Using a CT Angiography Artificial Intelligence Carotid Plaque Morphology Model to Predict Stroke

The gold standard for determining carotid artery stenosis is to calculate stenosis using the North American Symptomatic Carotid Endarterectomy Trial criteria and ultimately plan for medical vs surgical management based on percent stenosis and symptomatic status. Few studies have assessed plaque morphology as an additive tool for stroke prediction. Our study uses an artificial intelligence software in conjunction with a patient's computed tomography (CT) scan of the neck to create a three-dimensional (3D) model of the carotid artery and assess plaque morphology including calcification, intraplaque hemorrhage, matrix, and perivascular adipose tissue. Our goal was to create a predictive model inclusive of plaque morphology. This is a retrospective review of a single tertiary institution from 2010 to 2021. Patients with a CT angiography head/neck and a diagnosis of carotid artery stenosis were included in our analysis. Each CT scan was run through a third-party software to create a 3D image for plaque visibility and analysis. We used a stepwise backward regression to select variables for inclusion in our prediction models. Model discrimination was assessed with receiver operating characteristic curves (AUC) and the discrimination slope. Additionally, calibration was performed and the model with the least Akaike information criterion (AIC) was selected. Our primary outcome was all cause mortality and stroke. Our sample included 366 patients over the 11-year study. We created three models to predict mortality/stroke: model A using only clinical variables, model B using only plaque morphology software variables, and model C using both clinical and software variables. Model A was created using age, sex, peripheral arterial disease, hyperlipidemia, body mass index, chronic obstructive pulmonary disease, and history of transient ischemic attack or stroke and was found to have an AUC of 0.737and AIC of 285.4. Model B was created using perivascular adipose tissue volume, cross sectional lumen area, calcified volume, and target lesion length and was found to have an AUC of 0.644 and AIC of 304.8. Finally, model C combined both clinical and software variables and included age, sex, matrix volume, history of transient ischemic attack/stroke, body mass index, perivascular adipose tissue, lipid rich necrotic core, chronic obstructive pulmonary disease, and hyperlipidemia and was found to have an AUC of 0.759 and the least AIC of 277.6 (Figs 1 and 2). Our models demonstrate that combining both clinical factors and plaque morphology creates the best model to predict a patient's risk for all-cause mortality and stroke from carotid artery stenosis. Prospective studies are needed to validate our findings.Fig 2Model C: clinical and software variables calibration plot.View Large Image Figure ViewerDownload Hi-res image Download (PPT)