Prediction Model for Carotid Intima-Media Thickness Progression via Machine Learning and SHAP Analysis: A Three-Year Prospective Cohort Study

An Zhou¹Kui Chen²yonghui wei¹qun ye¹Yuanming Xiao²Rong Shi¹Jiangang Wang^2*Wei-Dong Li^1*

• ¹Department of Genetics, College of Basic Medical Sciences, Tianjin Medical University, Tianjin, China
• ²Health Management Center, Xiangya Hospital, Central South University, Hunan, Hunan Province, China

Background: Carotid intima-media thickness (CIMT) serves as a simple,noninvasive examination method for cardiovascular disease risk assessment and clinical treatment decision guidance. By predicting CIMT progression, we can advance the prediction window for atherosclerotic cardiovascular disease (ASCVD), enabling earlier detection and treatment to reduce morbidity risk. Objective: To develop and compare machine learning models for predicting CIMT thickening (defined as an absolute threshold ≥1.0 mm or significant progression ≥0.1 mm increase) on the basis of routine clinical biochemical parameters. Methods: From 128,938 records (31,158 individuals) collected since 2015 from the "XiangYa Third Hospital of Central South University Health Examination Cohort," 904 participants who completed three consecutive annual CIMT measurements were included. CIMT thickening was defined as a final examination CIMT ≥1.0 mm or an increase ≥0.1 mm across consecutive measurements. After Boruta algorithm preprocessing and feature selection, seven machine learning models (logistic regression, random forest, XGBoost, support vector machine, elastic net, decision tree, and neural network) were developed and evaluated in the test cohort. Platt scaling was applied to optimize calibration, and decision curve analysis (DCA) was used to assess clinical utility.Results: Among 904 participants, 227 (25.1%) developed CIMT thickening during follow-up. The elastic net model demonstrated superior performance (AUC=0.763, sensitivity=0.706, specificity=0.700). Standard calibration analysis revealed significant calibration bias in all the models, with expected calibration errors (ECEs) ranging from 0.1368--0.2889. After Platt scaling, the calibration improved significantly (ECE was reduced to 0.0131--0.0617), representing calibration error reductions of 71.7%-95.5%. DCA revealed that across a broad threshold probability range (0.2--0.6), the elastic net and support vector machine models provided greater net benefits than the "treat all" or "treat none" strategies did. Model interpretability was examined via SHapley Additive exPlanations (SHAP) analysis. Conclusion: Our machine learning approach effectively predicts CIMT progression within three years.The Platt-scaled elastic net model demonstrated optimal performance in both discrimination and calibration. SHAP analysis provides clinical interpretability by identifying key risk factors. DCA confirmed the models' clinical utility across reasonable threshold probability ranges. This predictive tool could facilitate targeted early intervention strategies by identifying individuals at highest risk for CIMT progression, thereby preventing cardiovascular disease.