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

Area of Science:

• Neurology
• Cardiovascular Medicine
• Artificial Intelligence

Background:

• Embolic stroke of undetermined source (ESUS) is linked to carotid artery plaques, even with <50% stenosis.
• Plaque vulnerability is influenced by intraplaque hemorrhage (IPH), lipid-rich necrotic core, perivascular adipose tissue (PVAT), and calcifications.
• Traditional plaque assessment lacks interpretability, especially with machine learning (ML) models.

Purpose of the Study:

• To apply an explainable ML approach using SHapley Additive exPlanations (SHAP) to classify culprit vs. nonculprit carotid plaques.
• To identify key plaque and calcification features predictive of stroke causality.
• To enhance clinical interpretability of ML models in stroke prediction.

Main Methods:

• Retrospective analysis of unilateral anterior circulation ESUS patients with calcified carotid plaques on CT angiography.
• Extraction of calcification-level and plaque-level features, including PVAT volume.
• Benchmarking of eight ML classifiers, with gradient-boosted decision tree (CatBoost) tuned and explained using SHAP.

Main Results:

• An ML model using five plaque/calcification features achieved ROC-AUC 0.79, outperforming plaque thickness (0.59) and IPH presence (0.51).
• SHAP analysis identified plaque thickness (>2.6 mm) and PVAT volume (≥112 mm³) as most influential features.
• The model demonstrated superior classification accuracy for culprit calcified carotid plaques.

Conclusions:

• ML models incorporating noncalcified plaque and calcification features offer improved classification of culprit carotid plaques in ESUS.
• Explainable ML (SHAP) provides clinically interpretable insights and suggests potential thresholds for plaque vulnerability.
• This approach enhances understanding of plaque characteristics associated with embolic stroke.