Introduction and Objective: Coronary angiography is the gold standard for diagnosing coronary artery stenoses, but is invasive and confers potential risks. This study aimed to develop a Machine Learning (ML) model to predict significant stenoses while minimizing false negatives, ensuring accurate risk stratification and better patient selection.Methods: Data from 2,310 patients undergoing coronary angiography were analyzed, with outcomes classified as no stenoses (X0), non-significant stenoses (X1), or significant stenoses (X2).Results: XGBoost, optimized through grid search and 5-fold cross-validation, emerged as the top-performing ML algorithm. Of 114 clinical and laboratory variables, fibrinogen, HbA1c, BMI, waist-hip ratio, TyG index, FGF23, ceramides, and vitamin D – all associated with insulin resistance and diabetes – were identified as key contributors to the ML model (figure). Overall, the model achieved 62.9% accuracy (95% CI: 57.8-67.9). Sensitivity, precision, and F1 score for X2 were 94.7%, 62.7%, and 74.5%. For X0, sensitivity was 37.3%, with precision and F1 scores of 67.6% and 48.1%.Conclusion: This ML-based approach has the potential to reduce unnecessary angiographies and optimize patient selection in clinical practice. Highlighting the relevance of diabetes-linked variables, the study also underscores the potential of metabolic profiling in coronary risk stratification.
A. Leiherer: None. L. Schnetzer: None. S. Mink: None. A. Muendlein: None. B. Bermeitinger: None. T. Plattner: None. A. Vonbank: None. A. Mader: None. B. Larcher: None. C.H. Saely: None. P. Fraunberger: None. H. Drexel: None.
Source link
Leave a Reply