Statistical and Machine Learning Methods for the Classification of Type -2 Diabetes Mellitus

Abstract

Diabetes mellitus is a chronic disorder of glucose metabolism either due to the relative deficiency of Insulin or due to the failure of body to response to the action of insulin. Diseases leads to critical life threatening complications, if it remains undiagnosed and untreated. There is growing incidence of this disease all over the world. It has assumed epidemic form in Asian countries, China and India. Changing to urban life style, metabolic syndrome and obesity are some factors responsible for this increased prevalence. It is most common (90%) among the different types of diabetes. It is recognised by raised fasting or post prandial blood sugar level. The review of literature shows that one model does not always give good performance in all situations and data frames. The study aimed to address this issue of classification by the use of statistical and machine learning models. Present study is conducted on Pima Indian diabetic dataset which is publicly available and is widely used for diabetes research. Unlike other datasets, medical datasets usually suffer from missing and unwanted data which needs proper handling so that classification accuracy is significantly improved. Therefore, in this study mean, median, and linear regression based imputing techniques are developed and outliers are handled using winsorizing approach. Classification accuracy on imputed data are compared with the accuracy on raw data. Results on mean and median based imputed data significantly outperformed as compared against the raw data.