Sequential Prediction of Bronchopulmonary Dysplasia from Birth Through the First Week of Life Using Machine Learning

Background/Objectives: Bronchopulmonary dysplasia (BPD) risk evolves as clinical information accumulates after birth. This study evaluated the incremental predictive value of routinely available data from birth through the first week of life using cumulative, time-aligned feature sets.

Methods

The primary analysis of this retrospective cohort study included 2153 preterm infants who survived to 36 weeks’ postmenstrual age and had ascertainable BPD status and complete predictor data. The outcome was clinically documented BPD at 36 weeks’ postmenstrual age, based on the treating neonatologists’ assessment of continued supplemental oxygen and/or respiratory-support requirements. Three cumulative feature sets were constructed using variables available at or immediately after birth (Step 1), additional clinical and laboratory variables available within the first 24 h (Step 2), and further clinical-course variables available through postnatal day 7 (Step 3). Logistic regression, random forest, XGBoost, LightGBM, and CatBoost models were developed using a stratified 70:30 training–test split and five-fold stratified cross-validation. Internal validation assessed discrimination, calibration, and Brier scores in the held-out test set. AUROC confidence intervals and paired differences between prediction windows were estimated using 2000 bootstrap resamples. Predictor contributions were examined using SHapley Additive exPlanations. A multiple-imputation sensitivity analysis evaluated logistic regression and random forest in all 2616 infants with ascertainable BPD status.