<![CDATA[Heart disease remains one of the leading causes of mortality worldwide, making early detection crucial to prevent severe complications. The limitations of conventional diagnostic approaches have encouraged the adoption of machine learning techniques to enable faster and more accurate predictions. Support Vector Machine (SVM) is widely recognized as an effective method for medical classification tasks; however, its performance is highly dependent on the choice of kernel function. This study evaluates three single-kernel SVM models (Linear, RBF, and Polynomial) and two hybrid kernel configurations, namely Linear–RBF and Linear–Polynomial, using the UCI Heart Disease Statlog dataset, which consists of 270 samples and 13 predictive features. In the hybrid approach, the probabilistic outputs of the individual base kernels are combined through an aggregation strategy to construct a decision function capable of capturing both linear and nonlinear patterns simultaneously. To ensure performance stability on the relatively small dataset, model evaluation was conducted using Stratified K-Fold Cross Validation, ensuring that the reported results do not rely on a single data split. Experimental results indicate that the SVM-Polynomial model achieved the highest ROC-AUC value of 0.9420; however, it did not outperform other models in terms of accuracy, precision, or F1-score. The hybrid approach demonstrated more consistent overall performance, with the Linear–RBF combination emerging as the best-performing model, achieving an accuracy of 0.8889, macro precision of 0.8896, and macro F1-score of 0.8886. These findings suggest that integrating linear and nonlinear kernel characteristics produces a more balanced decision function compared to single-kernel models. In contrast, the Linear–Polynomial combination did not yield significant performance improvements. The main contribution of this study lies in presenting a structured comparative analysis of kernel combination strategies in SVM for heart disease classification, which may support the development of more adaptive and stable clinical prediction systems.]]>
Copyrights © 2026
