Risk stratification in intermediate asymptomatic carotid artery disease: a multi-modal approach combining local flow dynamics and systemic clinical profiles.

Background/objectives: Carotid artery stenosis is a leading cause of ischemic stroke, yet accurate patient-specific prediction of its progression remains an unresolved clinical challenge. Neither baseline stenosis severity nor conventional cardiovascular risk factors can reliably identify which asymptomatic patients will progress. This study aimed to develop and validate a machine learning (ML) framework for predicting carotid stenosis progression by integrating patient-specific computational fluid dynamics (CFD) features with clinical and imaging data. Methods: A dataset of 146 carotid arteries from 129 asymptomatic patients enrolled in the multi-center TAXINOMISIS project was used. Stenosis progression was defined as a > 10% increase at the latest available follow-up assessment (56/146 arteries, 38.4%). Twenty-two features were derived from clinical risk factors and CFD-based hemodynamic metrics including time-averaged wall shear stress (TAWSS), oscillatory shear index (OSI), relative residence time (RRT), and disturbed flow indices. ML classifiers were trained using patient-level GroupShuffleSplit, Optuna Bayesian hyperparameter optimization (200 trials), patient-level 5-fold GroupKFold cross-validation, 1,000-iteration bootstrap confidence intervals, and calibration analysis. Performance was benchmarked against two clinically motivated baselines: baseline stenosis grade alone and clinical risk factors alone. Results: The proposed pipeline achieves substantially better discrimination than either clinical baseline: the best-performing model (RF + XGBoost Ensemble, Optuna-tuned) reached AUC=0.814 (95% CI: 0.702-0.917; Sensitivity=0.609; Specificity=0.828; Brier score=0.184), versus AUC=0.552 for stenosis grade alone and AUC=0.459 for clinical risk factors alone - neither of which provides meaningful discrimination. The combined stenosis + clinical model (AUC=0.388) also failed to improve upon either comparator, confirming that hemodynamic information is the necessary added value. Feature contribution analysis consistently identified the normalized area of low TAWSS (Low TAWSS %) as the dominant predictor across all classifiers, followed by the disturbed flow area (LOHA) and vessel-average OSI. Conclusions: A multimodal, physics-informed ML pipeline integrating patient-specific CFD features substantially outperforms all clinical benchmarks for predicting carotid stenosis progression. These findings demonstrate that hemodynamic phenotyping is necessary to advance beyond current risk stratification approaches in asymptomatic carotid disease and support the development of CFD-integrated ML as a clinical decision-support tool in vascular surgery.