Cross-Biosignal Pretraining for Wearable Health AI: Reflecting the Underlying Biosignal Structure

xMAE is a masked cross-modal reconstruction framework that leverages the physiological relationship between PPG (photoplethysmography) and ECG (electrocardiography). PPG, which is easily collected on wearable devices, is a downstream response to the heart's electrical signal, while ECG corresponds to its upstream cause. xMAE directly leverages this directional dependency relationship in training.

During pretraining, 10-second synchronized ECG-PPG pairs are used as input, masking continuous segments of ECG and training the model to reconstruct them from PPG. To prevent simple interpolation from bypassing the masking, continuous block masking is applied, and directional cross-attention (masked ECG tokens → query, PPG tokens → key/value) is used to infer cardiac electrical timing from the peripheral pulse. A curriculum masking strategy stabilizes training by keeping more ECG visible early on and gradually increasing reliance on PPG.

After pretraining, the ECG branch is discarded, and only the trained PPG encoder is used for downstream tasks. Pretrained on approximately 3.4 million synchronized recordings (~9,400 hours, 2,400 subjects), xMAE achieved higher classification performance across 6 studies and 19 downstream tasks (cardiovascular disease, sleep stage classification, blood test results, demographics) compared to existing open-source foundation models, using less pretraining data.