Novel Deep Learning-Based Models for Fetal Electrocardiogram

Abstract

This dissertation presents innovative deep learning approaches to enhance non-invasive fetal monitoring, addressing longstanding challenges in maternal and fetal healthcare. Central to this effort is W-net-transformer, a transformer-based architecture that extracts fetal electrocardiogram (ECG) from maternal abdominal signals with F1 score of 99.88% and 98.9% on the real ADFECGDB and PCDB datasets, respectively. Its robustness and efficiency are further demonstrated by identifying optimal abdominal lead placements during early and late gestational periods, ensuring reliable performance even in real-world settings with minimal computational demands. Building on these advancements, this work explores the prediction of delivery timing from maternal abdominal ECG, leveraging signal quality assessments to achieve remarkable accuracy within clinically significant windows. The study underscores the importance of time window selection and lead positioning in improving prediction outcomes, setting new benchmarks for obstetric care. Complementing these efforts, the research extends fetal monitoring to phonocardiography (PCG), a cost-effective alternative to traditional ECG. The proposed fetal heart sounds U-NetR (FHSU-NETR) model accurately extracts fetal and maternal heart beat locations from raw PCG signals with detecting arrhythmias, showcasing its potential for deployment in resource-limited settings. By integrating deep learning with non-invasive techniques, this model enhances the accessibility and reliability of fetal health assessments. Complementing these contributions is an investigation into maternal-fetal ECG coupling using deep learning, offering novel insights into the dynamic interplay between maternal and fetal heart activity. This study enhances the understanding of fetal development and maternal-fetal interactions, paving the way for new diagnostic tools and approaches in fetal monitoring. Together, these contributions underscore the transformative potential of advanced computational models in maternal and fetal diagnostics, offering scalable, real-time solutions that address diverse clinical needs across varied healthcare environments.

Date of Award	13 May 2025
Original language	American English
Supervisor	AHSAN Khandoker (Supervisor)