Objective: Bronchopulmonary dysplasia (BPD) is a costly chronic pulmonary disorder affecting preterm infants with extremely low gestational age (GA). This study aimed to overcome the limitations of traditional prediction models by employing Machine Learning (ML) to discern and compare predictive variables for BPD using routinely collected clinical data.



Methodology: This research, conducted at Shahid Akbarabadi Hospital, Tehran (2019–2024), included 86 neonates with BPD and 83 without BPD with GA < 35 weeks. Maternal and neonatal features were collected. Eleven diverse ML methods, including ensemble schemes, were analyzed using Monte Carlo Cross-Validation (MCCV) and robust metrics like accuracy, Cohen's kappa, and the Area Under the Receiver Operating Characteristic curve (AUC).



Results: Ensemble modeling (XGBoost and Random Forest) demonstrated substantially better and more stable performance than conventional models. The XGBoost model showed the best prediction performance: AUC of 0.94, kappa of 0.72, and accuracy of 0.86. Feature significance analysis (via RF) presented major predictors: GA (0.222), birth length (0.148), birth weight (0.129), and echocardiogram results (0.074). SHAP plots revealed that lower values for these factors (GA, weight, and length) were associated with a higher risk of BPD.



Conclusion: This study shows significant promise for applying ML algorithms (specifically XGBoost) in prospectively predicting BPD risk using initial clinical data. Identifying key perinatal predictors facilitates the development of automated, personalized screening tools in neonatal care, enabling earlier prophylaxis. Future directions include external validation using new datasets and incorporating dynamic perinatal data.