Prediction of stillbirth using machine learning methods.

Area of Science:

• Perinatal Medicine
• Artificial Intelligence in Healthcare
• Reproductive Health Research

Background:

• Stillbirth remains a significant global health concern, impacting numerous pregnancies annually.
• Predictive models are crucial for early identification and intervention to reduce stillbirth rates.
• Existing prediction methods often lack accuracy or are not tailored to specific demographic groups.

Purpose of the Study:

• To develop and validate a machine learning model for predicting stillbirth risk.
• To identify key predictive variables using SHAP values for model simplification.
• To assess the model's performance in a large, multi-center cohort of singleton pregnancies.

Main Methods:

• Retrospective analysis of 32,953 singleton pregnancies from South Korean multi-centers.
• Development of Extreme Gradient Boosting Machine models using baseline, E1 (pre-13 weeks), and T0 (pre-28 weeks) data.
• Validation using Area Under the Curve (AUC) and Area Under Precision-Recall Curve (AUPR) metrics, with model simplification via SHAP values.

Main Results:

• The model demonstrated good predictive performance for all stillbirths (AUC 0.720-0.740) and late stillbirths (AUC 0.781).
• A simplified model for late stillbirth achieved comparable performance (AUC 0.759), highlighting key predictive variables.
• The models showed effectiveness using data available before 28 gestational weeks.

Conclusions:

• The developed machine learning model offers a promising tool for predicting late stillbirth risk in singleton pregnancies.
• The model's utility may be particularly relevant for East Asian populations.
• Early risk assessment using this model could facilitate timely interventions and improve perinatal outcomes.