Unsupervised Domain Adaptation for Fetal R-peak Detection at Trans-Pregnancy Stages based on Multiview Mixing

Fetal heart rate (FHR) monitoring is essential for diagnosing hypoxic-ischemic encephalopathy, but traditional cardiotocography (CTG) suffers from interpretative variability, leading to suboptimal clinical decisions. Non-invasive fetal electrocardiography (fECG), which collects fetal heart electrical activity via maternal abdominal electrodes, offers a more effective alternative. However, deep-learning methods for fetal R-peak detection often require abundant reference signals invasively collected fetal scalp. The scarcity of labeled fECGs at the pregnant stage limits application during pregnancy. To overcome these limitations, we propose an unsupervised domain adaptation approach to enhance detection accuracy during pregnancy. This method incorporates a multiview enhancement strategy to analyze both time- and frequency-domain representations, improving adaptability across pregnancy stages. We further integrate FHR physiological characteristics and propose a misclassification correction mechanism to boost detection accuracy. Evaluations on multiple real datasets confirm the robustness and superior performance of our approach, establishing it as a reliable solution for fECG analysis across different pregnancy stages.