Developing and Validating a Machine Learning Model to Predict Brain Injury in Preterm Infants Using Multisource Data
Pu Xu1,2, Ying Li2, Ying Chen2
1Capital Institute of Pediatrics, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing 100020, China.
Background:
Moderate-to-severe preterm brain injury (PBI), including intraventricular hemorrhage (IVH) and periventricular leukomalacia (PVL), remains an important cause of adverse neurodevelopmental outcomes in preterm infants. Early risk stratification using routinely collected clinical data may help prioritize surveillance in vulnerable infants.
Methods:
We retrospectively included 318 preterm infants admitted between 2015 and 2024 as the development cohort. Thirty-three candidate predictors derived from perinatal factors, first laboratory tests within 24 h of admission, and selected early hospitalization variables were evaluated. Seven machine-learning algorithms were developed using stratified 10 × 5 nested cross-validation with prespecified preprocessing, class-balancing, and feature-selection procedures. Candidate models were compared primarily using the mean fold-level area under the receiver operating characteristic curve (AUROC). After model selection, the finalized LightGBM model was calibrated using Platt scaling, and its pooled out-of-fold (OOF) performance was summarized. Two prespecified thresholds (Youden and high-sensitivity) were used for risk stratification. A small independent temporal cohort of 35 infants was used for preliminary external validation.
Results:
PBI occurred in 62/318 infants (19.5%) in the development cohort and 6/35 infants (17.1%) in the temporal external cohort. During candidate-model comparison, LightGBM achieved the highest mean fold-level AUROC (0.768, 95% CI 0.708-0.825). The finalized 14-feature LightGBM model, evaluated using pooled OOF predictions after Platt calibration, yielded an AUROC of 0.747 (95% CI 0.679-0.811), a PR-AUC of 0.392, and a Brier score of 0.136. At the Youden threshold (0.18), sensitivity was approximately 0.70 and specificity approximately 0.85; at the high-sensitivity threshold (0.10), sensitivity was approximately 0.95 and specificity approximately 0.50. Key predictors included ventilation status and early physiologic and laboratory indicators. In the small temporal external cohort (n = 35), the AUROC was 0.897 (95% CI 0.672-1.000); however, this high point estimate should not be overinterpreted because of the limited sample size, wide confidence interval, and suboptimal calibration, and should therefore be considered preliminary.
Conclusions:
We developed an interpretable LightGBM model using routinely available early postnatal and early hospitalization data to support risk stratification for PBI in preterm infants. The model showed moderate internal discrimination and a positive net benefit across clinically relevant thresholds. Preliminary temporal external validation in a small cohort yielded highly uncertain estimates; larger multicenter studies are needed to confirm generalizability, refine calibration, and determine the most appropriate implementation strategy before routine clinical use.
