Objective To construct a model for predicting the risk of symptomatic ischemic cerebrovascular disease (ICVD) based on carotid plaque characteristics utilizing Automated Machine Learning (AutoML) technology, systematically identify key predictive factors, and provide evidence for clinical risk stratification and individualized intervention. Methods A single-center retrospective study design was employed, enrolling 626 patients with carotid plaques who were treated between January 2020 and December 2022. Structured electronic medical records (EMRs) were used to extract comprehensive clinical data, including: Demographic characteristics (gender, age); Cardiovascular risk factors (e.g., hypertension, diabetes mellitus); Lifestyle habits (smoking, alcohol consumption); Laboratory parameters (blood lipid profiles, C-reactive protein); Ultrasound-evaluated carotid plaque characteristics (stenosis severity, ulcer formation, plaque number, intraplaque neovascularization). The dataset was divided into a training set (501 patients, ∼80%) and a test set (125 patients, ∼20%). Utilizing the AutoML framework, we implemented the Improved Newton-Raphson Based Optimizer (INRBO) to optimize model hyperparameters. Feature importance was validated through dual-dimensional analysis employing LASSO regression and SHAP (SHapley Additive exPlanations) interpretability models. Furthermore, an interactive nursing decision support system was developed using MATLAB. Results Among the 626 patients, 375 (59.90%) developed symptomatic ICVD. The prediction model constructed in this study demonstrated significantly enhanced performance: On the training set: ROC-AUC rose to 0.9537 and PR-AUC improved to 0.9522. On the independent test set: ROC-AUC remained high at 0.9343 and PR-AUC was 0.9104. These results consistently surpassed all other comparative models. The model definitively identified six core variables predicting symptomatic ICVD onset: Stenosis Severity; Ulcerative Plaque; Plaque Number; Intraplaque Neovascularization; Age; Diabetes Status. LASSO regression analysis independently selected seven variables, achieving an 85.71% overlap rate (6 out of 7 features) with the features selected by the AutoML model. SHAP analysis confirmed the top three feature importance rankings: (1) Stenosis Severity, (2) Ulcerative Plaque, (3) Plaque Number. Conclusion By integrating multidimensional clinical data with interpretable machine learning techniques, this study confirms the pivotal role of carotid plaque morphological features and metabolic factors in symptomatic ICVD risk prediction. Crucially, it achieves the real-time translation of risk assessment into actionable intervention decisions, thereby providing innovative tools and methodological advances for the precision diagnosis and treatment of cerebrovascular diseases.
Chen et al. (Mon,) studied this question.