Explainable Machine-Learning Model to Classify Culprit Calcified Carotid Plaque in Embolic Stroke of Undetermined Source

Background and Purpose: 

Embolic stroke of undetermined source (ESUS) may be associated with carotid artery plaques with <50% stenosis. Plaque vulnerability is multifactorial, possibly related to intraplaque hemorrhage (IPH), lipid-rich necrotic core, perivascular adipose tissue (PVAT), and calcifications. Machine learning (ML)-based plaque classification is increasingly popular but often limited in clinical interpretability by black-box nature. We applied an explainable ML approach, using noncalcified plaque components and calcification features with the SHapley Additive exPlanations (SHAP) framework to classify plaques as culprit or nonculprit. 

Methods: 

This was a retrospective, cross-sectional study. Patients with unilateral anterior circulation ESUS with calcified carotid plaques in neck computed tomography (CT) angiography were analyzed. Calcification-level features were derived from manual segmentations. Plaque-level features were assessed by a neuroradiologist and by semi-automated software. Plaques were classified as culprit if ipsilateral to stroke side. Eight classifiers were benchmarked, and a gradient-boosted decision tree (CatBoost) was further tuned. SHAP explained model decisions. 

Results: 

Seventy patients yielded 116 calcified plaques (270 calcifications). Model based on five plaque- and calcification-level features achieved ROC-AUC (receiver operating characteristic area under the curve) 0.79 and precision-recall-AUC 0.86, outperforming classification based on plaque thickness ≥3 mm (ROC-AUC 0.59, p = 0.04) and IPH presence (ROC-AUC 0.51, p = 0.003). SHAP identified plaque thickness and PVAT volume as the most influential features with potential thresholds of >2.6 mm and ≥112 mm³, respectively.f. 

Conclusions: 

ML model trained with noncalcified plaque and calcification features can classify culprit calcified carotid plaque better than conventional criteria. Using clinically interpretable features with SHAP, the model explained its decisions and suggested hypothesis-generating thresholds.