Background: Preterm brain injury (PBI) is a prevalent complication in preterm infants, leading to the destruction of critical structural and functional brain connections and placing a significant burden on families. The timely detection of PBI is of paramount importance for the prevention and treatment of the condition. However, the absence of specific clinical manifestations in the early stages of PBI renders it susceptible to misdiagnosis and missed diagnoses. Moreover, once it occurs, there is no specific treatment available. The aim of this study was to develop and validate a machine learning (ML) based interpretable model for the early detection of PBI, as well as the assessment of patient-wide and individual risk factors for this disease.
Methods: This study utilized a cohort of premature infants provided by Northwest Women's and Children's Hospital in China, comprising medical records of 650 premature infants, spanning from 2019 to 2021. PBI were identified based on cranial magnetic resonance imaging (MRI). Fourteen machine learning models were employed with stratified 10-fold cross-validation method used to evaluate model performance. The Shapley Additive Explanations (SHAP) method was applied for model interpretation. Feature selection methods were used to determine the final model which was validated on the independent test set. Subsequently, risk factors for the entire cohort and individual patients were assessed.
Results: Among the fourteen machine learning models, the CatBoost model demonstrated the best discriminative ability. Following feature selection, the final model was constructed using seven features, designated as PBIPred (Preterm Brain Injury Predictor). PBIPred exhibited strong performance in both 10-fold cross-validation and independent test set (AUC = 0.8229) for accurately predicting PBI. The screening for risk factors in the cohort and individuals identified the following variables as positive risk factors for PBI: Mechanical ventilation (MV), Weight, Anemia of prematurity (AOP), Respiratory distress syndrome (RDS), Albumin (ALB), and White blood cell (WBC).
Availability and implementation: The PBIPred webserver and PBIPred tool were developed for clinical diagnosis and large-scale local medical record data prediction. They can be accessed freely at http://pbipred.liaolab.net and https://github.com/chikit2077/PBIPred, respectively.
Keywords: Machine learning; Model interpretation; Prediction model; Preterm brain injury.
Copyright © 2024 Elsevier B.V. All rights reserved.