Diabetic Retinopathy Detection: A Machine-Learning Approach Based on Continuous Glucose Monitoring Metrics

Diabetic Retinopathy (DR) is an extremely common complication of diabetes mellitus (DM) and a timely treatment may decelerate its progression before the occurrence of irreversible vision loss. Machine learning (ML) represents a powerful tool for addressing the massive screening burden, nowadays performed with the time consuming and operator dependent analysis of fundus photography. Continuous glucose monitoring (CGM) are wearable devices whose information could be exploited also in real-time. This study aimed to explore the potential of CGM and ML for DR detection. A classification task was pursued to identify DR class (n = 50) from the non-DR class (NDR, n = 28) based on data from anthropometric characteristics and extracted CGM metrics. Among the tested models, Logistic Regression achieved the best performances (72.7% of classification accuracy), with a balanced number of misclassifications accounting for less than 30% of misclassified cases. The approach could be suitable for real-time applications.