T he research focuses on two models with different goals. xMAE, short for Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning, learns temporal relationships between different biosignals. HiMAE, or Hierarchical Masked Autoencoder, learns health patterns across multiple time scales in wearable time-series data.
Samsung says xMAE was accepted to the International Conference on Machine Learning, while HiMAE was accepted to the International Conference on Learning Representations. The company describes both models as research into physiological relationships and temporal structures within biosignal data.
The models tackle different aspects of wearable-data analysis. xMAE connects two cardiac signals that capture related activity through different mechanisms. HiMAE examines data over both short and long intervals, allowing one pretrained model to support classification, numerical prediction, and data generation.
xMAE links continuous PPG data to ECG signals
Electrocardiograms, or ECGs, directly measure the heart’s electrical activity. Samsung says ECG is useful for measuring heart rate and heart-rate variability, as well as identifying abnormal heart rhythms and risks linked to conditions such as atrial fibrillation.
Wearable ECG readings generally require users to pause and take an active measurement. Photoplethysmography, or PPG, works differently. PPG detects changes in blood flow and can operate passively through sensors built into wearable devices such as smartwatches.
Both signals reflect cardiac activity, but they occur with a time difference. Samsung compares this relationship to hearing thunder after seeing lightning. xMAE learns that timing relationship by reconstructing masked portions of an ECG signal using PPG data.
The approach is designed to analyse cardiovascular-health features from continuously measured PPG data without requiring separate manual ECG measurements. The model was pretrained using about 9,400 hours of ECG and PPG data.
Subbu Venkatraman, Head of the Digital Health Research Lab at Samsung Research America, commented: “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.
“We remain committed to advancing foundational health AI research and translating it into healthcare solutions that meaningfully improve people’s health and wellbeing.”
Samsung reports that xMAE outperformed unimodal biosignal models and existing multimodal learning methods in 15 of 19 evaluation tasks. These tasks covered cardiovascular disease prediction, abnormal test-result detection, and sleep-stage classification. The company also says the learned features showed potential across different sensor devices, body locations, and data-gathering environments.
HiMAE analyses wearable data across time scales
Wearable data can reveal different information depending on the time period being examined. Short segments can capture fast-changing signals such as heartbeats, while longer segments can reveal patterns that develop over time, including sleep and physical activity.
HiMAE uses multiple encoders to analyse short and long data segments separately. Samsung says this setup allows the model to identify the time scale most useful for a particular health task. Heart-rate analysis and sleep prediction can therefore draw on different portions of the same time-series data.
The training process reconstructs masked sections of wearable data. Samsung says this helps HiMAE learn useful patterns from biosignals even when labelled data is limited. The resulting model can then support classification, numerical prediction, and data generation from a single pretrained system.
Samsung says HiMAE delivers strong performance while using a smaller model than existing alternatives. The company also reports that it can produce results in less than one millisecond on a smartwatch-class central processing unit.
That processing speed suggests the model’s analysis can take place directly on the device rather than relying on cloud servers. Foundation models trained on unlabelled physiological streams could help extract diagnostic markers, run predictive health classifications, and generate user guidance from consumer hardware, all without requiring continuous server connectivity.
Comments