Enhanced Detection of Diabetic Retinopathy Using Ensemble Machine Learning: A Comparative Study

Abstract

Diabetic Retinopathy (DR), a microvascular complication, stands as one of the leading causes of vision impairment among diabetic populations globally.This pathology is characterized by the occlusion of retinal vessels, thereby depriving the retinal tissue of essential nutrients.Given the progressive nature of DR and its potential to culminate in irreversible blindness, timely and accurate diagnosis is paramount for effective intervention.Conventionally, the detection of DR relies heavily on the expertise of ophthalmologists, a resource-intensive process that may be prohibitive in terms of cost and time.To address these limitations, automated detection systems have been developed, aiming to hasten diagnostic processes and democratize access to these crucial services.Nevertheless, the performance of such systems has been historically hampered by the scarcity of reliable data sources and medical records for this condition.In response to these challenges, this study explores an ensemble machine learning approach that synergizes multiple established classifiers into a cohesive diagnostic model.The proposed methodology demonstrates superior performance in accuracy compared to prevalent classification algorithms.Utilizing the Messidor dataset, the top-performing five and ten features were isolated into four subdatasets through InfoGainEval and WrapperSubsetEval methods.The accuracy achieved for the top five features via InfoGainEval was 70.7%, while for the complete feature set, it reached 75.1%.The employment of ensemble machine learning techniques in diagnosing DR represents a significant application of artificial intelligence within the medical domain, conferring advantages such as enhanced accuracy, robustness, efficient feature selection, early detection, scalability, and a reduction in human error, all while ensuring costefficiency and enabling continuous monitoring for improved patient outcomes.However, the approach is not without limitations.These include the quantity and quality of data, clinical variability, ethical and privacy concerns, scalability challenges, potential overfitting, intricate feature selection and engineering, bias in data collection, and issues related to cost and accessibility.The findings underscore the efficacy of the sub-datasets, which facilitate a less cumbersome classification process as compared to the full Messidor dataset, thereby streamlining the diagnostic pathway.