An Ensemble Learning Approach Towards Prediction of Diabetic Retinopathy

Diabetic retinopathy can be called an eye-related disorder where due to diabetes, the patient experiences problem in vision induced by blood vessel injuries in retina of the eye. Computational intelligence is a relevant domain in detecting as well as predicting the disease. Providing predictive competency to the model is a hectic task. Recent works have shown the use of deep neural networks in detecting and predicting diabetic retinopathy. This work focuses on use of ensemble learning techniques to predict positiveness of the concerned disease. Also, the proposed framework entails the use of ensemble classifiers which can be used in early detection of diabetic retinopathy. Techniques to tackle the problem of overfitting is also scrutinized. Specific works in this domain are classified based on the accuracy of classification. Even deep learning frameworks can aid in classifying imagery datasets. This work focuses on the uses of a dataset extracted from well-known Messidor data which specifies certain indispensable features which determine whether a sample is undergoing diabetic retinopathy. The features considered in the dataset are microaneurysms detection, exudates, macula, amplitude and frequeny modulation, and many such. The dataset is subjected to diverse classifiers and the performance is determined through ensemble learning with the aid of majority voting technique. The classifiers considered for this study are decision tree, linear support vector machines, Naïve Bayes, and random forest. The best fit achieved after majority voting is by linear SVM i.e., an accuracy of 77.48%. The paper further provides the need to switch to deep neural networks in future to predict diabetic retinopathy through imagery datasets.