Diabetes Mellitus Detection System Using Machine Learning with the Support Vector Machine (SVM) Method

Abstract

Background: Diabetes Mellitus (DM) remains a major non-communicable disease with a continuously rising prevalence, and delayed diagnosis increases the risk of severe complications such as kidney failure, cardiovascular disease, and neuropathy. Objective: This study aims to design and build a web-based Diabetes Mellitus detection system using the Support Vector Machine (SVM) algorithm and to measure the classification performance of the resulting model. Methods: The dataset consisted of 170 medical records from Rumah Sakit Cut Meutia collected over the 2023-2025 period, comprising eight predictor variables (gender, blood pressure, glucose level, insulin, skin thickness, Body Mass Index, diabetes family-history score, and age). The data were preprocessed through label encoding, median-based missing-value imputation, Interquartile Range (IQR) outlier clipping, and Min-Max normalization, then split into 80% training data (132 records) and 20% testing data (34 records). The SVM model, comparing Linear and Radial Basis Function (RBF) kernels, was optimized using GridSearchCV with Stratified 5-Fold Cross Validation. Results: The optimal model used an RBF kernel with C = 1 and gamma = 0.01, achieving a cross-validation ROC-AUC of 96.43%. On the test set, the model achieved an accuracy of 88.24%, precision of 94.12%, recall of 84.21%, and an F1-score of 88.89%, with a test-set AUC of 96.1%. Usability testing using the System Usability Scale (SUS) on 10 respondents produced an average score of 71.75, placed in the "Good" category. Conclusion: The SVM-based web system was able to classify Diabetes Mellitus status accurately and was assessed by users as easy to operate, indicating its feasibility as a decision-support tool for the early detection of Diabetes Mellitus.