Samsung Research has developed health foundation models that learn from wearable biosignals, aiming to power on-device insights without heavy sensors. The work spans two models: xMAE, which learns the temporal relationship between different cardiac signals, and HiMAE, which finds health patterns across multiple time scales.
xMAE reconstructs masked parts of an ECG signal from PPG data, letting a wearable infer heart rhythm features from passive optical readings instead of requiring an active ECG measurement. Pretraining used roughly 9,400 hours of ECG and PPG data, and the model was accepted to the International Conference on Machine Learning.
HiMAE, accepted to the International Conference on Learning Representations, analyzes data at short and long intervals so one pretrained model can support classification, numerical prediction and data generation. Both use self-supervised learning on unlabeled biosignal data.
Samsung describes the research as groundwork for its Connected Care vision, previewed at Galaxy Unpacked in July, where health technology and partnerships feed preventive, personalized care. The company says the models can operate on device with limited sensors and computing resources, a step toward continuous health insights without sending data to the cloud.
