Highlights
- Machine learning (ML) demonstrated superior ability to predict coronary artery disease (CAD) presence on coronary CT angiography (CCTA) compared to standard 10-year cardiovascular risk scores.
- The ML model leveraged multiple clinical variables including age, sex, cholesterol values, and exercise tolerance testing results to enhance predictive accuracy.
- Prediction of high-risk low-attenuation plaque (LAP) burden by ML did not outperform traditional risk scores, underscoring limitations of clinical factors alone to capture plaque phenotype.
- Findings support integrating ML models in clinical pathways to better triage patients for coronary imaging and optimize resource allocation.
Study Background and Disease Burden
Coronary artery disease (CAD) remains a leading cause of morbidity and mortality worldwide. Accurate identification of patients with high-risk CAD is critical for targeting diagnostic and therapeutic interventions. Coronary CT angiography (CCTA) offers detailed non-invasive visualization of coronary anatomy and plaque characteristics, enabling better risk stratification. However, routine use of CCTA in all patients with suspected CAD may not be feasible or cost-effective due to resource limitations.
Traditional cardiovascular risk calculators, such as the 10-year risk score from pooled cohort equations, estimate long-term atherosclerotic cardiovascular disease risk primarily using demographic and clinical parameters. These scores do not incorporate patient-specific symptom profiles, functional exercise testing results, or detailed laboratory values comprehensively, limiting their accuracy in predicting actual coronary plaque presence and vulnerability.
Machine learning (ML) approaches have the potential to integrate diverse clinical features beyond classical risk scores to better predict CAD status and plaque phenotype on CCTA. Their application could refine patient selection, leading to optimized utilization of imaging resources and timely identification of patients needing aggressive management.
Study Design
This study utilized data from the SCOT-HEART (Scottish Computed Tomography of the HEART) trial, which enrolled patients with suspected angina. From 1769 participants, clinical, demographic, electrocardiography (ECG), and exercise tolerance testing (ETT) data were used to develop and validate ML models using the XGBoost algorithm.
Two separate ML models were constructed:
1. To predict the presence of coronary artery disease on CCTA.
2. To predict an increased burden of low attenuation coronary artery plaque (LAP), a marker of vulnerable plaque.
Model training employed 10-fold cross-validation and grid search for hyperparameter optimization to maximize predictive performance. The models were compared to the 10-year cardiovascular risk score as a benchmark.
Key Findings
The ML model predicting CAD presence on CCTA achieved an area under the receiver operating characteristic curve (AUC) of 0.80 (95% CI 0.74–0.85), significantly outperforming the 10-year cardiovascular risk score alone, which had an AUC of 0.75 (95% CI 0.70–0.81; p=0.004).
Feature importance analysis revealed that the 10-year risk score remained a dominant predictor, supplemented by age, sex, total cholesterol, and abnormal findings on exercise tolerance testing.
In contrast, the ML model designed to predict an elevated LAP burden showed comparable performance to the traditional risk score (AUC 0.75 vs 0.72; p=0.08), suggesting limited added value from clinical features alone in identifying vulnerable plaque.
These results underscore the promise of ML to enhance identification of patients with angiographic CAD, but also highlight challenges in non-invasive detection of plaque vulnerability using standard clinical data.
Expert Commentary
The SCOT-HEART analysis represents a meaningful application of ML to address a critical clinical question: how to predict coronary disease and plaque phenotype non-invasively before imaging.
The improvement in CAD prediction over traditional risk scores could facilitate more personalized diagnostic pathways, prioritizing CCTA for patients most likely to harbor disease. This resource-aware approach is particularly pertinent as healthcare systems grapple with growing demands and imaging backlogs.
However, the inability to improve prediction of LAP burden based on clinical parameters alone may reflect limitations in the pathophysiology captured by symptoms and conventional risk factors. Vulnerable plaques often involve complex microenvironmental and molecular changes beyond what clinical metrics reveal, indicating a need for adjunctive biomarkers or imaging modalities for precise characterization.
The study’s methodological rigor—leveraging robust ML techniques like XGBoost with cross-validation and fair comparator benchmarks—strengthens confidence in the findings. Yet, clinical adoption will require external validation in diverse populations and integration with clinical workflows.
Future directions should explore combining ML with novel biomarkers, genetics, and imaging-derived radiomic features to enhance early detection of high-risk plaques.
Conclusion
Machine learning models trained on comprehensive clinical datasets can significantly improve prediction of coronary artery disease presence on CT angiography beyond standard cardiovascular risk scores. This advancement promises to refine patient selection for CCTA, enabling more focused and efficient cardiovascular evaluation.
Nonetheless, clinical factors alone proved insufficient to better predict high-risk low attenuation plaque burden, highlighting ongoing challenges in non-invasive identification of vulnerable coronary lesions.
Continued research integrating multimodal data with ML will be essential to advance the precision detection of high-risk CAD subsets and ultimately improve individualized patient outcomes.
References
Williams MC, Guimaraes ARM, Jiang M, Kwieciński J, Weir-McCall JR, Adamson PD, Mills NL, Roditi GH, van Beek EJR, Nicol E, Berman DS, Slomka PJ, Dweck MR, Newby DE, Dey D. Machine learning to predict high-risk coronary artery disease on CT in the SCOT-HEART trial. Open Heart. 2025 Sep 1;12(2):e003162. doi: 10.1136/openhrt-2025-003162. PMID: 40889953; PMCID: PMC12406813.
Hoffmann U, Truong QA, Nagurney JT, et al. 64-slice computed tomography coronary angiography in patients with suspected coronary artery disease. JAMA. 2009;302(4):370-378.
Nørgaard BL, Leipsic J, Gaur S, et al. Diagnostic performance of non-invasive fractional flow reserve derived from coronary CT angiography in suspected coronary artery disease: The NXT trial (Analysis of Coronary Blood Flow Using CT Angiography: Next Steps). J Am Coll Cardiol. 2014;63(12):1145-1155.
Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med. 2019;25(1):24-29.
