PPG-Based Heart Rate Estimation Using Unsupervised Domain Adaptation

Recent advancements in wireless sensors have introduced photoplethysmography (PPG) sensors for heart rate estimation. However, accurate estimation remains a challenge because of motion artifacts affecting signal precision. While deep learning-based approaches show promise in addressing MAs, they often require subject-specific training or fine-tuning to address inter-subject variability within PPG sensors, demanding extensive labeled data collection for each new subject. Therefore, this study explores the application of unsupervised domain adaptation (UDA) techniques to mitigate inter-subject variability within PPG sensors and enhance prediction performance on new subjects without individual labeling. Implementing five state-of-the-art UDA methods, we demonstrate their effectiveness in heart rate estimation compared to supervised learning methods. Moreover, we analyze and interpret these results based on the characteristics of each UDA method.