Inferring Optical Tissue Properties from Photoplethysmography Using Hybrid Amortized Inference

The smart wearable enables continuous tracking of validated biomarkers such as heart rate, heart rate variability, and blood oxygen saturation through photoplethysmography (PPG). Beyond these metrics, PPG waveforms contain rich physiological information, as recent research in deep learning (DL) shows. However, DL models often rely on features with unclear physical meaning, creating a tension between predictive power, clinical interpretation, and sensor design. We address this gap by introducing PPGen, a biophysical model that relates PPG signals to understandable physical and optical parameters. Building on PPGen, we propose hybrid amortized inference (HAI), which allows fast, robust, and scalable estimation of relevant life parameters from PPG signals while correcting for model misspecification. In extensive in-silico experiments, we demonstrate that HAI can accurately describe physiological parameters under various noise and sensory conditions. Our results show a path toward PPG models that maintain the fidelity required for DL-based features while supporting clinical interpretation and design of informed computing platforms.

• † Isomorphic Labs
• ** Work done while at Apple
• ‡ Share originality