Developing and Validating a Machine Learning Model to Predict Brain Injury in Preterm Infants Using Multisource Data...

Area of Science:

• Neonatal neurology
• Machine learning in medicine
• Computational biology

Background:

• Moderate-to-severe preterm brain injury (PBI), including intraventricular hemorrhage (IVH) and periventricular leukomalacia (PVL), is a major cause of poor neurodevelopmental outcomes in preterm infants.
• Early identification of at-risk infants through risk stratification is crucial for timely intervention and improved outcomes.

Purpose of the Study:

• To develop and validate a machine learning model for early risk stratification of preterm brain injury (PBI).
• To identify key clinical predictors for PBI in preterm infants using routinely collected data.

Main Methods:

• Retrospective analysis of 318 preterm infants (development cohort) and 35 infants (external validation cohort).
• Evaluation of 33 candidate predictors from perinatal factors, early laboratory tests, and hospitalization data.
• Development and comparison of seven machine-learning algorithms, with LightGBM selected for final analysis.
• Nested cross-validation, Platt scaling for calibration, and assessment using AUROC, PR-AUC, and Brier score.

Main Results:

• The LightGBM model demonstrated moderate internal discrimination with an AUROC of 0.747.
• Key predictors included ventilation status and early physiological and laboratory indicators.
• Preliminary external validation showed a high AUROC (0.897), though with limitations due to small sample size and wide confidence intervals.

Conclusions:

• An interpretable LightGBM model was developed for PBI risk stratification using readily available early data.
• The model shows potential for clinical utility, with moderate internal performance and positive net benefit.
• Larger multicenter studies are necessary to confirm generalizability and refine the model for routine clinical implementation.