venues in AI research, and says xMAE outperformed both single-signal biosignal models and existing multimodal approaches on 15 of 19 evaluation tasks.
Why on-device processing is the differentiator
Most consumer health AI today, including many features on Apple Watch and Fitbit devices, relies at least partly on cloud processing to run more computationally demanding models. Samsung’s pitch is that running foundation models locally on a smartwatch avoids sending continuous, highly sensitive biosignal data off the device, addressing privacy concerns that have dogged health wearables as they’ve become de facto medical monitoring tools. On-device processing also removes network latency and dependency, meaning health insights could theoretically work without an internet connection, though it demands models efficient enough to run on a smartwatch’s limited processor and battery.
Where this fits Samsung’s broader “Connected Care” push
Samsung previewed a “Connected Care” vision at a Health Forum held alongside Galaxy Unpacked in July 2026, describing an ambition to shift digital health from reactive treatment toward preventive, personalized monitoring. That framing places Samsung alongside Apple, whose Apple Watch AI screening was shown in the landmark EQUAL trial to quadruple atrial fibrillation detection in older adults, and hearing-aid makers like Starkey that have begun marketing their devices as fall-detection and health-tracking tools rather than purely audio devices — evidence that the entire wearables category is repositioning around passive, continuous health monitoring rather than single-purpose fitness tracking.
Why some researchers are withholding judgment
Because xMAE and HiMAE remain research models without a shipped consumer feature, independent clinicians have not yet had the chance to validate their real-world diagnostic performance outside Samsung’s own benchmark tasks. Academic peer review at ICML and ICLR evaluates methodological rigor and technical novelty, not clinical validity for detecting or predicting disease, so acceptance at those conferences does not equate to FDA clearance or clinical validation. Digital health skeptics have also noted a recurring pattern in the wearables industry of research papers preceding shipped products by years, or in some cases never translating into a consumer feature at all.
The competitive stakes
Wearable-based passive health monitoring has become one of the most closely watched fronts in consumer AI, with Whoop recently raising $575 million at a $10.1 billion valuation and Apple, Samsung and Google all racing to embed increasingly sophisticated AI health models directly into smartwatches and rings. Samsung’s emphasis on foundation models — broad, adaptable AI systems rather than narrow single-purpose algorithms — mirrors the industry-wide shift toward foundation-model architectures also visible in FDA-cleared clinical tools like Aidoc’s CARE imaging model, suggesting foundation models are becoming the default architecture across both clinical and consumer health AI.
What’s next
Samsung has given no timeline for when, or whether, xMAE and HiMAE will appear in a shipping Galaxy Watch, and the company has not disclosed plans for regulatory engagement should either model eventually support a diagnostic claim rather than general wellness tracking. The Health Forum’s stated ambition of “connected care” suggests Samsung’s next public step will likely come at a future Unpacked event, where the real test will be whether these research-stage models translate into features validated against real clinical outcomes rather than benchmark accuracy alone.
