<![CDATA[Heart disease classification using machine learning requires relevant features and predictive models capable of consistently generalizing clinical patterns. Previous studies on the Heart Failure Prediction dataset demonstrated that K-Nearest Neighbor (KNN) optimized with Particle Swarm Optimization (PSO) achieved an accuracy of 89.09% and an Area Under the Curve (AUC) of 0.935. However, the use of a fixed inertia weight and reliance on a single learner may limit the balance between exploration and exploitation, thereby reducing model robustness. This study proposes a feature selection approach based on Adaptive Particle Swarm Optimization (APSO), in which the inertia weight is gradually decreased from 0.90 to approximately 0.42 over 30 iterations. The optimal feature subset is subsequently utilized in a soft voting ensemble learning model. The dataset consists of 918 records, 11 predictive features, and one target class (HeartDisease). Experimental results indicate that the proposed APSO-based ensemble model achieved an accuracy of 89.71%, an F1-score of 0.8986, and an AUC of 0.9466. The confusion matrix yielded 90 true negatives, 12 false positives, 9 false negatives, and 93 true positives on 204 testing instances. Compared with the baseline KNN-PSO model, the proposed method improved classification accuracy by 0.62 percentage points and increased the AUC by 0.0116, while maintaining a disease-class recall of 91.18%. These findings demonstrate that combining adaptive search dynamics with heterogeneous ensemble learning enhances the discriminative capability of heart disease classification, although further validation using identical data partitioning strategies and external datasets is still required]]>
Copyrights © 2026
