Dynamic Risk-Stratification Models for Bronchopulmonary Dysplasia in Extremely Preterm Very Low Birth Weight Infants.

Abstract

Objective: This study aimed to identify independent risk factors for bronchopulmonary dysplasia (BPD) at multiple postnatal time points in extremely preterm (EP) or very low birth weight (VLBW) infants and to develop machine learning-based dynamic prediction models for early risk stratification and intervention.

Methods: This study utilized retrospective data from EP or VLBW infants (gestational age (GA) < 32 weeks or birth weight (BW) < 1500 g) admitted to the First Affiliated Hospital of Xinjiang between 2017 and 2022. The dataset was randomly divided into training (70%) and validation (30%) cohorts. Prospective data from six Xinjiang neonatal centers (January-October 2023) were collected for external validation. Infants were classified into three groups: no BPD, mild BPD, and moderate-to-severe BPD. Four machine learning algorithms-logistic regression (LR), random forest, XGBoost (XGB), and gradient boosting decision tree-were trained using clinical data from postnatal days 1, 3, and 7. The most predictive models were selected for external validation.

Results: The retrospective cohort included 554 infants (no BPD: 286; mild: 212; msBPD: 56), and the prospective cohort comprised 387 infants (no BPD: 208; mild: 138; msBPD: 41). Ordinal logistic regression identified significant independent risk factors for BPD severity, including GA, BW, prenatal steroids, umbilical flow interruption, severe Pre-eclampsia, FIO₂, C-reactive protein, red blood cell count, systemic inflammatory response index, prognostic nutritional index, platelet mass index, alveolar-arterial oxygen difference, and oxygenation index. The LR and XGB models demonstrated the highest predictive performance for BPD stratification on days 1, 3, and 7 (Area under the curve: day 1 = 0.810, day 3 = 0.837, day 7 = 0.813).

Conclusion: Machine learning-based dynamic prediction models for BPD were successfully developed and validated using data from postnatal days 1, 3, and 7. These models facilitate early identification of EP/VLBW infants at high-risk of BPD, supporting timely and targeted interventions to improve neonatal outcomes.