Objective
To investigate the independent risk factors for Bronchopulmonary Dysplasia (BPD) at different time points within the first week in extremely premature/very low birth weight infants and to construct an early stratification dynamic prediction model for BPD through machine learning, aiming to achieve dynamic prediction of BPD for the early identification of high-risk groups and preemptive prevention.
Methods
A retrospective collection of clinical data was conducted on premature infants admitted to the Neonatology Department of the First Affiliated Hospital of Xinjiang Medical University from January 2017 to December 2022, with gestational age (GA) < 32 weeks or birth weight (BW) < 1500g. Eligible subjects were randomly divided into training and validation sets in a 7:3 ratio for model building and internal validation. Prospective clinical data from preterm infants admitted to six neonatal rescue centers in various districts of Xinjiang from January to October 2023 were independently collected to validate the practical application value of each model. Clinical parameters were collected, and study participants were divided into three groups: no BPD, mild BPD, and moderate to severe BPD (msBPD). Machine learning predictive models for BPD stratification employing logistic regression (LR), random forest (RF), XGBoost (XGB), and gradient boosting decision tree (GBDT) were constructed for postnatal days 1, 3, and 7. Comprehensive evaluation was performed to select the optimal model at each time point and proceed to external validation.
Results
The study retrospectively gathered data from 554 preterm infants (286 no BPD, 212 mild, and 56 msBPD cases). Prospectively, 387 preterm infants (208 no BPD, 138 mild, and 41 msBPD cases). On ordinal logistic regression, GA, BW, prenatal steroids, interruption of umbilical blood flow, severe preeclampsia, FIO2, CRP, RBC, systemic inflammatory response index (SIRI), prognostic nutritional index, platelet mass index, alveolar-arterial oxygen difference, and oxygenation index were independent risk factors for BPD severity at different times after birth. After comprehensive evaluation, the LR and XGB models were identified as better BPD stratification prediction models for postnatal days 1, 3, and 7 (AUC = 0.810,0.837 and 0.813 respectively).
Conclusion
Early stratification dynamic prediction machine learning models for BPD have been constructed for postnatal days 1, 3, and 7 in extremely premature/very low birth weight infants. These may serve as effective tools for the screening of high-risk BPD populations.
