Interpretable Machine Learning Versus G-ROP Criteria for Screening Type 1 Retinopathy of Prematurity

Purpose: Retinopathy of prematurity (ROP) is a leading cause of preventable childhood blindness. Although the North-American (NA)/American Academy of Ophthalmology (AAO) and Growth and ROP (G-ROP) criteria are highly sensitive, they may trigger unnecessary examinations. This study evaluated whether an interpretable machine-learning (ML) model could reduce screen-positive classifications while identifying all documented Early Treatment for ROP (ETROP) type 1 ROP cases. Methods: A retrospective study included infants admitted to Hadassah Medical Center neonatal intensive care units (January 2023-February 2025). Of 2159 infants, 387 met local screening criteria (gestational age ≤32 weeks or birth weight ≤1500 g); 189 had complete G-ROP predictor data, and 5 were excluded because ROP status remained unavailable. ML models used gestational age, birth weight, and postnatal weight gain. L1-regularized logistic regression models used either gestational age and birth weight alone or these variables plus G-ROP postnatal weight-gain measures. Leave-one-out cross-validation assessed performance. Results: Among 184 infants, 21 had type 1 ROP; all approaches identified all 21 cases. Local, NA/AAO, modified G-ROP, and birth-based criteria classified 184 (100.0%), 174 (94.6%), 137 (74.5%), and 94 (51.1%) infants as screen-positive, respectively. The 5-variable ML model classified 58 (31.5%; specificity = 77.3%) as screen-positive, a 68.5% reduction versus local screening and 57.7% versus G-ROP (area under the receiver operating characteristic curve [AUROC] = 0.914). The birth-only ML model performed similarly (55/184 [29.9%]; specificity = 79.1%; AUROC = 0.909), indicating no incremental screening benefit from postnatal weight gain. Conclusions: Interpretable ML substantially reduced screen-positive classifications while identifying all documented type 1 ROP cases. Similar birth-only performance supports prospective validation of locally adapted screening. Translational Relevance: Routinely collected data and ML framework may enable more efficient type 1 ROP screening and neonatal ophthalmic care.