Machine Learning Model for Atherosclerosis Evaluation and Cardiovascular Risk Prediction Based on Coronary CT Angiography-Analysis From the CREATION Registry.
Authors
Affiliations (2)
Affiliations (2)
- Department of Cardiology, Fuwai Hospital, National Clinical Research Center for Cardiovascular Diseases, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China. (Y.S., N.X., C.C., Y.Z., L.G., Z.G., J.C., L.S., J.Y.).
- Department of Radiology, Fuwai Hospital, National Clinical Research Center for Cardiovascular Diseases, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China. (J.Z., S.J., B.L., H.Z.-H.).
Abstract
Current ASCVD risk prediction tools based on traditional risk factors and the coronary artery calcium score have limitations. The CREATION study includes suspected coronary artery disease patients who underwent coronary computed tomography angiography (CCTA) at Fuwai Hospital between 2016 and 2019. The primary outcome was major adverse cardiac events defined as a composite end point of all-cause death, acute myocardial infarction, coronary revascularization, or stroke. Six machine learning survival models were used to create an ASCVD prediction model. Overall, 8431 participants with analyzable CCTA data were included with a median follow-up of 3.68 years, and 319 major adverse cardiac events (3.8%) occurred (mean age: 54.73±10.21 years, 48.2% were male, 50.9% with symptomatic chest pain). Among 6 machine learning models trained with 48 CCTA parameters, XGBoost showed the best performance and was selected for model development. In the training cohort (n=5901, 70%), the XGBoost model significantly outperformed the clinical risk factors and coronary artery calcium score model (area under the curve, 0.903 versus 0.830; <i>P</i><0.001). Testing cohort showed similar performance (area under the curve, 0.899 versus 0.753; <i>P</i><0.001). The CCTA model demonstrates consistent predictive performance across gender (female or male), onset-age (early onset or late-onset), and symptom (asymptomatic or symptomatic) subgroup analysis. The final CCTA model included diameter stenosis, lipid plaque burden and volume, total plaque volume, high-risk plaque, and vessel volume as the most important features. Lipid plaque burden was most strongly associated with major adverse cardiac event (adjusted hazard ratio per 5% increase: 2.524 [95% CI, 2.157-2.996]; <i>P</i><0.001). The incremental value of machine learning CCTA features was consistent across different time points throughout the 1- to 5-year follow-up period. The findings remained unchanged when restricted to a secondary composite end point (death, myocardial infarction, or stroke). The machine learning model incorporating CCTA plaque quantification, characterization, and stenosis assessment significantly enhanced the predictive capacity for major adverse cardiac events. It provides direct visualization of coronary atherosclerosis and outperforms the traditional risk factors and the coronary artery calcium score model recommended in clinical practice.