Samsung Research Introduces xMAE and HiMAE Health AI Models for Wearable Biosignals

Samsung Research America has published two AI foundation models — xMAE and HiMAE — that learn to interpret biosignals from wearable devices without needing labelled training data, with one model small enough to run on a smartwatch CPU in under a millisecond, marking what the company claims is the first on-device health foundation model to run in real time without cloud connectivity.

xMAE — Teaching AI the Body's Own Timing

The first model, xMAE (Physiology-Aware Masked Cross-Modal Reconstruction), was accepted to the International Conference on Machine Learning (ICML) 2026 — one of the most competitive publication venues in AI. The model addresses a subtle but important limitation in previous wearable health AI: most self-supervised learning systems treat different biosignals as interchangeable views of the same data, ignoring the fact that those signals have a meaningful temporal relationship to one another.

A canonical example is the pairing of electrocardiography (ECG) and photoplethysmography (PPG). The ECG records the electrical activation that triggers each heartbeat; the PPG records the resulting peripheral pulse as blood reaches the fingertip or wrist — a process that typically lags the ECG by around 250 milliseconds. xMAE explicitly models that directional delay, learning how one signal anticipates and informs the other. The result is a representation model that understands the physiology behind the numbers, not just the statistical patterns. That distinction becomes critical when the downstream task is something clinically meaningful — detecting atrial fibrillation, predicting fainting risk, or characterising continuous cardiac state from a wrist-worn device.

HiMAE — Health AI That Runs on the Watch Itself

The second model, HiMAE (Hierarchical Masked Autoencoder), was accepted to the International Conference on Learning Representations (ICLR) 2026. Its central insight is that wearable biosignals carry useful information at radically different time scales simultaneously — a short-horizon view for beat-to-beat heart rate, a long-horizon view for sleep stage classification — and that a single model architecture should be able to serve both without retraining.

HiMAE achieves this with multiple parallel encoders that analyse short and long signal segments separately, then synthesise their representations. During training, it learns by reconstructing intentionally masked portions of the biosignal data — a self-supervised technique that allows the model to extract meaningful patterns even when labelled clinical data is scarce. The architecture supports classification, numerical prediction and data generation from a single pretrained model. Critically, it is compact enough to produce results in less than one millisecond on a smartwatch-class CPU. Samsung Research's paper describes HiMAE as demonstrating, for the first time, the potential of on-device health foundation models that analyse raw health signals in real time without relying on cloud servers.

Why Foundation Models Change the Wearable Health Equation

The term "health foundation model" describes an AI system pretrained on large-scale, unlabelled biosignal data using self-supervised learning — and then applied to a wide range of downstream health tasks without needing to be retrained from scratch for each one. That generalisation property is what separates foundation models from the traditional approach of building separate, narrow classifiers for each condition. A single pretrained xMAE or HiMAE model could, in principle, be fine-tuned to detect cardiac arrhythmias, track recovery from exercise, estimate blood oxygen saturation trends, or flag sleep-disordered breathing — all from the same base weights.

For Galaxy Watch users, that means each health feature Samsung ships could benefit from the same deep biosignal understanding rather than requiring a separate AI stack per application. The research was developed by the Digital Health Team at Samsung Research America, led by researchers including Sharanya Desai and Subbu Venkatraman, and ties directly into Samsung's Connected Care vision unveiled at Galaxy Unpacked July 2026 — a framework for moving healthcare from reactive treatment toward preventive, personalised and continuously connected experiences. This health AI trajectory parallels the broader AI-for-medicine push seen in Novo Nordisk and AWS's drug discovery partnership, though Samsung's focus is the consumer wearable layer rather than pharmaceutical R&D.

The On-Device Breakthrough and What It Unlocks

HiMAE's sub-millisecond inference speed on a smartwatch CPU is the detail that turns this from interesting research into a viable product architecture. Cloud-dependent health AI faces irreducible latency, privacy concerns about continuous health data transmission, and connectivity requirements that make continuous monitoring impractical. An on-device model sidesteps all three. It also opens the door to health insights that are genuinely continuous — not sampled every 30 seconds and batched to the cloud, but running persistently on the wrist, always available, always current.

The implications extend beyond Samsung's own ecosystem. Both xMAE and HiMAE are general-purpose biosignal architectures. Their acceptance at ICML and ICLR signals that the academic community views the underlying techniques — physiology-aware cross-modal pretraining and hierarchical multi-scale masked autoencoding — as genuine methodological contributions, not proprietary product engineering dressed up as research. That distinction matters for adoption: other wearable health platforms building on Samsung Health infrastructure, or competing devices pursuing similar foundation model approaches, will find the published techniques immediately applicable. The connection between frontier AI model safety work and wearable health is still emerging, but as on-device models gain clinical-grade capability, the governance questions around health data and model behaviour will follow. For now, the story is simpler: Samsung has shown that a foundation model that genuinely understands physiology can fit on your wrist and respond in the time it takes for a single heartbeat to complete.