<![CDATA[Opaque model decisions remain a practical barrier to the use of machine learning in heart disease screening. This study presents HeartXplain, a web-based AI Health Agent designed as a Clinical Decision Support System for early screening. The system combines XGBoost, Particle Swarm Optimization (PSO), SHAP, LIME, and interpretative narratives generated by a Large Language Model (LLM). The model was trained on 69,105 records represented by 19 clinical and engineered features. PSO optimized a recall-oriented objective, followed by decision-threshold adjustment. At a threshold of 0.4997, the model achieved 71.01% accuracy, 80.08% recall, 67.39% precision, a 73.19% F1-score, 79.29% AUC-ROC, and 78.05% PR-AUC; false negatives decreased from 1,367 to 1,360. SHAP and LIME exposed global and patient-level feature contributions, while the LLM translated those outputs into a more readable narrative. MAPIE reached 95.01% empirical coverage and 83.02% accuracy on singleton sets. Fairlearn indicated minor sex-based differences but wider gaps across age groups. HeartXplain therefore provides a transparent supporting layer for early screening, not a substitute for clinical examination, diagnosis, or professional judgment. External validation on real clinical data remains necessary.]]>
Copyrights © 2026
