Machine Learning Applications in Retinopathy of Prematurity Diagnosis Using the ROP Retinal Image Dataset

Purpose: Retinopathy of Prematurity (ROP) is a leading cause of childhood blindness, particularly in preterm infants. This study aims to evaluate the performance of machine learning (ML) models for ROP diagnosis using a publicly available retinal image dataset. Methods: The study utilized the Retinal Image Dataset of Infants and Retinopathy of Prematurity (ROP) from the University Hospital Ostrava, Czech Republic. The dataset comprised 6,004 retinal images and clinical metadata, including gestational age, birth weight, and diagnosis codes. Two ML models, Random Forest and Subspace Discriminant, were implemented. Preprocessing included metadata normalization and stratified sampling into training and testing sets. Model evaluation metrics included accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Results: The Random Forest model achieved an accuracy of 78.62%, sensitivity of 100%, and specificity of 92.21%, indicating strong performance in predicting ROP diagnoses. In contrast, the Subspace Discriminant model underperformed, with an accuracy of 26.90% and no reliable predictions. Feature importance analysis identified gestational age and device type as key predictors. Conclusion: The Random Forest model demonstrated significant potential for automated ROP diagnosis. Future research should explore multimodal data integration, larger balanced datasets, and advanced deep learning models to enhance predictive accuracy and clinical applicability.