EVALUATION OF MACHINE LEARNING CLASSIFIERS FOR EARLYSTAGE TYPE 2 DIABETES RISK PREDICTION

Abstract

Early diagnosis of chronic metabolic problems is crucial to improve preventive healthcare and reduce the long-term effects of Type 2 Diabetes Mellitus (T2DM). Since diabetes is mostly undetected in its early stages, accurate prediction models can help for timely therapeutic intervention. The study assesses the performance of three commonly used machine learn-ing algorithms, Logistic Regression, Support Vector Machine (SVM) and Random Forest, for early prediction of diabetes risk utilizing demographic, clinical, lifestyle and biochemical data. The study used a standardized healthcare dataset which consisted of patient demographics, body mass index (BMI), blood pressure, fasting plasma glucose (FPG), lipid profile, kidney function markers, lifestyle factors and family history. The data pre-processing involved missing value treatment, duplicate record removal, feature scaling, and balancing the training data using the Synthetic Minority Oversampling Technique (SMOTE). Models were evaluated based on Accuracy, Precision, Recall, F1-score and Receiver Operating Characteristic Area Under Curve (ROC-AUC). In general, the Support Vector Machine classifier was the best one with about 89% accuracy and a ROC-AUC of 0.95. Logistic Regression and Random Forest also had a good predictive power. Feature importance analysis showed that fasting plasma glucose (FPG) and age were the most important predictors of diabetes, followed by HDL, LDL and systolic blood pressure. These results indicate that machine learning approaches can be reliable and efficient tools to assist the decisions related to early screening and risk assessment of diabetes. The proposed framework can be used in the digital healthcare system to help doctors in detecting high-risk individuals and advocating timely preventative actions.