Risk stratification of asymptomatic carotid plaque: a comparison of symptom onset and treatment eligibility prediction models

Identifying high-risk, vulnerable plaques in asymptomatic patients is a significant clinical challenge for stroke prevention. This study compares two distinct machine learning approaches for risk stratification: one model predicts future biological events (symptom onset), while the other models complex clinical decisions (treatment eligibility). We used an integrated dataset from 18 patients diagnosed with asymptomatic carotid stenosis, combining multi-modal MRI radiomics derived from 3D Turbo Spin Echo (TSE) and 3D Time-of-Flight (TOF) sequences with key clinical features. A 'symptom onset model' was developed to predict future ischemic stroke. This model used patient-level labels (7 symptomatic vs. 11 asymptomatic), employed Sequential Feature Selection (SFS) for feature optimization, and was rigorously evaluated using leave-one-out cross-validation. Concurrently, a 'treatment eligibility model' was developed to predict the clinical decision for surgical intervention within one year. This second model used artery-level labels (13 surgical vs. 23 follow-up), applied mutual information (MI) for feature selection, and was evaluated via stratified 5-fold cross-validation. For symptom onset prediction, logistic regression was most accurate (AUC = 0.909, Recall = 0.857), demonstrating robust strength in minimizing false negatives. For treatment eligibility prediction, a Random Forest model performed best (AUC = 0.933, Recall = 0.923), indicating high predictive power for clinical decision-making. The symptom onset model effectively forecasts the direct clinical outcome of a stroke. In contrast, the treatment eligibility model excels at modeling the complex clinical reasoning for intervention. We conclude these two approaches are complementary, holding significant potential to improve the precision of personalized medicine for managing asymptomatic patients.