Machine learning analysis of pregnancy-related factors and stillbirth: a retrospective cohort study of 65,000 pregnant women in Shuyang, Suqian, Jiangsu, China, 2019-2024.

INTRODUCTION: Stillbirth remains a critical global health challenge, particularly prominent in developing regions. While China’s stillbirth rate falls below the global average, persistent disparities suggest opportunities for further reduction. This study aims to analyze pregnancy-related factors affecting stillbirth through machine learning to support risk stratification and to inform future studies evaluating targeted prevention strategies. MATERIALS AND METHODS: This retrospective cohort study analyzed electronic medical records (EMR) from 65,884 pregnant women who delivered in Shuyang. Three machine learning models were employed to train on the dataset and identify key variables for stillbirth. The Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set to address the class imbalance. Using the key predictors identified, seven distinct machine learning models were retrained. Model performance was evaluated, with an analysis of area under the curve (AUC) values conducted across both test and validation datasets. To enhance the random forest model's performance, grid search combined with fivefold cross-validation for hyperparameter tuning was utilized and the classification decision threshold was refined. RESULTS: Fetal growth restriction (FGR), poor pregnancy history, birth defects, hyperglycemia in pregnancy, gestational age, folic acid supplementation, and number of prenatal examinations were identified as key predictors. The random forest model demonstrated superior predictive performance for stillbirth, with an AUC of 0.807 in the test set and 0.749 in the validation set, alongside a low missed diagnosis rate of 0.3% and a false positive rate of 0.1%. Through hyperparameter optimization, the model achieved a livebirth recall rate of 99.7% ± 0.1%, with the optimal decision threshold set at 0.217 to balance sensitivity and specificity. CONCLUSION: This study developed and internally validated a machine learning–based prediction model for stillbirth risk using pregnancy-related factors from EMR, which may support risk stratification in the local setting and inform future studies evaluating targeted prevention strategies; external validation is warranted before broader implementation.