A Survey of Deep Learning for Non-Invasive Fetal Electrocardiogram Analysis

Fetal electrocardiogram (fECG) signals recorded from maternal abdominal electrodes are emerging as a promising non-invasive technique for monitoring fetal health. The principal challenge lies in effectively extracting the weak fetal signal from the overlapping maternal electrocardiogram (mECG) components present in abdominal recordings. This review examines the application of deep learning models to address these challenges, enhancing both signal extraction and analysis. We assess various model architectures—including convolutional, recurrent, hybrid, autoencoder, and generative adversarial networks—evaluate their strengths and weaknesses, and discuss the critical settings necessary for optimal functionality, together with a review of open-source datasets. We also outline challenges and future directions, including the need for larger and more diverse datasets, standardized evaluation protocols, improved interpretability, efficient real-time implementations, and advanced modeling approaches such as transformer-based foundation models. These developments will enable deep learning–based fECG analysis to truly provide clinically actionable insights and scalable maternal–fetal monitoring.