Computed tomography coronary angiography features integration with gradient boosted decision tree for predicting cardiovascular event
A Choo, C K Wong, F C C Tam, W C Wong, S C W Cheung, C W S Chan, W Wong, H F TseAbstract
Background
Computed tomography coronary angiogram (CTCA) is commonly performed to diagnose coronary artery disease. Machine learning has enabled automated computation of quantitative plaque volume, CT-fractional flow reserve (CT-FFR), and perivascular fat attenuation index (FAI). These parameters have been individually used to stratify patient prognosis. However, limited work has been done to integrate all these novel parameters using machine learning methods.
Purpose
To develop a predictive model that integrates advanced biomarkers from CTCA using machine learning, enhancing prognosis in coronary artery disease by combining automated quantification of plaque volumes, CT fractional flow reserve and the perivascular fat attenuation index.
Methods
A retrospective observational single-center study was conducted involving patients who underwent CTCA from January 2014 to October 2024. Propensity score matching was used to subset and divide the cohort into 2 groups, including those with cardiovascular event in the follow up period (CV Group) and control (Control Group). Cardiovascular events analyzed included incident myocardial infarction, heart failure, ventricular tachycardia and fibrillation. Data for age, sex, quantitative plaque volumes, including total plaque volume (TVL), calcium volume, fibrous volume, lipid volume, and fibro-lipid volume, CT-FFR, and FAI were extracted. XGBoost, a machine learning algorithm based on gradient boosting that builds an ensemble of decision trees, was trained to predict cardiovascular events. Primary outcome was accuracy of predicting cardiovascular event, represented by the area under receiver operating characteristic curve (AUROC). Hazard ratios were estimated using cut points derived from the Youden index.
Results
In total, 7940 patients (47.50% male; mean age 63.90 ±13.20) were included in the study. After a median follow-up period of 36 months, 366 patients (4.61%) developed cardiovascular event. After propensity score matching, 310 and 6900 patients were allocated to the CV Group and Control Group. When used individually, TVL, CT-FFR, and FAI predicted cardiovascular event with AUROC 0.61, 0.61, and 0.55. Using Youden based thresholds, including TVL 67.73 cm3, CT-FFR 0.80, and FAI -79.5, patients were stratified into high-risk and low-risk group. The HR for cardiovascular event in the high-risk group were 0.96 for TVL, 1.27 for CT-FFR, and 1.09 for FAI. An XGBoost model integrating demographics and all CTCA parameters achieved AUROC 0.67 and a high risk hazard ratio of 2.74.Figure(1-2)
Conclusion
Integrating CT derived parameters with machine learning modestly improves risk stratification after CTCA. In this large single center cohort, a combined model outperformed any individual marker and identified a subgroup with more than double the risk of cardiovascular events.Figure1.ROC CurveFor image description, please refer to the figure legend and surrounding text.Figure 2.KM CurveFor image description, please refer to the figure legend and surrounding text.