ABSTRACT
OBJECTIVE: This study aimed to develop and validate a machine learning model incorporating both dual-energy computed tomography (DECT) angiography quantitative parameters and clinically relevant risk factors for the identification of symptomatic carotid plaques to prevent acute cerebrovascular events. METHODS: The data of 180 patients with carotid atherosclerosis plaques were analysed from January 2017 to December 2021; 110 patients (64.03±9.58 years old, 20 women, 90 men) were allocated to the symptomatic group, and 70 patients (64.70±9.89 years old, 50 women, 20 men) were allocated to the asymptomatic group. Overall, five machine learning models using the XGBoost algorithm, based on different CT and clinical features, were developed in the training cohort. The performances of all five models were assessed in the testing cohort using receiver operating characteristic curves, accuracy, recall rate, and F1 score. RESULTS: The shapley additive explanation (SHAP) value ranking showed fat fraction (FF) as the highest among all CT and clinical features and normalised iodine density (NID) as the 10th. The model based on the top 10 features from the SHAP measurement showed optimal performance (area under the curve [AUC] .885, accuracy .833, recall rate .933, F1 score .861), compared with the other four models based on conventional CT features (AUC .588, accuracy .593, recall rate .767, F1 score .676), DECT features (AUC .685, accuracy .648, recall rate .667, F1 score .678), conventional CT and DECT features (AUC .819, accuracy .740, recall rate .867, F1 score .788), and all CT and clinical features (AUC .878, accuracy .833, recall rate .867, F1 score .852). CONCLUSION: FF and NID can serve as useful imaging markers of symptomatic carotid plaques. This tree-based machine learning model incorporating both DECT and clinical features could potentially comprise a non-invasive method for identification of symptomatic carotid plaques to guide clinical treatment strategies.
