Precision Risk Model Using Quantitative Assessment of Vascular Severity in Telemedicine-Based Screening.

Abstract

Retinopathy of prematurity (ROP) screening requires frequent examinations to avoid missed treatment-requiring disease, but this approach is burdensome for infants, families, and health systems. Whether precision risk models could reduce examination burden without compromising safety is not known. To develop and externally validate an interpretable risk model integrating gestational age (GA), postmenstrual age (PMA), and vascular severity (artificial intelligence-derived VSS or clinician-assigned P-score) to predict 2-week risk of treatment-requiring retinopathy of prematurity (TR-ROP) and estimate its impacts in screening frequency. This diagnostic study used data from the Imaging and Informatics in ROP (i-ROP) consortium (2011-2022) and the Stanford University Network for Diagnosis of ROP (SUNDROP; 2013-2021). The i-ROP dataset was split into training, validation, and test subsets, and SUNDROP served as an external validation cohort. A subset of i-ROP examinations with clinician assessment scores (P-scores) was analyzed for clinical adaptability. Data were analyzed from September 1, 2024, to September 1, 2025. GA, PMA, and vascular severity (VSS or P-score). Discrimination (area under the receiver operating characteristic curve [AUROC], area under the precision-recall curve [AUPRC]), sensitivity, specificity for predicting TR-ROP within 2 weeks, and the proportion of examinations that could be deferred in a retrospective simulation while maintaining 100% sensitivity. Among 559 infants in the i-ROP training dataset and 1544 in SUNDROP, infants who developed TR-ROP had a mean GA of 2.9 (95% CI, 2.3-3.5) weeks lower and birth weight of 391 (95% CI, 328-454) g than those who did not. Adding vascular severity improved discrimination vs GA alone (AUROC difference 0.13 [95% CI, 0.06-0.19] in i-ROP; 0.08 [95% CI, 0.03-0.14] in SUNDROP). A decision threshold achieved 100% sensitivity with moderate specificity (i-ROP 63%; SUNDROP 73%). Simulated risk-based scheduling reduced 28% (376 of 1384) of examinations in i-ROP and 39% (2356 of 6090 ) in SUNDROP without missing TR-ROP. Substituting P-scores for VSS preserved model performance (AUROC 0.87; 95% CI, 0.76-0.97; sensitivity 100%; 95% CI, 63%-100%). In this study, this validated clinically adaptable model provided individualized visit-level TR-ROP risk assessment with potential to improve screening efficiency by reducing unnecessary examinations without missing TR-ROP. The model will be made publicly available for further validation; however, prospective evaluation within a defined clinical workflow is required prior to routine implementation.