Samsung’s Vision: Preventive, Personalized, Connected Care
Samsung Research America’s Digital Health Team has introduced a technical foundation for its long-term “Connected Care” vision, first articulated at the Health Forum during the July 2026 Galaxy Unpacked event. The company’s roadmap describes a future of preventive, personalized, and connected healthcare, underpinned by advanced health technology and strategic partnerships. Central to this vision is the development of AI models that can efficiently process data from consumer wearables.
“This research is significant because it lays the technical groundwork for delivering health insights that are efficient, precise, and continuous through a health foundation model,” said Sharanya Desai, Head of Digital Health Algorithms at Samsung Research America. “We will continue to develop and advance health foundation models that can be applied to a variety of biosignals and health features that can operate on-device with limited sensors and computing resources.”
The Core Concept: Health Foundation Models
Samsung’s research pivots on the concept of a “health foundation model.” Unlike models trained for a single, specific task, a foundation model uses self-supervised learning to identify inherent features in vast amounts of unlabeled data. In this context, the model is pretrained on large datasets of physiological signals captured by wearables. This pretraining allows one versatile model to support multiple downstream applications, from analyzing specific biosignals and developing new biomarkers to predicting potential health issues.
The company has developed two distinct models, each published and accepted at a prestigious machine learning conference:
- xMAE (Physiology-Aware Masked Cross-Modal Reconstruction): Accepted to the International Conference on Machine Learning (ICML), this model focuses on learning the temporal relationships between different types of biosignals.
- HiMAE (Hierarchical Masked Autoencoder): Accepted to the International Conference on Learning Representations (ICLR), this model is designed to learn health patterns across multiple time scales within wearable time-series data.
xMAE: Linking Continuous PPG to Clinical-Grade ECG
A key challenge in cardiac monitoring is the difference between data collection methods. Electrocardiograms (ECG) measure the heart’s electrical activity directly and are excellent for detecting irregular rhythms like atrial fibrillation. However, they typically require an active, pause-and-measure action from the user. Photoplethysmography (PPG), the technology behind optical heart rate sensors in most smartwatches, detects blood flow changes passively and continuously.
Although PPG and ECG signals originate from the same cardiac event, they are captured through different mechanisms and with a slight time delay. Samsung likens this to seeing lightning before hearing thunder. The xMAE model is trained to learn this precise temporal relationship. It does this through a “masked reconstruction” technique, where it learns to reconstruct missing portions of an ECG signal using only the corresponding PPG data.
“Biosignals are inherently dynamic, with unique time-varying physiological properties. The key contribution of this research lies in proving the viability of health foundation models capable of capturing both the inter-signal relationships and their underlying temporal structures,” explained Subbu Venkatraman, Head of the Digital Health Research Lab at Samsung Research America.
The model was pretrained on approximately 9,400 hours of combined ECG and PPG data. In evaluations, Samsung reports that xMAE outperformed both unimodal (single-signal) models and existing multimodal learning methods in 15 out of 19 tasks. These tasks included predicting cardiovascular disease, detecting abnormal test results, and classifying sleep stages. Crucially, the features learned by xMAE showed potential for transferability across different sensor devices, body locations, and data-gathering environments.
HiMAE: Analyzing Health Patterns Across Short and Long Time Scales
Wearable data contains information at different temporal resolutions. Heartbeats occur in fractions of a second, while sleep cycles and activity patterns unfold over hours. The HiMAE model is engineered to parse both. It uses multiple encoders to analyze short and long data segments separately, allowing it to determine the appropriate time scale for a given health task—whether analyzing rapid heart rate variability or long-term sleep quality.
Like xMAE, HiMAE employs a masked reconstruction training method, enabling it to learn from abundant unlabeled biosignal data. From this single pretrained foundation, the model can then perform a variety of tasks: classification (e.g., sleep stage), numerical prediction (e.g., heart rate), and data generation.
A significant advancement cited by Samsung is HiMAE’s efficiency. The company states it achieved high performance with a smaller model size than existing approaches and can deliver results in less than one millisecond on a smartwatch-class CPU. This performance metric is critical for the feasibility of on-device processing, where computing resources and battery life are limited.
Implications: On-Device AI for Continuous Health Monitoring
The development of these foundation models signals a move towards sophisticated, AI-driven health analysis that resides directly on user devices. By processing data locally, systems can potentially offer more immediate feedback and maintain user privacy by minimizing the need to send raw health data to cloud servers continuously.
Samsung’s research positions these foundation models as a component for new consumer health experiences. The stated goal is to extract diagnostic markers, run predictive health classifications, and generate personalized user guidance—all from consumer hardware. This aligns with the broader industry trend toward leveraging AI to transform passive wearable data into actionable, preventive health insights.
See also: Google AI health coach to use Abbott glucose data