PREDICTING DIABETES MELLITUS USING CATBOOST CLASSIFIER AND SHAPLEY ADDITIVE EXPLANATION (SHAP) APPROACH

Abstract

Diabetes mellitus as a metabolic disease characterized by hyperglycemia can be dangerous if it cannot be handled properly. Early detection of existing symptoms can reduce the impact of delays in treatment. This study aims to carry out early-detection patients with diabetes mellitus using a machine learning approach through data from MIT’s GOSSIS (Global Open Source Severity of Illness Score). By using Shapley Additive Explanation (SHAP) which enables prioritization of feature that determine compound classification, this study shows that the CatBoost classifier has 14 features that significantly can be used for classification with feature ‘d1_glucose_max’ or the highest glucose concentration of the patient in their serum or plasma during the first 24 hours of their unit stay has the highest impact to classify diabetes mellitus patients, then followed by age and glucose APACHE. The selected features are then classified and get the validation AUC score of 86.86%.