Predicting carotid plaques in metabolic dysfunction-associated steatotic liver disease using machine learning and SHAP interpretation

Abstract

Abstract Cardiovascular disease (CVD) remains the most common cause of death worldwide. Carotid plaque is an indicator of subclinical CVDs. Metabolic dysfunction-associated steatotic liver disease (MASLD) is a risk factor for atherosclerotic CVDs. We aimed to develop and validate a predictive model for carotid plaque occurrence in annual health check-up populations, to integrate health check-up indicators with machine learning (ML) algorithms and LASSO-based feature selection and leverage advanced interpretability frameworks to elucidate the contribution of individual risk factors. In this retrospective cohort study, we enrolled 4,973 MASLD patients, among whom 1,178 were diagnosed with carotid plaques using carotid ultrasound. Collected baseline data included ​demographic indicators, ​clinical histories, blood ​biochemical parameters, and liver function test indicators. A predictive model for carotid plaques was developed and validated using five ML algorithms. Model performance was evaluated based on the​ area under the curve, ​sensitivity, ​specificity, ​accuracy, and ​F1 Score. For model interpretability, we adopted the ​Shapley Additive Explanations (SHAP) framework to quantify the contribution of individual features to the prediction outcomes. Among the five ML algorithm models, the support vectors machine model demonstrated superior discriminative capability, higher goodness-of-fit, and greater clinical utility compared to other ML algorithm models. Moreover, age, systolic blood pressure, total cholesterol, sex, and fasting plasma glucose were the most important risk factors associated with carotid plaques in the MASLD population. This study demonstrated the feasibility of constructing a predictive model for carotid plaques in MASLD populations using health check-up indicators combined with ML algorithms. The application of SHAP methods enhanced model interpretability by quantifying the contribution of individual risk factors to prediction outcomes, enabling clinicians to identify high risk MASLD patients prone to carotid plaque development, so that they can adjust interventions accordingly.