Leakage-Safe Benchmarking of Tempered Fractional Optimization and Swarm-Driven Feature Construction for Heart Disease Prediction on Structured Clinical Data

Abstract Prediction of heart disease from structured clinical data requires both predictive performance and methodological rigor. This paper presents a leakage-safe benchmark of tempered fractional optimization and swarm-driven feature construction for heart disease prediction. The proposed framework combines a tempered fractional gradient-based logistic learner with particle swarm optimization (PSO)-driven nonlinear feature construction and is evaluated against tuned strong baselines, including elastic-net logistic regression, Extra Trees, histogram-based gradient boosting, XGBoost, LightGBM, and a stacking ensemble. The experimental protocol nests preprocessing, hyperparameter tuning, and PSO construction strictly within training data to avoid leakage. Across 15 repeated stratified hold-out splits, the proposed TFGD_PSO consistently improves over plain TFGD. However, the stacking ensemble achieves the strongest overall performance (mean ROC-AUC 0.9318 ± 0.0146, mean F1-score 0.8917±0.0210, mean MCC 0.7564±0.0463), while Extra Trees reaches the highest mean PR-AUC. TFGD_PSO remains competitive and computationally attractive. The study provides a rigorous benchmark showing that PSO-enhanced tempered fractional learning improves over plain tempered fractional optimization, while tuned tree-based ensembles remain the strongest predictors for this dataset.