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

Area of Science:

• Computational intelligence and machine learning applied to biomedical informatics.
• Development and evaluation of optimization algorithms for predictive modeling.

Background:

• Feature selection and hyperparameter optimization are crucial for accurate heart disease prediction.
• Comparative studies are often limited by variations in preprocessing, search space, and evaluation budgets.
• Need for standardized evaluation of advanced optimization techniques in this domain.

Purpose of the Study:

• To compare quantum-inspired particle swarm optimization (QI-PSO) and evolutionary algorithms (QI-EA) with their binary variants (QI-BPSO, QI-BEA).
• To evaluate their effectiveness in joint feature selection and hyperparameter tuning for heart disease prediction.
• To analyze performance under a unified evaluation protocol across different classifiers.

Main Methods:

• Employed QI-PSO, QI-EA, QI-BPSO, and QI-BEA for simultaneous optimization of feature masks and classifier hyperparameters.
• Utilized the UCI Cleveland Heart Disease dataset with consistent preprocessing, inner-validation, and evaluation budget.
• Evaluated Decision Tree (DT), Logistic Regression (LR), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN) classifiers using a fixed train-test split and analyzed various performance metrics.

Main Results:

• Classifier-dependent performance differences were observed among QI variants.
• QI-PSO excelled for DT and K-NN accuracy, while QI-BPSO was best for SVM.
• QI-BEA improved K-NN performance over QI-EA; QI-BPSO achieved the highest DT ROC-AUC. Runtime and CV accuracy varied, with QI-BPSO being fastest and QI-BEA yielding high median CV accuracy for DT, K-NN, and LR.

Conclusions:

• The relative effectiveness of quantum-inspired variants is contingent on the specific classifier and the performance metric prioritized (e.g., test accuracy, runtime, validation accuracy).
• Joint optimization in a unified search space provides a robust comparison framework.
• No-feature selection baselines offered competitive or superior results for LR and SVM, indicating potential limitations of selected-feature models in certain contexts.