Abstract
<title>Abstract</title> <p>Cardiovascular disease (CVD) remains a leading cause of global mortality, demanding timely and precise risk stratification alongside transparent decision support. While machine learning models excel at risk prediction, they lack clinical interpretability. Conversely, Large Language Models (LLMs) provide narrative clinical reasoning but suffer from hallucinations and ungrounded recommendations. To overcome these limitations, we propose CardioTrust, an integrated trustworthy AI decision support framework that unifies Machine Learning risk stratification, prediction guided Hybrid Retrieval Augmented Generation (Hybrid RAG), SHAP explainability, and an automated Trust Validation module. Extensive 5 fold cross validation across multi center cardiovascular benchmarks demonstrates that CardioTrust achieves superior predictive performance (95.84% ± 0.45% accuracy, 0.982 ± 0.003 AUROC, 0.941 ± 0.005 precision, and 0.952 ± 0.004 recall), significantly outperforming baseline models including Logistic Regression (84.21% ± 0.82%), Support Vector Machines (86.50% ± 0.74%), and XGBoost (92.15% ± 0.51%). Furthermore, our prediction guided Hybrid RAG and Trust Validation module ensure that generated recommendations remain factually grounded, verifiable, and aligned with feature importance attributions. CardioTrust provides a robust, transparent, and clinician aligned solution for personalized cardiovascular care.</p>