Novel Application of Artificial Intelligence Algorithms to Develop a Predictive Model for Major Adverse Neurologic Events in Patients With Carotid Atherosclerosis

Abstract

Current risk assessment for patients with carotid atherosclerosis relies primarily on assessing the degree of stenosis. More reliable risk stratification could improve patient selection for more aggressive treatment. We used artificial neural network algorithms to generate a model to predict major adverse neurologic events (MANE; stroke, transient ischemic attack, and amaurosis fugax) incorporating a combination of carotid plaque geometry (two-dimensional and three-dimensional), plaque tissue composition, patient demographics, and clinical information. Images from computed tomography angiograms of 70 patients asymptomatic at baseline, and follow-up, from four institutions were analyzed for carotid plaque morphology using a semiautomatic image-analysis program (vascuCAP, Elucid Bioimaging, Boston, Mass), yielding 193 analyzable arteries. We recorded patient demographic and clinical information. All visible plaque in each carotid artery bifurcation was analyzed regardless of size. Measurements included two-dimensional geometry (plaque thickness, percent stenosis by diameter and area), three-dimensional geometry (volume), and tissue composition (calcium, lipid-rich necrotic core, and intraplaque hemorrhage). All demographic, clinical, and plaque morphologic features were used in the analysis. First, univariate correlation and supervised/unsupervised clustering was performed. Next, several high-performance multivariate models in an artificial neural network were implemented (logistic regression, glmnet penalization, support vector machine with radial basis function, C5.0, partial least squares, and averaged neural nets) for the response variable of MANE. Finally, models were selected after cross-validation by optimizing the area under the receiver operating characteristic curve (AUC/ROC). The mean age was 69.2 ± 8.46 years and 65.7% were male; other demographic and clinical features in the model are listed in the Table. After baseline imaging, 32 carotid artery plaques were associated with an ipsilateral MANE on follow-up. Unsupervised clustering of clinical, demographic, and morphologic features identified putative univariate predictors of MANE (Fig). Artificial neural network analyses generated ROC curves for various combinations of clinical and plaque features that best predicted MANE (Fig). A combination of morphologic features at baseline (volumes for plaque matrix, lipid-rich necrotic core, calcium and intraplaque hemorrhage) best predicted MANE (AUC/ROC, 0.86) while the percent diameter stenosis performed worst (AUC/ROC, 0.45). Adding longitudinal data on plaque progression did not improve the AUC/ROC. We present a novel application of artificial intelligence algorithms for risk stratification in carotid atherosclerosis. A composite of carotid plaque geometry and tissue composition, patient demographics, and clinical information has better predictive performance for MANE than the traditionally used degree of stenosis alone. Implementing this predictive model on asymptomatic patients in the clinical setting will help identify those at a high-risk for future major MANE.TablePatient demographic and clinical features tested in the predictive modelVariable (at baseline)Mean ± SD or percentAge, years69.2 ± 8.46Male sex65.7Race White60 Black37 Asian3Diabetes mellitus46Coronary artery disease50Peripheral arterial disease27Hypertension93Hyperlipidemia87Smoking (past or present)64Total cholesterol (mg/dL)163.77 ± 40.88 Triglycerides132.41 ± 75.72 LDL88.37 ± 37.17 HDL49.08 ± 17.83Body mass index (kg/m2)28.69 ± 4.80Systolic blood pressure, mm Hg142.2 ± 22.04Diastolic blood pressure, mm Hg74.34 ± 12.58Resting heart rate, per minute73.10 ± 11.91Statin use86Anticoagulant use14Low-dose aspirin use77HDL, High-density lipoprotein; LDL, low-density lipoprotein; SD, standard deviation. Open table in a new tab