A Hybrid Recursive Feature Elimination Principal Component Analysis Framework With Optimized Support Vector Machine For Accurate Diabetes Prediction

Abstract

Timely diagnosis of diabetes supports earlier clinical action, yet standard machine-learning pipelines often struggle with overlapping features, inter-correlated clinical measurements, and models whose decisions are hard to explain. In this paper, we present an explainable hybrid framework based on Recursive feature elimination – Principal Component Analysis – Radial Basis Function Support Vector Machine (RFE-PCA-RBF-SVM). The study exploits Pima Indians Diabetes Dataset (PIDD), which consists of 768 cases and eight clinical attributes. Our framework comprises handling of invalid data based on median imputation, Min-Max normalization of the dataset, RFE feature selection, PCA dimensionality reduction, and hyperparameter optimization of RBF-SVM. The experiment was designed to use 80:20 stratified train-test scheme as well as 10-fold stratified cross-validation, ablation analysis, and statistical validation methods. It was found out that hybrid framework performed better than the base model in terms of measured characteristics and achieved accuracy of 71.08% (standard deviation = 2.71%). Considering that Histogram Gradient Boosting surpassed the other methods analyzed in terms of the predictive performance of the model, the hybrid framework proposed in this study managed to remain highly effective offering a well-structured process for feature optimization and nonlinear classification. In this context, the explainability based on the SHAP methodology determines the importance of Glucose, BMI, Age, and Diabetes Pedigree Function for the model prediction results. Consequently, the proposed method provides a neat, statistically verified, and interpretable technique for diabetes prediction, demonstrating the significance of the combination of predictive evaluation with component-wise ablation and explainability analysis.