Cross-Subject Domain Adaptation for Generalizable Photoplethysmography-Based Health Monitoring

Abstract

Photoplethysmography has become a widely used optical sensing technique for continuous cardiovascular and respiratory monitoring in wearable devices. Despite its practical advantages, the translation of photoplethysmographic sensing into reliable population-scale health monitoring remains constrained by substantial cross-subject variability arising from skin properties, tissue perfusion, sensor placement, motion artifacts, ambient lighting, and device-specific characteristics. Although deep learning models trained on large datasets can improve average estimation accuracy, their performance often degrades when deployed on unseen subjects because the underlying data distribution shifts across anatomical, demographic, and environmental contexts. This article examines cross-subject domain adaptation as a systems-level strategy for building generalizable photoplethysmography-based health monitoring. The discussion moves beyond narrow algorithmic formulations and considers domain adaptation within a broader sociotechnical infrastructure that includes data governance, sensor architecture, deployment logistics, fairness, regulatory accountability, and sustainability. The paper analyzes structural trade-offs among centralized training, on-device adaptation, federated learning, and hybrid inference arrangements. It further explores how unlabeled target-domain data, adversarial alignment, contrastive pretraining, attention mechanisms, and quality-aware signal processing can be integrated into a robust monitoring pipeline. The central argument is that generalizability is not merely a statistical property of a model but an emergent outcome of system design, data curation, institutional oversight, and continuous evaluation. This perspective has important implications for clinical validation, health equity, and the responsible scaling of wearable diagnostics beyond controlled laboratory settings.