Samsung built AI that reads your Galaxy Watch data on-device

Samsung Research America's Digital Health Team unveiled two AI foundation models built to analyze biosignal data from Galaxy Watch devices, extending the company's push to run more of its AI workload locally rather than in the cloud, the company said, processing heart activity, sleep, and movement data directly on the device rather than sending it to the cloud.

The first model, xMAE, learns the relationship between different types of biosignals rather than just recognizing patterns within a single signal type. It connects electrocardiogram (ECG) readings with photoplethysmography (PPG) data, the optical sensor readings a smartwatch uses to track heart rate, to build a more complete picture of cardiovascular health. Samsung pretrained xMAE on roughly 9,400 hours of ECG and PPG data and reported that it outperformed both single-signal models and existing multimodal approaches on 15 of 19 evaluation tasks, including cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification.

The second model, HiMAE, tackles a different problem: health signals mean different things depending on the time scale you look at them, whether that's seconds, minutes, or hours. HiMAE uses multiple encoders to read biosignals across several time scales simultaneously, and Samsung says it produces results in under one millisecond on smartwatch-class processors, fast enough to run locally without a network connection. xMAE was accepted to the International Conference on Machine Learning, and HiMAE was accepted to the International Conference on Learning Representations, both peer-reviewed venues for machine learning research.

HiMAE demonstrates the potential of on-device health foundation models for the first time.

• xMAE links ECG and PPG signals to build a fuller cardiovascular health picture, trained on ~9,400 hours of data
• xMAE beat existing methods on 15 of 19 evaluation tasks, including cardiovascular disease prediction
• HiMAE reads biosignals across multiple time scales at once and runs in under 1 millisecond on watch-class hardware
• Both models process data on-device rather than sending biosignals to Samsung's servers
• xMAE and HiMAE were presented at the Health Forum during Galaxy Unpacked in July 2026, with no confirmed consumer rollout date yet

Keeping the analysis on-device is as much a business decision as a technical one: health data is among the most sensitive categories a company can collect, and processing it locally sidesteps a chunk of the regulatory and trust questions that come with streaming biosignals to a corporate cloud. It also fits the direction Samsung signaled at this year's Galaxy Unpacked, where the company leaned hard into agentic AI across its foldable lineup; xMAE and HiMAE extend that same on-device AI push into health tracking specifically. Neither model has a confirmed date for reaching consumer Galaxy Watch software, and Samsung hasn't disclosed real-world accuracy figures, battery impact, or a commercialization timeline, so the gap between a peer-reviewed research result and a feature that actually ships on your wrist remains open. Both papers still matter beyond Samsung's own product line: publishing the underlying architecture through ICML and ICLR peer review means other wearable makers and researchers can evaluate, critique, and eventually build on the same on-device approach, rather than taking Samsung's performance claims on faith.

None of this should be read as personalized investment advice.