Bridging the Domain Gap in Wearable Photoplethysmography: A Comparative Study of Domain Adaptation Strategies for Cross-Domain Generalisation in Atrial Fibrillation Detection
Main article
Abstract
Deep learning models for atrial fibrillation (AF) detection from wearable photoplethysmography (PPG) frequently fail to maintain performance when deployed on data whose acquisition conditions differ from those of the training corpus, a deterioration that is the direct consequence of the gap between the source domain on which a classifier is trained and the target domain encountered at inference. Domain adaptation offers a principled route to recover the lost performance by realigning the two domains, and a broad spectrum of strategies has emerged, ranging from statistical moment matching through adversarial feature learning to generative signal reconstruction. This article presents a comparative study of representative domain adaptation strategies for cross-domain AF detection from wrist-worn PPG. We organise the strategies into a unified taxonomy, evaluate them within a common framework in which a single downstream classifier is held frozen while only the adaptation component varies, and benchmark each family under controlled distribution shift induced by graded signal-quality degradation. Across two target cohorts, every adaptation strategy improved cross-domain discrimination relative to an unadapted baseline, with generative reconstruction and adversarial alignment yielding the largest and best-calibrated gains, while statistical alignment offered a favourable accuracy-to-complexity trade-off. The findings substantiate the position that, when a significant domain gap exists, adaptation improves generalisation, and that several distinct adaptation approaches are available to practitioners. We close with practical guidance and open challenges for trustworthy deployment.
