The Application of Machine Learning Models to Predict Stillbirths.

Background and

Objectives

This study aims to evaluate the predictive value of comprehensive data obtained in obstetric clinics for the detection of stillbirth and the predictive ability set of machine learning models for stillbirth. Material and Method: The study retrospectively included all stillbirths followed up at a hospital between January 2015 and March 2024 and randomly selected pregnancies that resulted in a live birth. The electronic record system accessed pregnant women's maternal, fetal, and obstetric characteristics. Based on the perinatal characteristics of the cases, four distinct machine learning classifiers were developed: logistic regression (LR), Support Vector Machine (SVM), Random Forest (RF), and multilayer perceptron (MLP).

Results

The study included a total of 951 patients, 499 of whom had live births and 452 of whom had stillbirths. The consanguinity rate, fetal anomalies, history of previous stillbirth, maternal thrombosis, oligohydramnios, and abruption of the placenta were significantly higher in the stillbirth group (p = 0.001). Previous stillbirth histories resulted in a higher rate of stillbirth (OR: 7.31, 95%CI: 2.76-19.31, p = 0.001). Previous thrombosis histories resulted in a higher rate of stillbirth (OR: 14.13, 95%CI: 5.08-39.31, p = 0.001). According to the accuracy estimates of the machine learning models, RF is the most successful model with 96.8% accuracy, 96.3% sensitivity, and 97.2% specificity.

Conclusions

The RF machine learning approach employed to predict stillbirths had an accuracy rate of 96.8%. We believe that the elevated success rate of stillbirth prediction using maternal, neonatal, and obstetric risk factors will assist healthcare providers in reducing stillbirth rates through prenatal care interventions.