Abstract / Summary
Abstract Background Carotid atherosclerosis, an important subclinical manifestation of vascular injury, is closely associated with stroke, cardiovascular events, and systemic atherosclerotic burden. This study aimed to identify stable predictive factors for carotid atherosclerosis, characterize the interrelationships among these factors, and quantify their probabilistic impact on carotid atherosclerosis. Methods A total of 22,012 participants who underwent carotid ultrasound examinations at the Health Management Center of a tertiary hospital in Lanzhou, Northwest China, between 2023 and 2024 were included. Carotid atherosclerosis was defined as the presence of carotid plaques detected by ultrasonography. The dataset was randomly divided into training and testing sets at a ratio of 7:3. Five machine learning models were developed: LASSO-regularized logistic regression, support vector machine, random forest, decision tree, and CatBoost. Model performance was assessed using the area under the receiver operating characteristic curve, precision-recall curves, and decision curve analysis. Variables consistently identified by the best-performing models were used to construct a Bayesian network. Network structure was learned using the Maximum-Minimum Hill-Climbing algorithm, and conditional probabilities were calculated to quantify changes in the posterior probability of carotid atherosclerosis under different risk-factor patterns. Results Among 22,012 participants, 8,987 (40.8%) were classified as having carotid atherosclerosis, while 13,025 (59.2%) were classified as having no carotid atherosclerosis. Among the five models, LASSO and CatBoost showed the best discrimination, each achieving an AUC of 0.803. LASSO provided higher sensitivity and more balanced classification performance, whereas CatBoost achieved the best overall accuracy, calibration, specificity, and positive predictive value. The two models identified 10 common predictors for Bayesian network construction. The final network contained seven nodes and 14 edges, with carotid atherosclerosis related conditional dependencies mainly involving older age, male sex, overweight, dyslipidemia, and elevated fasting blood glucose. The highest posterior probability of carotid atherosclerosis, 78.3%, was observed in individuals aged ≥ 60 years who were overweight and had elevated triglycerides and fasting blood glucose. Age and fasting blood glucose emerged as the most prominent nodes: age ≥ 60 years produced the largest increase from the prior probability, while elevated fasting blood glucose was consistently associated with higher carotid atherosclerosis probability across combinations with and without age. Gender-stratified analyses showed acceptable performance in both sexes, although discrimination and calibration were weaker in women. Conclusion Compared with conventional risk prediction alone, this integrated prediction-inference framework may improve identification of individuals at high risk of carotid atherosclerosis while enhancing clinical interpretability. However, prospective studies and external validation in independent populations are required before clinical application. Clinical trial number Not applicable.