Majority Voting Ensemble Approach for Predicting Diabetes Mellitus in Female Patients from Unbalanced Dataset

Diabetes is a common yet deadly disease among women. In 2017, about one in every nine women were found to be exposed to diabetes in the USA alone. It is needless to say the importance to predict it before it can turn deadlier for the body. In this paper, we used machine learning algorithms to predict the possibility of diabetes in women based on the behavioral and medical diagnosis data related to diabetes. The dataset used here was the Pima Indians Diabetes Dataset by National Institute's of Diabetes and Digestive and Kidney Diseases. The unbalance in the dataset was dealt using synthetic minority oversampling technique (SMOTE). Various machine learning algorithms were opted as base for the prediction. The research work focused on applying hard voting or majority voting ensemble technique on the various combinations of those base algorithms. The aggregation of XGBoost, KNN and Random Forest gave the best performance. According to the classification reports, the precision, recall and, f-1 scores were found to be very good compared to most of the recent and earlier works. The model could predict diabetes with an accuracy of 86%. SHAP analysis was then used to finding out the impact of features on the model performance.