DETECTION AND DATA MINING OF DIABETIC RETINOPATHY USING CLASSIC MACHINE LEARNING

Abstract

Diabetic retinopathy (DR) is a serious vision threat to the increasing diabetic population worldwide. Many studies employed machine learning for DR classification. However, the merits and demerits of various machine learning algorithms were rarely compared. Moreover, some well-established risk factors were rarely integrated into DR risk prediction scores to improve the model performance.

We implemented 11 machine learning algorithms to detect DR cases among 2791 diabetic patients from the Singapore Epidemiology of Eye Diseases Study. Six datasets were constructed by including different subsets of the 5 feature categories: traditional risk factors, extended risk factors, retinal parameters, serum metabolites, and genetic profiles.

We found that ensemble learning, gradient boosting decision tree, and regularized logistic regression had the best performance. We also identified the most important features associated with DR from each feature category.

This study illustrated the great potential of various machine learning algorithms in DR detection and biomarker selection.