Samsung health AI models analyse wearable biosignal data
Samsung Research America introduced two AI foundation models, xMAE and HiMAE, designed to extract health insights from wearable biosignal data using self-supervised learning on unlabeled physiological streams. xMAE learns temporal relationships between PPG and ECG cardiac signals, enabling cardiovascular analysis through continuous passive PPG monitoring without requiring active ECG measurements, and outperformed existing methods in 15 of 19 evaluation tasks. HiMAE employs a hierarchical masked
Analysis
TL;DR
- Samsung Research America introduced two AI foundation models, xMAE and HiMAE, designed to extract health insights from wearable biosignal data using self-supervised learning on unlabeled physiological streams.
- xMAE learns temporal relationships between PPG and ECG cardiac signals, enabling cardiovascular analysis through continuous passive PPG monitoring without requiring active ECG measurements, and outperformed existing methods in 15 of 19 evaluation tasks.
- HiMAE employs a hierarchical masked autoencoder architecture with multiple encoders to analyze wearable data across short and long time scales, achieving high performance with a smaller model that runs in under one millisecond on a smartwatch-class CPU.
- Both models were pretrained on large-scale unlabeled biosignal datasets, demonstrating that foundation models can support diverse downstream tasks including classification, numerical prediction, biomarker development, and data generation from consumer-grade wearable hardware.
- The research positions on-device health AI as a viable path toward preventive, personalized, and connected care, reducing reliance on cloud connectivity and enabling continuous health monitoring directly from smartwatches.
Why It Matters
Samsung's health foundation models represent a significant step toward deploying clinically useful AI directly on consumer wearable devices, bridging the gap between passive biosignal collection and actionable health insights without requiring cloud dependency. For AI practitioners and healthcare technology developers, this work demonstrates that self-supervised pretraining on large unlabeled physiological datasets can produce general-purpose models capable of handling multiple downstream health tasks, setting a precedent for how foundation models might be adapted beyond text and vision into the biomedical domain. The emphasis on on-device inference at sub-millisecond latency also addresses critical privacy and connectivity constraints that have historically limited real-world deployment of health AI systems.
Technical Details
xMAE (Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning): A multimodal foundation model that learns the temporal relationship between photoplethysmography (PPG) and electrocardiogram (ECG) signals by reconstructing masked portions of ECG data from continuous PPG input. Pretrained on approximately 9,400 hours of paired ECG and PPG data, it enables cardiovascular feature extraction from passive PPG measurements alone, eliminating the need for manual active ECG readings. Evaluated across 19 tasks including cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification, outperforming unimodal and existing multimodal baselines in 15 tasks.
HiMAE (Hierarchical Masked Autoencoder): A single foundation model using multiple encoders operating at different temporal resolutions to analyze both short-segment signals (e.g., heartbeats) and long-segment patterns (e.g., sleep cycles, physical activity trends). Trained via masked reconstruction on unlabeled wearable time-series data, it supports classification, numerical prediction, and data generation from one pretrained system. Achieves high performance with a smaller model footprint and runs in under one millisecond on smartwatch-class hardware, enabling fully on-device inference.
Self-Supervised Pretraining Paradigm: Both models leverage self-supervised learning to identify features in unlabeled biosignal data, addressing the scarcity of labeled medical data. The pretraining approach allows transfer to diverse downstream health tasks without task-specific retraining from scratch.
Publication Venues: xMAE was accepted to the International Conference on Machine Learning (ICML); HiMAE was accepted to the International Conference on Learning Representations (ICLR).
Industry Insight
- The convergence of foundation model architectures with wearable biosignal processing signals a shift from reactive, clinic-based diagnostics toward continuous, preventive health monitoring powered by consumer devices. Companies building health AI pipelines should prioritize self-supervised pretraining strategies that maximize the utility of abundant unlabeled wearable data before fine-tuning on smaller labeled datasets.
- Sub-millisecond on-device inference for a foundation model is a notable engineering achievement that removes cloud dependency, addressing both latency and privacy concerns. This sets a competitive bar for other hardware manufacturers and health-tech companies aiming to deploy sophisticated AI directly on resource-constrained edge devices.
- Samsung's Connected Care vision, positioning health foundation models as one component of broader healthcare partnerships, suggests that future competitive advantage will depend less on isolated model performance and more on ecosystem integration—combining on-device AI, clinical validation, and healthcare provider networks to deliver end-to-end preventive care solutions.
Disclaimer: The above content is generated by AI and is for reference only.