Preterm Birth Prediction Using GraphSAGE and Electrohysterogram Data

Preterm birth (PTB) affects around 15 million infants every year and causes major health issues, making it a worldwide health concern [4]. The demand for precise early prediction methods has motivated researchers to explore the use of Electrohysterogram (EHG) signals. EHG is a promising approach for predicting preterm labor as it is a non-invasive and affordable method for monitoring uterine activity. This study aims to build a strong PTB prediction model with explainable AI using the publicly accessible TPEHG dataset (300 recordings, 38 preterm). The validation of the model is performed with the TPEHGT dataset (13 preterm and 13 term samples). We employed a graph neural network (GNN) that efficiently aggregates local neighborhood information in graphs and is based on the GraphSAGE architecture. The GraphSAGE model achieved $99.2 \%$ accuracy by training with stratified 10 -fold cross-validation. This was validated through confusion matrices, ROC, and precision-recall curves. In order to guarantee model transparency, the most important factors influencing predictions were found using Shapley Additive explanations (SHAP) analysis, which offered insightful information about the decision-making process. This study emphasizes the value of explainable AI in enhancing clinical decision-making, as well as the possibility of GNNs for PTB prediction.