Comparison of key performance metrics of ensemble learning algorithms for diagnosis of diabetic retinopathy

Around the world, Blindness in diabetic patients are seen commonly because of Diabetic Retinopathy (DR). Sadly, the suggested annual testing of the diabetic patients' eye fundus is too complicated or not a regular manner. It is needed to provide information for the doctors about critical patients who would need regular checkup and others who can be considered as low risk and can be screened after considerable time. This paper explores various performance factors from most popular and leading ensemble algorithms that play key role in classifying and grading diabetic retinopathy. Three ensemble algorithms most widely used in Medical-IT industry used for comparison here are Stacking algorithm, AdaBoost algorithm and XGBoost ensemble classifier. The experiment demonstrates how various performance factors used to compare each classifier helps in deciding about the kind of dataset to be used and also for selecting the attributes. This demonstration uses a reduced set of attributes obtained thru correlation based feature extraction (CFS) method in comparison with original dataset with all characteristics that constitute significant risk factors for deciding whether a patients who are at high risk of diabetic retinopathy. Comparison of specificity and sensitivity levels gained are provided for a deeper comprehension of the behaviors of different ensemble algorithms. This research is therefore a first productive step towards developing a customized system helping in making a good decision.