A robust deep learning framework for automated fetal behavioral state classification: leveraging multi-center datasets to improve antepartum heart rate monitoring

Fetal behavioral states reflect the developing brain's capacity to organize behavior into distinct sleep state patterns and are a window to evaluate the maturation of the nervous system throughout gestation. This study presents a deep learning-based methodology for the identification of behavioral states from fetal heart rate traces (FHR) recorded via cardiotocography (CTG). A Residual Attention U-Net was pre-trained on a large unlabeled dataset of over 7000 FHR recordings and fine-tuned on a multi-center dataset comprising 236 FHR signals with labeled states. These signals were collected across multiple countries using different cardiotocographs and annotated by clinicians with diverse backgrounds. The model showed high agreement with expert clinicians on a stratified hold-out test set, averaging a Macro F1-Score of 91% and Balanced Accuracy of 93% in distinguishing between Active and Quiet states. The model has the potential to enhance the evaluation of fetal behavioral states in clinical practice by assisting clinicians and ensuring objective, rapid, and reproducible results, as well as supporting future enhancements in electronic fetal monitoring.