Abstract
BACKGROUND: Coronary artery disease is a critical public health issue. While traditional biomarkers like cardiac troponin show high sensitivity, their utility in unstable angina (UA) patients is limited. Recent research has identified novel biomarkers, including ceramides, which may correlate with atherosclerotic plaque stability and acute coronary syndrome (ACS) risk. This study aimed to develop a machine learning-based model to evaluate ceramide's diagnostic efficacy in patients with UA. METHODS: We conducted a prospective study from April 2021 to August 2022 at two medical centers, comparing data from UA patients with controls (stable angina and non-coronary patients). Data were split into training and validation sets. We assessed seven machine learning algorithms, including Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Distributed Random Forest (DRF), Extremely Randomized Trees (XRT), Deep Neural Networks (DNN), and Generalized Linear Models (GLM), using stacked ensemble models for improved predictions. Predictive accuracy was evaluated with the receiver operating characteristic area under the curve (ROC AUC), and model calibration was assessed using calibration curves. Feature importance was analyzed using the SHAP method. RESULTS: A total of 835 participants were recruited for the study, comprising 649 individuals diagnosed with unstable angina (UA) and 186 control subjects. The dataset was randomly divided into a derivation cohort consisting of 664 participants and a validation cohort comprising 171 participants. After comparing predictive models, we selected a generalized linear model with L1 regularization, demonstrating a strong ability to distinguish UA patients from controls (ROC AUC: 0.85, 95% CI 0.81-0.88). Calibration curve analysis indicated an intercept of 0.00 (95% CI - 0.21-0.21) and a slope of 1.05 (95% CI 0.86-1.25). Cer16:0 and Cer24:1 emerged as key predictors based on SHAP values. CONCLUSION: The machine learning model combining ceramide and conventional clinical variables effectively diagnoses UA, improving clinical efficacy. Future studies will validate and refine this model through expanded multicenter research.