Abstract
INTRODUCTION: Retinopathy of prematurity (ROP) is a leading cause of childhood blindness. However, current screening guidelines may be overly broad, necessitating better models to detect high-risk infants. METHODS: From a multicenter cohort of 103,701 infants (3,301 [3.2%] treated for ROP) discharged from 298 neonatal intensive care units from 2006 to 2017 with birth weight ≤1,500 grams or gestational age ≤30 weeks, we used clinically relevant variables to develop machine learning (ML) models at 2-week intervals from postnatal day 14 to 98 to stratify infants by ROP treatment timing. We assessed model performance by concordance index, area under the receiver operating characteristic curve (AUROC), and average precision (AP), validated performance in a cohort of 25,105 infants across 231 sites from 2018 to 2020, and compared model performance to a logistic regression (LR) model. RESULTS: In the validation cohort, the day 28 ML model outperformed the LR model by AUROC (0.916 [0.905-0.926] vs. 0.903 [0.892-0.914]; p < 0.001) and AP (0.190 [0.167-0.217] vs. 0.160 [0.140-0.183]; p < 0.001). Using the ML model at a 100% sensitivity threshold would have negative predictive value of >99.9% and could reduce the number of infants needing screening by 14% compared to current guidelines. CONCLUSION: ML models can effectively predict the need for ROP treatment and stratify infants by risk, potentially reducing unneeded screening. Future work is needed to translate model-based ROP predictions to the clinical setting.