Abstract
Glucose is the primary source of energy for cells, which are the building blocks of life. It is given to the body by insulin that carries out the metabolic tasks that keep people alive. Glucose level imbalance is a sign of diabetes mellitus (DM), a common type of chronic disease. It leads to long-term complications, such as blindness, kidney failure, and heart disease, having a negative impact on one's quality of life. In Saudi Arabia, a ten-fold increase in diabetic cases has been documented within the last three years. DM is broadly categorized as Type 1 Diabetes (T1DM), Type 2 Diabetes (T2DM), and Pre-diabetes. The diagnosis of the correct type is sometimes ambiguous to medical professionals causing difficulties in managing the illness progression. Intensive efforts have been made to predict T2DM. However, there is a lack of studies focusing on accurately identifying T1DM and Pre-diabetes. Therefore, this study aims to utilize Machine Learning (ML) to distinguish and predict the three types of diabetes based on a Saudi Arabian hospital dataset to control their progression. Four different experiments have been conducted to achieve the highest results, where several algorithms were used, including Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbor (K-NN), Decision Tree (DT), Bagging, and Stacking. In experiments 2, 3, and 4, the Synthetic Minority Oversampling Technique (SMOTE) was applied to balance the dataset. The empirical results demonstrated promising results of the novel Stacking model that combined Bagging K-NN, Bagging DT, and K-NN, with a K-NN meta-classifier attaining an accuracy, weighted recall, weighted precision, and cohen's kappa score of 94.48%, 94.48%, 94.70%, and 0.9172, respectively. Five principal features were identified to significantly affect the model accuracy using the permutation feature importance, namely Education, AntiDiab, Insulin, Nutrition, and Sex.
•A novel stacking ensemble was constructed to identify three types of Diabetes Mellitus.•The novel stacking compromises K-NN, Bagging K-NN, and Bagging DT with K-NN as a meta-classifier.•The novel stacking ensemble demonstrated promising results achieving an accuracy of 94.48%.•The Synthetic Minatory Oversampling Technique was utilized to balance the dataset.