Prediction of Type 2 Diabetes Mellitus Using Machine Learning in Dupti Referral Hospital

Abstract Diabetes is a chronic, metabolic disease characterized by elevated levels of blood glucose, which leads over time to damage to the heart, vascular, eyes, kidneys and nerves. Numbers of study have proposed machine learning algorithm for diabetes disease detection and prediction. However, there has been used very small number of dataset and the feature they used also has been little discussion about the relationship between diabetes symptoms. Generally, diabetes is classified into three different types which include type 1 diabetes, type 2 diabetes and gestational diabetes among them this study focused on Type 2 diabetes disease prediction. The primary goal of this study is to give a concise overview of a real-world dataset & Dupti referral hospital of diabetes illness symptoms. The data for develop our type 2 diabetes mellitus (T2DM) predictive model is collected from dupti referral hospital which is found in Afar region that contain 6000 recorded patient laboratory result. From those row data we applied different preprocessing tasks handling missing value; remove redundancy, data transformation, and feature selection for improve the model performance. To predict the T2DM disease we compare seven machine learning algorithms by taking the selected feature and all features of our dataset and the predicted model achieved better in all features of our data. The machine learning algorithm that used for build our predictive model were Random Forest, KNN, SVM, CAT Boost, XgBoost, Gradient Boosting, and AdaBoost algorithms achieved 98.87%, 94.89%, 98.77%, 98.46%, 97.85%, 98.26%, and 98.06% accuracy respectively using selected feature and 98.16%, 93.77%, 97.85%, 98.16%, 98.16%, 98.16%, and 97.44% accuracy respectively using all feature. Among those model metric Random Forest achieved better accuracy for both selected feature and all feature that is 98.87% and 98.16% respectively. In addition to accuracy all those models are evaluated by using other evaluation metric those are confusion matrix, Precision Recall with Area under the Curve (PR-AUC), and Receiver Operating Curve with Area under the Curve (ROC-AUC) metrics.