Transformer-based deep learning for estimating bidirectional maternal-fetal cardiac coupling.

Introduction: Maternal-fetal cardiac coupling has emerged as a promising non-invasive marker of fetal autonomic regulation and neurocardiac development. However, existing approaches are limited by unidirectional modeling, handcrafted coupling metrics, and a lack of interpretable and clinically validated abnormality assessment frameworks. Methods: We propose a transformer-based deep learning framework for bidirectional maternal-fetal cardiac coupling analysis using heart rate variability (HRV) features. The model learns fetal-to-maternal (FTM) and maternal-to-fetal (MTF) interactions using two independently trained UNETR-based models with identical architecture and no shared weights. The framework is evaluated on a real clinical dataset, incorporating time-delay, gestational, and classification analyses. Results: Time-delay analysis revealed directional asymmetry, with MTF coupling peaking at 3 seconds and FTM at 5 seconds. Gestational analysis showed relatively stable MTF coupling, while FTM coupling was consistently reduced in abnormal cases. Classification using support vector machines demonstrated strong discrimination, with the full-feature model (MTF + FTM + gestational week) achieving ROC-AUC = 0.91 ± 0.08. Under class imbalance, the coupling-only model (MTF + FTM) achieved the highest PR-AUC, indicating robust minority-class sensitivity. Discussion: These findings suggest that bidirectional coupling analysis may provide useful insight into maternal-fetal physiological interactions. However, the results should be interpreted as proof-of-concept evidence due to the limited and heterogeneous abnormal cohort. Further validation on larger and more homogeneous datasets is required to assess the potential of this approach for non-invasive prenatal monitoring.