A Hybrid CNN-Transformer with Cross-Attention for Automated Fetal Distress Detection from Cardiotocography

Cardiotocography (CTG) is the standard method for intrapartum fetal monitoring, yet its visual interpretation suffers from high inter-observer variability and limited sensitivity for detecting fetal acidosis. Existing deep learning approaches predominantly treat fetal heart rate (FHR) and uterine contraction (UC) signals independently, ignoring the clinical reality that these signals interact over time. We propose CTG-CrossFormer, a hybrid CNN-Transformer architecture designed to capture this FHR-UC temporal interaction via bidirectional cross-attention. This design directly mirrors how clinicians identify late decelerations in practice. The network uses dual CNN branches with Squeeze-and-Excitation blocks for local feature extraction, followed by a cross-attention module to fuse the signals, and a Transformer encoder to model long-range dependencies. Evaluated on the CTU-UHB Intrapartum Cardiotocography Database (552 recordings, 5-fold patient-level cross-validation), CTG-CrossFormer yields an AUC-ROC of 0.822, outperforming CNN-BiLSTM (0.567), InceptionTime (0.717), and PatchTST (0.735) (p < 0.001, DeLong’s test). Ablation testing shows a 0.087 AUC gain over channel-independent models. Moreover, attention weight visualizations highlight FHR changes linked to uterine contractions, offering clinicians an interpretable view of the model’s decisions.