Accurate Diabetes Prediction Using a Robust Framework Based on a Hard-Voting Classifier

Abstract

Diabetes is one of the most dangerous diseases that a significant number of people worldwide. An accurate and timely diagnosis of diabetes helps to minimise the overall prevalence of the illness and save the lives of diagnosed individuals. Researchers have proposed a number of diagnostic procedures for the identification of diabetes, but such methods should be improved to guarantee accurate and effective diagnosis. This study aims to develop accurate and timely predictions about diabetes to save the lives of diabetic patients. A three-stage integrated methodology was developed for accurate diagnosis and applied to a Pima Indian Diabetes (PID) Dataset. SMOTE technique was used to balance the dataset and ensure the lack of bias during the training process. The proposed methodology is mainly based on a hard voting classifier that predicts whether a patient will develop diabetes or not. Finally, a set of metrics, namely, accuracy, k-fold cross validation, AUC, precision, recall and f1 score, was used to test the performance of the proposed methodology in diagnosing diabetes. Results showed the superiority of our proposed methodology, with values of 90%, 83.9% with 10-fold cross-validation, 0.901, 0.871, 0.926 and 0.898, respectively.