Coronary CT Angiography-Based Prediction Model for Hemodynamically Significant Coronary Stenosis Integrating Morphological and Plaque Characteristics.
Authors
Affiliations (9)
Affiliations (9)
- Affiliated Hospital of Jining Medical University, Jining, China.
- National Heart Centre Singapore, 5 Hospital Drive, Singapore, 169609, Singapore.
- Duke-NUS Medical School, Singapore, Singapore.
- Department of Cardiology, National University Heart Centre, Singapore, Singapore.
- Yong Loo Lin School of Medicine, National University of Singapore, Singapore, Singapore.
- Department of Diagnostic Imaging, National University Hospital, Singapore, Singapore.
- National Heart Centre Singapore, 5 Hospital Drive, Singapore, 169609, Singapore. [email protected].
- Duke-NUS Medical School, Singapore, Singapore. [email protected].
- Department of Biomedical Engineering, National University of Singapore, Singapore, Singapore. [email protected].
Abstract
A rising trend involves the application of machine learning to coronary computed tomography angiography (CCTA) for predicting physiological significance of coronary lesions. This study aimed to formulate a clinical model using CCTA-derived features and the least absolute shrinkage and selection operator (LASSO) regression method to predict hemodynamically significant coronary stenosis. The study population comprised individuals prospectively recruited from two tertiary medical centres with suspected or known coronary artery disease. The model was developed using patients from Centre 1 and independently validated in patients from Centre 2. All participants underwent CCTA followed by invasive coronary angiography with fractional flow reserve (FFR) measurements. The LASSO model incorporated coronary morphological and plaque features derived semi-automatically from CCTA, along with clinical and demographic variables, as input predictors. Model performance was assessed against the reference standard of invasive FFR ≤ 0.80. The analysis included 210 diseased vessels from 133 patients-141 vessels from 84 patients in Centre 1 (development cohort) and 69 vessels from 49 patients in Centre 2 (validation cohort). Using the LASSO algorithm, nine predictive variables were selected from an initial set of 33 candidate features: heart rate, minimal lumen diameter (MLD), area stenosis, diameter stenosis ≥ 50%, lesion length/MLD<sup>4</sup>, necrotic core plaque volume, lesion entrance angle, remodeling index, and vessel-specific calcium score. In the validation cohort, the LASSO model demonstrated strong diagnostic performance at both vessel and patient levels. Per-vessel accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were 93%, 86%, 98%, 96%, and 91%, respectively. Comparable performance was observed at the per-patient level, with corresponding values of 92%, 89%, 95%, 96%, and 88%, respectively. The area under the receiver operating characteristic curve (AUC) for the LASSO model at the vessel level was 0.950, significantly higher than that of CT-derived FFR (AUC = 0.857, P = 0.046) and diameter stenosis ≥ 50% (AUC = 0.641, P < 0.0001). A prediction model integrating coronary and plaque parameters outperformed existing CCTA-based methods in identifying hemodynamically significant coronary stenosis, offering improved accuracy for ischemia detection.