Multimodal Preterm Birth Prediction Using Electrohysterogram Signals and Clinical Data via a Hybrid CNN-BiLSTM Ensemble Framework

Preterm delivery is one of the major contributors to mortality and morbidity among infants. Infants born prematurely are more susceptible to respiratory, neurological, and other complications that can cause developmental issues throughout their lives. It is therefore crucial to identify pregnancies at high risk of preterm delivery as early as possible to take timely interventions, reducing mortality and complications among infants. Electrohysterography, which detects the electrical activity of uterine muscles, is a promising tool for the early prediction of preterm delivery. It detects changes in uterine contractions, which are precursors to delivery. However, existing machine learning approaches for analyzing electrohysterography signals, which are based on statistical features, are inadequate for effectively extracting features from uterine muscle signals. The current research examines the application of EHG signals along with maternal clinical features for preterm birth forecasting. Specifically, a novel combination of CNN-BiLSTM networks was introduced to generate both spatial and temporal features from time-frequency EHG spectrograms. The second branch consists of extracting manually crafted EHG features and clinical features using a Random Forest classifier. At last, predictions obtained from the two branches were fused by stacking techniques. The experimental results on the TPEHGDB database achieved ROC-AUC scores of 0.88 and 0.90 for the hybrid and ensemble models, respectively.