ECG is accurate, but you have to stop what you're doing and press your finger on the watch crown to get a reading. PPG (photoplethysmography) is convenient, but it's long been criticized for lacking precision. This time, Samsung Research America isn't picking a side—instead, it's having the two signals teach each other.

Taipei time, August 14 — according to foreign outlet Artificial Intelligence News, Samsung Research America's digital health team has unveiled two health foundation models: xMAE and HiMAE. Targeting heart activity, sleep, and physical activity data collected from wearables like smartwatches, the models use self-supervised learning to extract features directly from massive amounts of unlabeled signals.

xMAE: Teaching PPG the Timing of ECG

Short for "Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning," xMAE does something very specific — it learns the temporal relationship between PPG and ECG. The team pre-trained the model on over 9,400 hours of PPG and ECG data, allowing it to memorize how the two signals correspond to each other along the time axis. The result: watches can now infer near-ECG-grade cardiovascular features purely from the PPG data they already routinely collect, without requiring users to stop and manually take an ECG reading. According to Samsung, xMAE outperformed existing multimodal approaches on 15 out of 19 evaluation tasks, and the paper has already been accepted at ICML, a top machine learning conference.

HiMAE: Hitting 1 Millisecond on a Smartwatch CPU

If xMAE tackles the question of "what to measure," HiMAE tackles "how to do it fast." This "Hierarchical Masked Autoencoder" uses separate encoders to handle signals at different time scales — fast, short-term changes like heartbeat versus slow, cumulative patterns like sleep quality are processed independently. The model is smaller than existing solutions, yet can produce results in under 1 millisecond on smartwatch-grade CPUs. No cloud computing connection is needed, meaning analysis happens in real time and sensitive health data never has to be uploaded to a server. This research has been recognized at ICLR, the International Conference on Learning Representations.

Sharanya Desai, Digital Health Algorithms Lead at Samsung Research America, said this lays the technical groundwork for delivering efficient, accurate, and continuous health insights, with plans to extend the technology to more devices with limited computing resources going forward. Subbu Venkatraman, Head of the Digital Health Research Lab, emphasized that this proves health foundation models can capture the temporal structure behind dynamic signals, and the team will continue turning research into medical applications that can genuinely improve health outcomes.

Both technologies align with Samsung's "Connected Care" vision, first proposed in July 2026. No specific product launch timeline or compatible device models have been announced yet.