426-P: Predicting Coronary Stenoses Using Machine Learning to Reduce Unnecessary Angiographies
Journal
Diabetes
ISSN
0012-1797
Type
journal article
Date Issued
2025-06-20
Author(s)
Leiherer, Andreas
;
Schnetzer, Laura
;
Mink, Sylvia
;
Muendlein, Axel
;
;
Thomas Plattner
;
Vonbank, Alexander
;
Mader, Arthur
;
Larcher, Barbara
;
Saely, Christoph
;
Peter Fraunberger
;
Drexel, Heinz
Abstract
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.
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.
Language
English
HSG Classification
contribution to scientific community
Refereed
Yes
Volume
74
Number
Supplement_1
Start page
426-P
Official URL