Automated early preterm detection using hybrid multimodal attention network with gated fusion

Preterm birth is a major pregnancy complication associated with increased maternal and neonatal risks, making early detection of labor essential. Electrohysterography (EHG), provides a non-invasive approach for monitoring uterine dynamics. However, global synchronization patterns and propagation dynamics remain insufficiently understood, and many existing methods rely on manual annotation, limiting clinical applicability. This study evaluates synchronization-driven propagation patterns in multi-electrode EHG signals and proposes an automated framework for early preterm detection, beyond the commonly studied one-week prediction window before delivery. The analysis included 452 recordings from pregnancies between 21 and 36 weeks of gestation, covering spontaneous, cesarean, induced, and induced-cesarean cases. Four electrode configurations (S1-S2, S1-S3, S2-S3, and Multichannel) were examined. A hybrid deep learning model integrating multi-head attention and a gated fusion mechanism was used to combine EHG segments with synchronization features. Feature selection was performed using random forest top-k selection, and class imbalance was handled through inverse-frequency weighting. Results showed that non-linear synchronization measures, including non-linear interdependence directionality index and normalized permutation cross mutual information, were statistically significant (p<0.05) and provided strong discrimination. Our results suggest increased synchronization and spatial coordination near term, with relatively stronger activity in the upper right uterine region. In contrast, preterm activity appeared weaker and less organized. The model achieved an accuracy, sensitivity, specificity, F1-score, and an area under the curve (AUC) of 0.98, 0.93, 0.99, 0.94, and 0.99, respectively, indicating robust performance. These findings highlight synchronization as a key marker for early and clinically relevant preterm detection.