While machine learning can sharpen coronary artery disease (CAD) prediction using standard clinical data, it falls short in detecting plaques most likely to cause future cardiac events in patients receiving coronary CT angiography (CCTA), according to the results of a new study.

For patients with stable symptoms undergoing CCTA, detecting any CAD and identifying high-risk plaque phenotypes are both clinically important. The presence of coronary disease supports starting medical therapy, while recognition of high-risk plaque may indicate a need for more intensive treatment.

This secondary analysis of the original SCOT-HEART trial, led by researchers from the University of Edinburgh, evaluated 1,769 patients aged 18-75 years with stable chest pain who were randomised to undergo CCTA.

Machine learning prediction models

From this, two machine learning models were developed using the XGBoost algorithm: one to predict the presence of CAD and another to identify increased low-attenuation plaque (LAP) burden – a marker of high-risk atherosclerosis – on CCTA. The dataset was split into training (80%) and testing (20%) cohorts.

The CAD prediction model achieved an area under the curve (AUC) of 0.80, significantly outperforming both the 10-year cardiovascular risk score (AUC 0.75, p=0.004) and the European Society of Cardiology (ESC) pre-test probability score (AUC 0.73, p<0.001). Important predictors included age, sex, total cholesterol and abnormal exercise tolerance test results.

By contrast, the LAP model did not significantly outperform the 10-year risk score (AUC 0.75 vs 0.72, p=0.08), though it slightly exceeded the ESC score (AUC 0.69, p=0.005). Key variables for LAP prediction included body mass index and high-density lipoprotein cholesterol.

The researchers acknowledged that a lack of external validation, reliance on routine clinical data and applicability restricted to patients with stable chest pain undergoing CCTA limited the study. They also highlighted that the models may not generalise to asymptomatic or acute care populations, and scanner variability or missing data could have influenced the results.

Clinical potential in CCTA

Machine learning could be used to better select patients for CCTA, the researchers concluded, but they cautioned that clinical data alone were not sufficient to predict high-risk plaque burden. This highlights the ongoing need for imaging-based biomarkers in cardiovascular risk assessment.

They added that incorporating advanced imaging metrics, such as radiomic features, may enhance the ability of future models to detect high-risk plaque and further refine risk stratification.

Overall, the findings highlight the potential of machine learning to refine diagnostic pathways for CAD while emphasising the need for further development of this machine learning approach to CCTA before clinical implementation.

Reference
Williams MC et al. Machine learning to predict high-risk coronary artery disease on CT in the SCOT-HEART trial. Open Heart 2025;12:e003162.