A Stacked Generalization Approach for Diagnosis and Prediction of Type 2 Diabetes Mellitus

Abstract

Diabetes mellitus is a global health problem that occurs due to metabolic disorders. It is a lifestyle disease that enhances the likelihood of the development of many serious health complications. In the past, several learners have been applied for prediction of diabetes but there is still enough space to develop classifiers with higher accuracy. The study utilizes Pima Indian Diabetes secondary dataset. In this paper, individual approaches, viz., linear-SVM, kernel methods including polynomial, radial basis function, and sigmoid have been used while among ensembles majority voting and stacking strategies have been applied. Stacked ensembling is based on various types of meta-learners such as C4.5, NB, k-NN, SMO, and logit-boost. The stacking approach with meta-learner SMO (ST-SMO) achieves accuracy 79%, sensitivity 78.9%, false positive rate 31.5%, precision 78.5%, F-measure 77.8%, and AUC 73.2% demonstrating that it is the best classifier as compared to any of the individual and ensemble approaches.