Development and validation of machine learning models for predicting bronchopulmonary dysplasia risk in preterm neonates based on antenatal determinants: a retrospective cohort study

Abstract Background Bronchopulmonary dysplasia (BPD) remains one of the most common and serious complications in preterm infants. This study aimed to develop and validate machine learning (ML) algorithms to predict the risk of BPD in preterm neonates based on potential antenatal risk factors. Methods This retrospective study included 644 preterm infants with gestational age (GA) < 32 weeks who were hospitalized between January 2019 and December 2022. Of these, 450 and 194 infants were assigned to the training and internal validation cohorts, respectively. An independent external validation external validation cohort comprising 141 infants born between January 2023 and December 2023 was also included. Clinical data, including maternal prenatal history and early neonatal parameters, were obtained from the electronic medical records system. Least absolute shrinkage and selection operator was used for feature selection. Ten ML models were developed using the selected variables. Model performance was assessed based on discrimination (AUROC), calibration, and discrimination were evaluated in both internal and external validation cohorts. Results The overall incidence of BPD among preterm infants with GA < 32 weeks was 35.09%. Univariate logistic analysis identified 13 statistically significant predictors of BPD (p < 0.05). Multivariable logistic regression analysis confirmed GA, gravidity, and prenatal immunosuppressants use as independent risk factors. Among the ten ML models evaluated, LR algorithm demonstrated the best overall performance, with the highest F1 score and consistent results across both internal and external validation cohorts. Conclusions GA, gravidity, and prenatal immunosuppressants use are independent predictors of BPD in preterm infants. The LR model showed superior predictive performance and robustness across datasets, suggesting its potential utility in clinical decision-making for early BPD risk stratification.