Classification of diabetic retinopathy using stacked machine learning approach on low resource dataset

This study focuses on classification of Diabetic Retinopathy (DR) using Machine Learning (ML) algorithms on low resource dataset. We conduct the computation on a publicly available dataset comprising of just 19 features detailing 1151 instances. We deploy eight popular ML algorithms for classification and observe the best accuracy of 77.8%, and an F1-Score of 0.712 using the Logistic Regression algorithm. Further stacking of all the ML algorithms present an accuracy of 81.3% with an F1-Score of 0.784. Moreover, the confusion matrix reflects false negative values of about 12% only (37 points out of 306 instances). The results fare better in comparison with prior art. Prior art in this context relates to those proposals which are purely ML based. A certain amount of proposals in classification of DR exhibit initial feature extraction using Deep Learning (DL) models followed by classification using ML algorithms. We exclude these proposals containing partly engaged DL models. The performance metrics used in our evaluation of our proposed model are accuracy, precision, recall, F1-Score, Cohen-Kappa score, and RoC-AuC curve. Our model being based purely on ML algorithms perform the task much faster than the DL counterparts.