Predicting stillbirth and identifying key maternal risk factors using machine learning

Area of Science:

• Maternal health
• Public health
• Medical informatics

Background:

• Stillbirth is a critical global health issue, especially in low-resource settings.
• Identifying maternal and obstetric risk factors is crucial for stillbirth prevention.
• Machine learning (ML) models offer advanced predictive capabilities beyond traditional logistic regression.

Purpose of the Study:

• To predict stillbirth risk using ML models.
• To identify key maternal and obstetric predictors of stillbirth in Ethiopia.
• To compare the performance of various ML models for stillbirth prediction.

Main Methods:

• Retrospective cross-sectional study using maternal and obstetric records.
• Applied Random Forest (RF), Gradient Boosting Machines, Support Vector Machines, and logistic regression.
• Evaluated models using accuracy, ROC-AUC, and balanced accuracy; utilized SHAP for interpretability.

Main Results:

• Random Forest (RF) achieved 92% accuracy and 0.95 ROC-AUC, outperforming other models.
• Maternal age was the strongest predictor, followed by labor and delivery factors.
• SHAP analysis confirmed predictor importance and explained variable-specific risk effects.

Conclusions:

• Maternal age is a dominant stillbirth determinant, alongside labor/delivery and maternal characteristics.
• ML models, especially RF, accurately predict stillbirth risk and offer interpretable insights.
• Findings support using ML for targeted prenatal care to reduce stillbirths in resource-limited areas.