Abstract
Introduction
Retinopathy of prematurity (ROP) remains an important cause of preventable visual impairment in preterm infants. Current screening relies primarily on gestational age and birth weight, which may not fully capture the heterogeneous postnatal clinical course of very preterm infants. This study aimed to develop and internally validate an interpretable machine-learning framework using routinely collected clinical variables for clinical-course-based in-hospital ROP risk stratification.
Methods
A total of 236 infants were retrospectively included, comprising 106 infants with ROP and 130 infants with normal ROP screening findings during hospitalization. Fifteen routinely collected maternal and neonatal clinical variables were used in stratified five-fold internal validation. Within each fold, AutoGluon trained 13 individual candidate models and automatically constructed its built-in WeightedEnsemble_L2 from selected base-model predictions. Missing-data imputation and classification-threshold selection were performed within the training data of each fold. A separate 15-predictor LightGBM model was trained using the same clinical-variable framework and evaluated for TreeSHAP-based interpretation and local research-prototype implementation.
Results
WeightedEnsemble_L2 was selected as the primary analytical model because of its overall discrimination and screening-oriented performance. It achieved an AUC of 0.823, sensitivity of 0.764, specificity of 0.677, positive predictive value of 0.659, and negative predictive value of 0.779. The Brier score was 0.171, with a calibration intercept of −0.095 and a calibration slope of 1.046. In the separately trained LightGBM model, mechanical ventilation duration, oxygen therapy duration, birth weight, bronchopulmonary dysplasia, length of hospital stay, and prematurity-related encephalopathy were the most influential predictors.
Discussion
The proposed framework provides an interpretable, clinical-course-based approach to in-hospital ROP risk stratification by integrating routinely collected NICU data. It may enhance individualized risk assessment by incorporating clinical-course information beyond conventional static birth characteristics. Multicenter external and prospective validation is warranted to further evaluate its clinical applicability.