Construction and Assessment of Machine Learning Models of Brain Injury Risk Prediction in Preterm Twin Neonates

Objective: To construct an interpretable model based on prenatal clinical data to predict the risk of brain injury in preterm twin neonates. Methods: This retrospective study enrolled 1415 parturients with twin pregnancies (gestational age: 28–36⁺⁶ weeks) who received cesarean delivery from 2018 to 2022. Neonatal brain injury was diagnosed by postnatal cranial ultrasound, covering intraventricular hemorrhage, periventricular leukomalacia and hypoxic–ischemic brain injury, and graded into mild, moderate and severe levels. Taking twin pregnancy as the analysis unit, participants were divided into brain injury group (n=484) and non-brain injury group (n=931). The dataset was split into training and internal test sets at a 7:3 ratio via stratified sampling. We screened predictive variables using the LASSO method and established four models: random forest, SVM, logistic regression and XGBoost. Model predictive efficacy was assessed by AUC, and calibration curves, PR curves, DCA and SHAP analysis were applied for visualization and interpretability evaluation. Results: Multivariate logistic regression demonstrated that fetal growth restriction (OR=6.74, 95% CI: 2.83–16.08) and placental adhesion (OR=1.72, 95% CI: 1.30–2.69) were independent risk factors for brain injury in preterm twin neonates, while increased gestational age (OR=0.58, 95% CI: 0.50–0.69) and elevated maternal hemoglobin level in the third trimester (OR=0.80, 95% CI: 0.70–0.93) were independent protective factors. All four models exhibited moderate discrimination in the test set, with AUC values ranging from 0.733 to 0.776. The random forest model was selected as the optimal model after comprehensive assessment of sensitivity and AUC. SHAP analysis revealed that the contribution to model prediction ranked as follows: gestational age, pre-pregnancy BMI, maternal hemoglobin level in the third trimester, gestational weight gain, and gestational diabetes mellitus. Conclusion: The random forest model constructed with prenatal risk factors presents favorable predictive value for brain injury in preterm twin neonates, and can serve as an auxiliary tool for neurological risk stratification.