Abstract Body: Introduction: Embolic stroke of undetermined source (ESUS) may be associated with nonstenotic carotid artery plaques. Noncalcified plaque components such as intraplaque hemorrhage (IPH) as well as perivascular adipose tissue (PVAT) are associated with increased stroke risk, but the role of plaque calcifications is unclear. We examine a machine-learning approach using eXtreme Gradient Boosting (XGBoost) to classify carotid plaques as vulnerable or stable using non-calcified plaque features and calcification morphology.
Methods: Patients with neck CT angiography and unilateral anterior circulation ESUS with calcific carotid plaque were retrospectively analyzed. Derived by a combination of manual scoring by a neuroradiologist blinded to stroke side and semi-automated plaque composition segmentation software (Elucid), plaque-level features included plaque thickness, ulceration, composition volumes (IPH, lipid-rich necrotic core, matrix) and PVAT. Calcification-level features, segmented/scored manually with 3D Slicer, included spotty calcification (arc <90 degrees, thickness <3 mm), rim-sign (adventitial site, >90 degrees arc, and 2 mm non-calcified plaque), volume, surface area, roundness, flatness and density. Plaques ipsilateral to stroke side were defined as vulnerable and stable if contralateral. XGBoost models were trained to classify plaques as vulnerable/stable using 1) plaque-level, 2) calcification-level and 3) combination of both features. Data were split into training (80%) and test (20%) sets with 5-fold cross-validation (CV) to ensure consistency. Performance was measured by ROC AUC and accuracy.
Results: 71 patients were included [116 calcific carotid plaques; 60 ipsi-, 56 contralateral to stroke; 270 calcifications (146 ipsi-, 124 contralateral)]. 11 plaque-level and 16 calcification-level metrics were extracted. Plaque-level model achieved AUC 0.38 and accuracy 0.38. Calcification-level model achieved AUC 0.60 and accuracy 0.56. Combined model achieved AUC 0.97 and accuracy 0.93, with 5-fold CV score range 0.70-0.93 and mean 0.80. Five most important features for the Combined model were plaque ulceration, calcification surface area, ratio of PVAT to total plaque volume, plaque thickness and PVAT volume.
Conclusions: XGBoost model trained with a combination of noncalcified plaque and calcification features can classify plaque as vulnerable with > 90% accuracy, superior to models exclusively trained with plaque-level or calcification-level features.