Samsung has introduced two AI foundation models, xMAE and HiMAE, designed to analyze wearable health data like heart activity and sleep patterns directly on consumer devices.
Key facts
- •Samsung Research America developed xMAE and HiMAE to analyze wearable biosignals like heart rate and sleep data.
- •xMAE learns the relationship between ECG and PPG signals to analyze cardiovascular health without manual ECG measurements.
- •HiMAE processes data across multiple time scales and can run on a smartwatch CPU in less than one millisecond.
- •The models were trained using self-supervised learning on approximately 9,400 hours of ECG and PPG data.
- •xMAE outperformed other models in 15 of 19 evaluation tasks related to disease prediction and health classification.
Samsung Research America’s Digital Health Team has unveiled two new AI foundation models, xMAE and HiMAE, designed to process biosignal data from wearable devices. These models aim to provide continuous, precise health insights by analyzing data such as heart activity, sleep, and physical movement. The research, which was discussed during the July 2026 Galaxy Unpacked event, focuses on enabling advanced health diagnostics to run locally on smartwatches with limited computing resources.
By the numbers
The xMAE and HiMAE Models
The xMAE (Physiology-Aware Masked Cross-Modal Reconstruction) model focuses on temporal relationships between different biosignals. It specifically connects ECG data, which measures electrical heart activity, with PPG data, which tracks blood flow. By reconstructing masked ECG signals from PPG data, xMAE allows for the analysis of cardiovascular features without requiring users to manually pause for active ECG readings. HiMAE (Hierarchical Masked Autoencoder) is designed to analyze health patterns across multiple time scales. It uses multiple encoders to process both short-term segments, such as heartbeats, and long-term segments, such as sleep patterns. This allows a single pretrained model to support various tasks, including classification and numerical prediction, while operating on smartwatch-class hardware in under one millisecond.
Research Performance and Implementation
Samsung reports that xMAE outperformed existing unimodal and multimodal learning methods in 15 out of 19 evaluation tasks, including cardiovascular disease prediction and sleep-stage classification. The models were trained using self-supervised learning on unlabeled biosignal data, which the company notes is particularly useful when labeled data is limited. By enabling these models to run on-device, Samsung aims to reduce the need for continuous cloud server connectivity. The company states that this approach supports its broader 'Connected Care' vision, which seeks to deliver preventive and personalized health experiences through consumer technology.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by Artificial Intelligence News.
