Samsung just made a major leap in wearable health tech. The company's research division unveiled two AI foundation models - xMAE and HiMAE - that can analyze biosignals like heart rate and sleep patterns directly on smartwatch hardware, no cloud required. Both models earned acceptance at top AI conferences ICML and ICLR, signaling Samsung's push to deliver real-time, personalized health insights as part of its broader Connected Care vision announced at Galaxy Unpacked in July 2026.
Samsung is making a serious play to own the future of wearable health AI. The company's Digital Health Team at Samsung Research America just pulled back the curtain on two foundation models that could fundamentally change how smartwatches understand your body - and they're doing it without sending your data to the cloud.
The timing isn't coincidental. At Galaxy Unpacked in July 2026, Samsung laid out its Connected Care vision - a shift from reactive sick care to preventive, personalized health monitoring. These new AI models, xMAE and HiMAE, are the technical foundation making that vision possible.
Here's what makes them different. Traditional health AI models analyze one biosignal at a time or rely on cloud processing. Samsung's approach does neither. xMAE (Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning) learns the temporal dance between two cardiac signals: electrocardiogram (ECG) and photoplethysmography (PPG). Think of it like teaching AI to predict thunder based on lightning.
ECG directly measures your heart's electrical activity with clinical precision, but it requires you to stop and actively take a reading on your Galaxy Watch. PPG measures blood flow changes passively and continuously through optical sensors. The signals originate from the same heartbeat but appear with a slight time lag. According to Samsung's research published in ICML proceedings, xMAE reconstructs missing ECG data using PPG readings, enabling continuous cardiac monitoring without manual measurements.
The researchers trained xMAE on approximately 9,400 hours of paired ECG and PPG data. The results caught attention in academic circles - the model beat unimodal approaches and existing multimodal methods in 15 of 19 evaluation tasks. We're talking cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification. More impressive, the learned features transferred across different sensor devices, body locations, and environments.
"This means you could get ECG-level cardiac insights 24/7 from just your wrist sensor," the Samsung researchers explained in their published work. That's a game-changer for detecting conditions like atrial fibrillation before they become emergencies.
But xMAE is only half the story. HiMAE (Hierarchical Masked Autoencoder) tackles a different problem: time scale. Your heartbeat happens in milliseconds. Sleep patterns unfold over hours. Physical activity trends emerge across days. Most health AI models pick one time scale and miss the rest.
HiMAE uses multiple encoders to simultaneously analyze short bursts and long segments of biosignal data. According to Samsung's ICLR paper, this hierarchical approach lets the model zoom in on rapid signals like individual heartbeats while also catching slow-burning patterns like sleep quality or activity trends. During training, it reconstructs masked portions of data across these different time scales, learning robust features even with limited labeled datasets.
The efficiency numbers are wild. A single pretrained HiMAE model handles classification, numerical prediction, and data generation - all while being smaller than comparable models. Samsung claims it produces results in under one millisecond on a smartwatch-class CPU. That's fast enough for real-time on-device analysis.
And that's the strategic move here. Apple processes some health data on-device but still leans heavily on cloud infrastructure for complex AI tasks. Google's Fitbit relies on server-side processing for deeper health insights. Samsung is betting that privacy-conscious consumers want their biosignal data analyzed locally, never leaving their wrist.
"HiMAE demonstrates the potential of on-device health foundation models for the first time," Samsung researchers noted in materials reviewed by The Tech Buzz. "Analyzing raw health signals in real time, without relying on cloud servers."
Both models use self-supervised learning, meaning they discover meaningful patterns in unlabeled biosignal data without needing massive manually-tagged datasets. After pretraining on large-scale health data, they can tackle downstream tasks like biomarker development and health issue prediction. It's the same foundational approach that powers large language models, now applied to your body's electrical and optical signals.
The academic validation matters too. Acceptance to ICML (International Conference on Machine Learning) and ICLR (International Conference on Learning Representations) signals peer recognition that Samsung's doing legitimate AI research, not just product marketing dressed up as science.
This puts Samsung in direct competition with emerging health AI players. Meta has explored biosignal analysis for AR/VR applications. Startups like Whoop and Oura have built businesses on advanced wearable analytics. But Samsung has an advantage: vertical integration. They control the hardware sensors, the operating system, the AI models, and the health app ecosystem.
The Connected Care vision Samsung outlined at Unpacked positions these models as infrastructure for partnerships across the healthcare ecosystem. Imagine third-party apps building on Samsung's foundation models to detect early signs of diabetes, monitor chronic conditions, or optimize athletic training - all running locally on Galaxy Watch hardware.
There are obvious questions Samsung hasn't fully answered yet. How do these models perform across diverse populations? Health AI has a documented bias problem when training data skews toward certain demographics. What about battery life when running complex AI on-device continuously? And perhaps most critically, how will Samsung navigate the regulatory landscape as these tools inch closer to medical-grade diagnostics?
The FDA has shown willingness to clear software-based health features, like Apple's ECG and atrial fibrillation detection. But foundation models that analyze multiple biosignals simultaneously and predict health outcomes venture into murkier territory. Samsung will need to balance innovation speed with the cautious pace of medical device regulation.
For now, the research represents a technical milestone. Two researchers from Samsung's Digital Health Team, including those working on xMAE and HiMAE, are highlighted in Samsung's announcement as driving this work forward. Their bet is that the future of health AI lives on your wrist, not in distant data centers.
Samsung's xMAE and HiMAE models represent more than incremental improvement in wearable health tech - they're a strategic bet that the next battleground in consumer AI happens on-device, not in the cloud. By processing biosignals locally with sub-millisecond latency while matching or exceeding cloud-based approaches in accuracy, Samsung is carving out differentiation in a crowded wearables market. The real test comes next: translating peer-reviewed research into features consumers can actually use, navigating medical regulations, and proving these foundation models work across diverse populations. If Samsung pulls it off, every Galaxy Watch becomes a continuous health monitoring station that never phones home. That's a compelling pitch in an era where health data privacy concerns are mounting and consumers increasingly question what happens to their body's signals once they leave their wrist.
