A Comparative Evaluation of Machine Learning Algorithms for Diabetes Risk Prediction

Abstract

Diabetes mellitus is a chronic metabolic disorder whose global prevalence continues to rise, creating an urgent need for scalable, low-cost tools for early risk identification. This study evaluates the effectiveness of five machine learning algorithms (Random Forest, XGBoost, Support Vector Machine [SVM], CatBoost, and TabNet) for predicting diabetes risk from routinely available clinical and lifestyle variables. Using the Pima Indians Diabetes dataset, a preprocessing pipeline was applied that included median imputation of physiologically implausible zero values, standardized (Z-score) feature scaling, Boruta-based feature selection, and class-imbalance handling through class-weight adjustment and the Synthetic Minority Oversampling Technique (SMOTE). Models were trained on an 80/20 train-test split and assessed using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (ROC-AUC). Random Forest achieved the strongest overall performance (accuracy 0.753; F1-score 0.689; ROC-AUC 0.810), followed by CatBoost, SVM, and XGBoost, whereas TabNet performed worst with very low recall for the diabetic class. The best-performing model (Random Forest) was deployed in a lightweight Flask web application that returns a probability-based diabetes risk assessment, categorising each prediction as low, moderate, or high risk together with a tailored recommendation. The findings confirm that ensemble tree-based methods, particularly Random Forest, provide a reliable, interpretable, and deployable basis for diabetes risk screening, especially in resource-constrained settings. Key limitations include dataset homogeneity, residual class imbalance, and limited feature coverage.