Artificial intelligence-enabled comprehensive coronary phenotyping in patients with suspected CAD

Abstract

Introduction: The capabilities of artificial intelligence (AI) are rapidly progressing and the research community is getting increasingly interested in its possibilities. AI algorithms are able to work continuously and at high speed, reducing human workload and saving time that physicians can spend on more complex data or rarer cases. However, many clinical AI applications are currently only used in a research setting and lack proper testing and validation. Objectives: This study aimed to determine the accuracy and performance of a novel AI-based software tool for CCTA analysis compared to conventional expert evaluation. Methods: We evaluated 100 CCTA exams from a cohort of symptomatic patients with mild-to-moderately abnormal non-invasive ischemia test. Stenosis severity assessed by AI-based analysis (automatic evaluation, AEv) was compared with a level III expert CCTA interpretation (manual evaluation, MEv). AI-based analysis reported exact % stenosis and obstructive CAD was considered if maximal stenosis was ≥50%. Plaque phenotype was also estimated using AI algorithms. Results: The study cohort was as follows: 52% male, mean age 68±10 years. The prevalence of hypertension, dyslipidemia and diabetes was 77%, 81% and 23%, respectively, and 10-year cardiovascular risk was 19±10% as predicted by Framingham risk score. Typical angina was present in 33%, of which 67% had a Canadian Cardiovascular Society angina grade ≥2. Overall prevalence of obstructive CAD determined by MEv and AEv was 25% and 21%, respectively, with a significant association between both assessments (p<0.001). When compared to MEv as reference, AEv method performed with a sensitivity, specificity, positive and negative predictive values of 0.56, 0.91, 0.58 and 0.86, respectively. Area under the curve was 0.871 (p<0.001) demonstrating high accuracy. AEv atherosclerosis quantification revealed significant differences between patients with and without obstructive CAD according to MEv: Median total plaque volume (569 vs 115 mm3, p<0.001), calcified plaque volume (297 vs 19 mm3, p<0.001), non-calcified plaque volume (235 vs 71 mm, p<0.001), low-density non-calcified plaque volume (2.8 vs 1.0 mmi, p=0.023) and percent atheroma volume (16.1 vs 3.8 mm3, p<0.001). Conclusion: In patients with suspected CAD and mild-to-moderately abnormal ischemia tests, a diagnostic strategy using AEv as a gatekeeper is effective, providing a quantitative stenosis evaluation with similar diagnostic performance for obstructive CAD when compared to MEv. AI-enabled approach additionally allows a fully automated quantification of coronary plaque volumes and composition, which would further enhance risk stratification.