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

Key Points

• This research aims to develop an explainable machine learning model to classify calcified carotid plaques as either culprit or nonculprit in embolic stroke of undetermined source.
• Conducted a retrospective, cross-sectional study on patients with unilateral anterior circulation ESUS and calcified plaques from CT angiography.
• Derived calcification-level features through manual segmentations and assessed plaque-level features via neuroradiologist evaluation and semi-automated software.
• Used eight different classifiers, with a focus on a tuned gradient-boosted decision tree (CatBoost) for plaque classification.
• Applied SHapley Additive exPlanations (SHAP) to explain model decisions.
• Model achieved ROC-AUC of 0.79 and precision-recall-AUC of 0.86, outperforming conventional classification methods.
• Plaque thickness ≥3 mm had ROC-AUC of 0.59, while the presence of intraplaque hemorrhage had ROC-AUC of 0.51.
• SHAP identified plaque thickness >2.6 mm and PVAT volume ≥112 mm³ as influential features with potential clinical thresholds.