Towards Emotion-Aware Wearables: Distilling PPG Foundation Models for Real-Time Affective and Physiological Monitoring

Abstract

Photoplethysmography (PPG) is a physiological signal natively suited for affective computing: non-invasive, continuously acquired from wearables, and directly linked to autonomic responses underlying stress, arousal, and emotional states. Yet deploying high-performance PPG foundation models in real-world settings remains constrained by their computational demands. This work presents a knowledge distillation (KD) framework that compresses the Pulse-PPG foundation model into ultra-lightweight 1D-MobileNetV1 and V2 architectures, enabling emotion-aware physiological inference directly on resource-constrained wearable edge devices. Through soft-target and feature-alignment losses, the framework transfers rich affective representations to compact student networks evaluated on three benchmark datasets - WESAD (stress and affective state recognition), PPG-DaLiA (heart rate estimation and activity classification), and SDB (sleep disturbance detection) - covering five downstream tasks. Distilled models with as few as 0.15M parameters achieve competitive performance, with compression ratios up to 190× in model size and 300× in computational load (MFLOPs), while maintaining near-teacher accuracy on affective state classification (AUROC 0.90 vs. 0.81 for the teacher on WESAD 4-class). These results establish a concrete pathway toward privacy-preserving, real-time affective monitoring on wearables - a critical enabler for human-centered health and mental well-being applications.