Development and validation of a multimodal machine learning model integrating ultrasound imaging and serum biomarkers for vulnerable carotid plaque prediction

OBJECTIVE: To establish and validate the radiomics model of carotid vulnerable plaque. METHODS: 182 patients who underwent carotid endarterectomy in the First Affiliated Hospital of Soochow University from January 2019 to June 2022 were retrospectively analyzed. Plaque ultrasound images were acquired and segmented, and features were extracted using the R package "EBImage". Six machine learning algorithms were used to construct a plaque vulnerability classifier model. The relationship between clinical biomarkers and outcomes was evaluated using linear regression and the False Discovery Rate (FDR) correction method. RESULTS: Diabetes, neutrophils, monocytes, high-sensitivity C-reactive protein, high-density lipoprotein, and stenosis rate were found to have a strong association with plaque vulnerability. The Naive Bayes algorithm performed well in the training set using image features alone, with an AUC of 0.840, and an AUC of 0.762 in the test set. The Decision Tree algorithm had certain performance in the training set using image features alone, with an AUC of 0.609, and an AUC of 0.626 in the test set. The Naive Bayes algorithm achieved excellent performance in the training set using both plaque ultrasound imaging features and clinical laboratory indicators, with an AUC of 0.922, and an AUC of 0.928 in the test set. CONCLUSION: The above radiomics model can be used to predict the vulnerable carotid plaque before surgery. By combining plaque ultrasound imaging with clinical blood biomarkers, a more comprehensive and accurate assessment of plaque vulnerability can be achieved, thereby reducing the risk of cardiovascular events.