Samsung Research America’s Digital Health Team is charting a new course in wearable technology with the introduction of two advanced AI foundation models. These models are engineered to derive profound insights from biosignals captured by smartwatches, focusing on critical metrics like heart activity, sleep patterns, and physical exertion. This development is a cornerstone of Samsung’s broader “Connected Care” vision, an initiative aiming to usher in an era of preventive, personalized, and interconnected healthcare solutions, bolstered by strategic health technology and healthcare partnerships.
The company’s ambition, as articulated at its Health Forum during Galaxy Unpacked in July 2026, positions these health foundation models as pivotal components in crafting novel consumer health experiences. Sharanya Desai, Head of Digital Health Algorithms at Samsung Research America, underscored the significance of this research: “This work lays the essential technical groundwork for delivering health insights that are not only efficient and precise but also continuous, all powered by a health foundation model.” She further elaborated, “We are committed to developing and refining health foundation models that can be adapted to a wide spectrum of biosignals and health features, capable of operating on-device with minimal sensor and computational resources.”
Samsung’s Groundbreaking Health AI Foundation Models
At its core, a health foundation model leverages self-supervised learning to discern intricate patterns within unlabeled biosignal data. Samsung highlights that by pretraining these models on vast health datasets, a single model can effectively support a multitude of downstream applications. These range from sophisticated biosignal analysis and biomarker development to the prediction of potential health issues, demonstrating a remarkable versatility and efficiency.
Samsung’s research encompasses two distinct models, each designed with unique objectives. The first, xMAE (Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning), is adept at uncovering temporal relationships between disparate biosignals. The second, HiMAE (Hierarchical Masked Autoencoder), excels at identifying health patterns across multiple temporal scales within wearable time-series data.
The academic community has recognized the merit of this research, with xMAE accepted into the prestigious International Conference on Machine Learning and HiMAE slated for presentation at the International Conference on Learning Representations. Both models represent significant advancements in understanding the intricate physiological relationships and temporal structures inherent in biosignal data.
These models address different facets of wearable data analysis. xMAE, for instance, is engineered to correlate two distinct cardiac signals that measure related activity through different physiological mechanisms. HiMAE, on the other hand, delves into data analysis across both short and long temporal intervals, empowering a single pretrained model to adeptly handle tasks such as classification, numerical prediction, and data generation.
xMAE: Bridging Continuous PPG with ECG Signals for Advanced Cardiovascular Insights
Electrocardiograms (ECGs) offer a direct measurement of the heart’s electrical activity, proving invaluable for assessing heart rate, heart-rate variability, and detecting anomalies such as atrial fibrillation. However, traditional wearable ECG readings often necessitate an active, paused measurement by the user.
In contrast, Photoplethysmography (PPG) sensors, commonly integrated into smartwatches, offer a passive approach. PPG monitors changes in blood flow, enabling continuous, unobtrusive data collection. While both ECG and PPG originate from cardiac activity, they exhibit a distinct temporal lag, much like the perceptible delay between seeing lightning and hearing thunder.
Samsung’s xMAE model ingeniously learns this temporal relationship by reconstructing masked segments of ECG data using PPG signals. This innovative design aims to facilitate comprehensive cardiovascular health analysis through continuously collected PPG data, thereby minimizing the need for separate, manual ECG measurements. The model was pretrained on an extensive dataset comprising approximately 9,400 hours of combined ECG and PPG data.
Subbu Venkatraman, Head of the Digital Health Research Lab at Samsung Research America, emphasized the core contribution of this research: “Biosignals are inherently dynamic, possessing unique time-varying physiological properties. The key breakthrough of this research lies in validating the efficacy of health foundation models capable of capturing both inter-signal relationships and their underlying temporal structures.” He further affirmed, “We remain steadfast in our commitment to advancing foundational health AI research and translating it into healthcare solutions that genuinely enhance people’s health and well-being.”
Samsung reports that xMAE demonstrated superior performance compared to unimodal biosignal models and existing multimodal learning techniques across 15 out of 19 evaluation tasks. These tasks spanned critical areas such as cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification. Notably, the features learned by xMAE showed significant potential for application across a diverse range of sensor devices, body locations, and data-gathering environments, hinting at a truly universal applicability.
HiMAE: Unlocking Multiscale Analysis of Wearable Data
Wearable data is rich with information that can manifest differently over various timeframes. Short data segments can capture rapid fluctuations, such as individual heartbeats, while longer segments reveal patterns that develop over extended periods, like sleep cycles or sustained physical activity.
HiMAE employs a sophisticated architecture featuring multiple encoders designed to analyze short and long data segments independently. This allows the model to dynamically identify the most relevant time scale for a given health task. Consequently, heart-rate analysis and sleep prediction can draw upon distinct temporal perspectives within the time-series data, leading to more accurate and context-aware insights.
The model’s training methodology involves reconstructing masked portions of wearable data. Samsung explains that this process enables HiMAE to discern complex patterns from biosignals even when labeled data is scarce. Once pretrained, the model can adeptly support classification, numerical prediction, and data generation from a unified system, streamlining the analytical workflow.
Samsung further highlights HiMAE’s impressive efficiency, noting its ability to achieve high performance with a smaller model footprint than many existing solutions. Crucially, the company reports that HiMAE can deliver results in under one millisecond when running on a smartwatch-class central processing unit. This remarkable on-device processing capability moves critical health analysis away from cloud servers and directly onto the user’s wearable device. Foundation models trained on unlabelled physiological streams thus provide a powerful mechanism for extracting diagnostic markers, performing predictive health classifications, and generating personalized user guidance directly from consumer hardware, all without the need for continuous server connectivity. This paradigm shift promises enhanced privacy, reduced latency, and greater autonomy for users in managing their health.
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/24895.html
