Pre-Conference Talk by PHAM Hung Manh | PPG-IDR: Leveraging User Identity for Robust Cross-user PPG Sensing via Disentangled Representations

Wearable devices are increasingly able to monitor our health continuously, using signals such as photoplethysmography (PPG) to estimate heart rate, blood pressure, stress, sleep, and other physiological states. However, a fundamental challenge remains: a model that performs well for people seen during training may perform substantially worse when deployed to a new user. This cross-user generalization gap limits the reliability and scalability of wearable health systems in real-world settings. In this talk, I will examine this problem from an identity perspective. Beyond measuring physiological states, PPG signals also contain surprisingly strong biometric characteristics that can reveal who a person is. We show that such identity information remains encoded even in modern PPG foundation-model representations, suggesting that models may inadvertently rely on user-specific patterns rather than purely physiological information. Motivated by this observation, we introduce PPG-IDR, a representation-learning framework that explicitly separates medical information from identity-specific characteristics. PPG-IDR uses dedicated medical and identity representations together with orthogonality constraints, adversarial learning, and self-supervised learning to suppress identity leakage while preserving physiologically meaningful information. Across multiple datasets and six health-monitoring tasks, PPG-IDR consistently improves performance when models are evaluated on previously unseen users.

This is a Pre-Conference talk for ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp/ISWC 2026).