AI-Guided Quantitative Plaque Staging Predicts Long-Term Cardiovascular Outcomes in Patients at Risk for ASCVD

The increasing use of AI has extended to reading CCTA results, now referred to as AI-guided quantitative analysis (AI-QCT). This now enables rapid analysis of atherosclerotic plaque burden and characteristics. The ability to incorporate plaque burden, types of plaque (ie, vulnerable plaque volume), high-risk plaque features (ie, remodeling and spotty calcification), and location/severity of stenosis will undoubtedly provide clinicians with more information than coronary calcium scoring (CCS) or simple interpretation of stenosis severity based on the Coronary Artery Disease Reporting and Data System 2.0 (CAD-RADS 2.0).

The current study used stored scans from a decade earlier to evaluate the incremental prognostic value of AI-QCT compared with CCS, clinical risk factors, and CAD RADS 2.0. Not surprising, AI-QCT improved a model with clinical risk factors and CCS at 10-year follow-up. Perhaps surprising, there was little incremental value to knowing maximum stenosis severity, with net reclassification improvement being modest (CAD-RADS 2.0, 0.090 [95% CI, −0.02 to 0.29]; and manual QCT, 0.04 [95% CI, −0.05 to 0.27]).

It is likely that, as we learn more about plaque, location, and remodeling, AI can continue to improve risk prediction. It already affords the only practical way to measure plaque burden over time clinically for CCTA, as manual or semi-automated methods take up to 1 hour to perform. The time for AI-QCT is measured in minutes and, as computing gets better, may continue to improve. Thus, patients can get baseline CCTA and have plaque, stenosis, and high-risk features measured by AI-QCT and then follow-up in 1 to 2 years for a repeat test, evaluating whether atherosclerosis has progressed and if high-risk features have developed or resolved over time.

One must remember that it is possible to optimize the AI programs for any given dataset, so prospective evaluation of this performance is necessary. Despite this, we have entered the age of AI in cardiac imaging, which will lead to less reading time, faster and more accurate risk assessment, and the ability to combine multiple different metrics on CCTA to predict events and track atherosclerosis. The role of the expert reader and how they he/she interacts with these interpretations still needs to be worked out as well, both from a liability and work-flow perspective.

Development of these AI programs are just in the infancy, as there are literally thousands of "-omics" available on any given CT scan, involving location (ie, left main vs distal right coronary), plaque assessment (CCS, plaque volume, and type), stenosis severity and location, and eventually other metrics of risk on the scan, such as pericoronary fat attenuation, epicardial (visceral) fat, bone density, aortic atherosclerosis, and valve calcifications.