A comparative study of quantum-inspired PSO and EA with their binary variants for heart disease classification

Feature selection and hyperparameter optimization are widely used in heart disease prediction, but comparative findings are often affected by differences in preprocessing, search space design, and evaluation budget. This study compared quantum-inspired particle swarm optimization and evolutionary algorithms with their binary variants, namely QI-PSO, QI-EA, QI-BPSO, and QI-BEA, under the same evaluation protocol for joint feature selection and discrete hyperparameter tuning on the UCI Cleveland Heart Disease dataset. The feature mask and classifier hyperparameters were optimized simultaneously in a unified search space using the same preprocessing scheme, inner-validation protocol, and evaluation budget. Decision Tree (DT), Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN) classifiers were evaluated using a fixed train-test split. Held-out test performance was reported with 95% confidence intervals. In addition, a no-feature selection baseline was evaluated for each classifier using the final hyperparameter configuration of the best selected feature QI model for the corresponding classifier while retaining all 13 input variables. Precision-recall curves, threshold sensitivity profiles, outer-trial CV Accuracy, runtime, selected feature counts, convergence behavior, and cost-performance distributions were also analyzed. Statistical comparisons of outer-trial distributions were performed using Kruskal-Wallis tests followed by Holm-adjusted Mann-Whitney U post-hoc comparisons. The held-out test results showed classifier-dependent differences among the QI variants. For DT and K-NN, the highest selected-feature test Accuracy was obtained with QI-PSO, whereas for SVM it was obtained with QI-BPSO. In the K-NN setting, QI-BEA, defined in this study as the modified binary implementation of the QI-EA update strategy, produced higher selected-feature Accuracy, F1-score, and MCC than QI-EA. For DT, QI-BPSO yielded the highest ROC-AUC, while for LR all four QI methods produced identical held-out test metrics. Compared with the corresponding no-feature selection baselines, the selected feature QI models showed a clear held-out advantage for K-NN. In DT, the no-feature selection baselines matched the best selected feature accuracy, whereas QI-BPSO yielded the highest DT ROC-AUC. In LR and SVM, the no-feature selection baselines produced stronger held-out results than the selected feature configurations. Precision-recall and threshold sensitivity analyses further showed classifier-dependent overlap and separation patterns beyond single point estimates. At the outer-trial level, QI-BPSO occupied the lowest runtime range, while QI-BEA yielded the highest median CV accuracy for DT, K-NN, and LR. These findings show that the relative behavior of QI variants depends on the classifier family and on whether emphasis is placed on held-out test performance, operating-threshold behavior, inner-loop validation level, or computational cost.