Machine Learning Models to Identify Discriminatory Factors of Diabetes Subtypes

Abstract

Diabetes Mellitus is a chronic illness that can be defined by high glucose levels in the blood due to its outreach cells in the body. Numerous efforts were happened to detect diabetes however most works were not considered critical factors of diabetes. In this study, proposed machine learning model creates various diabetes subtypes and pinpoints discriminating features. First, we gathered Pima Indians diabetic dataset (PIDD) and Sylhet Diabetes Hospital Datasets (SDHD) from the University of California Irvine (UCI) machine learning repository. Then, only diabetes records were extracted, and created some subtypes from both datasets using k-means clustering and silhouette method. These subtypes data were balanced employing Random Over-Sampling (ROS) and Synthetic Minority Over-sampling Technique (SMOTE). Then, we employed various classifiers such as Decision Tree (DT), K-Nearest Neighbor (KNN), Naïve Bayes (NB), Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting (GB), Adaboost (AdaB), Extreme Gradient Boosting (XGB), Multi-Layer Perceptron (MLP) and Logistic Regression (LR) into the subtypes of PIDD, SDHD, and their balanced datasets. In this case, RF showed the best performance for SMOTE dataset of PIDD subtypes. In addition, LR provided the best performance for ROS dataset of SDHD subtypes. Then, Local Interpretable Model-Agnostic Explanations (LIME) was employed to identify discriminatory factors of these diabetes subtypes by ranking their features.