Samsung's xMAE and HiMAE: On-Device Health AI Models That Analyze Biosignals Without the Cloud

xMAE was pretrained on approximately 9,400 hours of paired ECG–PPG data from about 2,400 subjects. It was then evaluated across six studies covering 19 downstream tasks, including hypertension prediction, ectopic beat detection, sleep staging, and blood lab estimation. The model outperformed all prior open-source foundation models in 15 of 19 tasks by AUROC . A key practical insight: after pretraining, the ECG branch is removed and the model runs on PPG alone, making it well-suited for wearable deployment .

HiMAE: Hierarchical analysis across time scales

HiMAE (Hierarchical Masked Autoencoder) takes a different approach: it analyzes wearable time-series data — heartbeats, sleep stages, activity patterns — across multiple temporal resolutions using a hierarchical convolutional encoder–decoder . Rather than treating all time scales uniformly, HiMAE discovers resolution-specific structure: coarse patterns at longer scales (e.g., sleep/wake cycles) and fine-grained patterns at shorter scales (e.g., individual heartbeats) . The model achieves sub-millisecond inference on a smartwatch CPU, enabling continuous on-device analysis without cloud offloading .

The shared self-supervised learning approach

Both xMAE and HiMAE are masked autoencoders (MAE) trained via self-supervised learning on unlabeled biosignal data . The core idea is simple but powerful: mask a portion of the input signal, then train the model to reconstruct it. This allows the models to capture meaningful physiological structure without requiring manually labeled training data — a critical advantage in health AI, where labeled clinical data can be scarce and expensive to produce.

Role in Samsung's Connected Care vision

At Galaxy Unpacked in July 2026, Samsung introduced its Connected Care strategy, a shift from population-averaged health metrics to preventive, personalized health monitoring powered by on-device AI . Dr. Pak described health foundation models as the key enabler — moving "from population averages to something truly personal" by using self-supervised learning to extract meaningful features from each individual's biosignals .

xMAE and HiMAE are the first concrete Samsung Research models that realize this vision: they run entirely on the smartwatch (no cloud required), learn user-specific physiological patterns, and can be fine-tuned for tasks like blood pressure estimation, arrhythmia detection, and sleep analysis directly on the device . Together, they represent a practical step toward making wearable health AI both private and continuously available — without needing an internet connection.