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

Area of Science:

• Biomedical Engineering
• Obstetrics and Gynecology
• Signal Processing

Background:

• Preterm birth poses significant maternal and neonatal risks, necessitating early labor detection.
• Electrohysterography (EHG) offers non-invasive uterine monitoring, but understanding synchronization patterns and automating analysis remains challenging.
• Current methods often require manual annotation, limiting clinical utility for preterm labor prediction.

Purpose of the Study:

• To evaluate synchronization-driven propagation patterns in multi-electrode EHG signals.
• To develop an automated framework for early preterm birth detection beyond a one-week window.
• To investigate the clinical applicability of EHG synchronization for predicting preterm labor.

Main Methods:

• Analysis of 452 EHG recordings from 21 to 36 weeks of gestation across various birth types.
• Utilized four electrode configurations and a hybrid deep learning model with multi-head attention and gated fusion.
• Employed random forest for feature selection and inverse-frequency weighting for class imbalance.

Main Results:

• Non-linear synchronization measures demonstrated significant discriminative power (p<0.05).
• Near-term labor showed increased uterine synchronization and spatial coordination, particularly in the upper right region.
• Preterm labor exhibited weaker, less organized EHG activity.
• The automated model achieved high performance: 0.98 accuracy, 0.93 sensitivity, 0.99 specificity, 0.94 F1-score, and 0.99 AUC.

Conclusions:

• Synchronization patterns in EHG signals are key indicators for early and clinically relevant preterm birth detection.
• The proposed automated framework shows robust performance for predicting preterm labor.
• Findings suggest potential for improved non-invasive monitoring and management of high-risk pregnancies.