Clinical Value of Machine Learning Model Based on Non-Invasive Myocardial Work Parameters and Glycolipid Metabolism Indicators for the Early Diagnosis of Coronary Heart Disease

Abstract

Objective This study aims to explore the clinical value of machine learning models constructed by combining non-invasive Myocardial Work (MW) parameters and clinical glycolipid metabolism indicators for the early diagnosis of coronary heart disease (CHD). Methods A total of 183 suspected CHD patients scheduled for coronary angiography (CAG) admitted to the Affiliated Hospital of Youjiang Medical University for Nationalities from October 2024 to June 2025 were enrolled in this study. Subjects were randomly divided into a training set (n=129) and a validation set (n=54) at a ratio of 7:3. General clinical data, laboratory test results and echocardiographic parameters of all enrolled subjects were collected. Univariate Logistic regression analysis, multicollinearity test and multivariate Logistic regression analysis were performed on the training set to screen independent predictive factors for CHD, based on which six distinct machine learning models were established. The optimal machine learning model was selected comprehensively according to the area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity, specificity, F1-score and Youden’s index. Receiver operating characteristic (ROC) curves, calibration curves and clinical decision curves were adopted to evaluate the diagnostic performance and clinical benefit of the model. Finally, SHapley Additive exPlanations (SHAP) algorithm was applied to conduct interpretability analysis of the final model. Results Multivariate Logistic regression analysis revealed that Metabolic Syndrome(MetS) (OR=3.070, P=0.004), Triglyceride-Glucose Index(TyG index) (OR=24.710, P<0.001), High Density Lipoprotein Cholesterol(HDL) (OR=0.224, P=0.012), Global Constructive Work(GCW) (OR=0.999, P=0.004) and Global Wasted Work(GWW) (OR=1.005, P=0.027) were independent predictive factors for CHD. The Multilayer Perceptron (MLP) model achieved the optimal performance, with an AUC of 0.799, an ACC of 0.796, an F1-score of 0.820, a sensitivity of 0.833, a specificity of 0.750 and a Youden’s index of 0.583 in the validation set. Calibration curve and clinical decision curve analyses demonstrated that the predicted probabilities of the MLP model were in good agreement with the observed actual probabilities, and the model yielded considerable clinical net benefit across a wide range of threshold probabilities. The SHAP algorithm was adopted to visualize variable importance within the MLP model. The results indicated that GCW exerted the greatest predictive contribution, followed by MetS and GWW with moderate contributions, HDL with moderate-to-low contribution, and the TyG index with the lowest contribution. Conclusions The MLP machine learning model constructed based on GCW, GWW, MetS, HDL and TyG index exhibits favorable diagnostic performance for the early identification of CHD, which can provide reference for clinical decision-making.