Prediction of retinopathy of prematurity using machine learning models.

backgroundEarly identification of neonates at risk of retinopathy of prematurity (ROP) is essential to prevent vision loss. The goal of this study was to develop, test, and validate machine learning (ML) models to predict which newborns will develop ROP and require treatment.

methodsA real-world, de-identified, clinical dataset including 23,404 medical records from 9205 individuals was abstracted, curated, and processed from 25 newborn care units in Odisha, India. A sequential two-step strategy was used: STEP 1 used demographic, perinatal, laboratory, and NICU variables to predict the development of any form of ROP; STEP 2 incorporated structured findings from specialist ophthalmic examinations, including stage, zone, plus/pre-plus status, and related severity descriptors, to identify which ophthalmically assessed infants would require treatment for ROP. The models used structured clinical variables rather than raw retinal images. Eight ML models were trained and tested for each step. Model performance was evaluated prospectively in a separate cohort from the same population.

resultsThe highest-performing models were Random Forest for STEP 1 and LightGBM for STEP 2. In the prospective validation cohort from the same population, the sequential framework had 84.9% sensitivity (437/515) for treatment-requiring ROP; 78 treatment-requiring infants were not identified.

conclusionThe sequential framework shows promise as a structured-data, workflow-aligned approach for risk stratification in a resource-constrained public-health setting. Its current performance supports further development as adjunctive prioritization support within existing guideline-based screening pathways; model outputs should not be used to defer specialist ophthalmic assessment without additional threshold optimization, calibration, external validation, and implementation testing. IMPACT: Early identification of neonates at risk of retinopathy of prematurity (ROP) is essential to prevent vision loss. We developed a two-step machine learning framework using a real-world clinical dataset of 23,404 records from 9205 infants. In a prospective validation cohort, the sequential framework had 84.9% sensitivity (437/515) for treatment-requiring ROP. These findings define the framework's role as adjunctive risk-stratification and prioritization support within existing guideline-based screening pathways; model output should not be used to defer indicated ophthalmic examinations. With further model optimization, such a staged approach may help resource-constrained networks prioritize referrals while preserving the safety net of routine screening.