AI-based Assessment of Risk Factors for Coronary Heart Disease in Patients With Diabetes Mellitus and Construction of a Prediction Model for a Treatment Regimen

Abstract

Background: This study aimed to construct a prediction model for a treatment plan for patients with coronary artery disease combined with diabetes mellitus using machine learning to efficiently formulate the treatment plan for special patients and improve the prognosis of patients, provide an explanation of the model based on SHapley Additive exPlanation (SHAP), explore the related risk factors, provide a reference for the clinic, and concurrently, to lay the foundation for the establishment of a multicenter prediction model for future treatment plans.

Methods: To investigate the relationship between concomitant coronary heart disease (CHD) and diabetes mellitus (DM), this study retrospectively included patients who attended the Beijing Anzhen Hospital of Capital Medical University between 2022 and 2023. The processed data were then input into five different algorithms for model construction. The performance of each model was rigorously evaluated using five specific evaluation indicators. The SHAP algorithm also provided clear explanations and visualizations of the model's predictions.

Results: The optimal set of characteristics determined by the least absolute shrinkage and selection operator (LASSO) regression were 15 features of general information, laboratory test results, and echocardiographic findings. The best model identified was the eXtreme Gradient Boost (XGBoost) model. The interpretation of the model based on the SHAP algorithm suggests that the feature in the XGBoost model that has the greatest impact on the prediction of the results is the glycated hemoglobin level.

Conclusions: Using machine-learning algorithms, we built a prediction model of a treatment plan for patients with concomitant DM and CHD by integrating patients' information and screened the best feature set containing 15 features, which provides help and strategies to develop the best treatment plan for patients with concomitant DM and CHD.

Keywords: SHapley Additive exPlanation; coronary heart disease; diabetes mellitus; machine learning; predictive modeling.

Fig. 1.

Flowchart of AI framework establishment. The population with coronary heart disease combined with diabetes mellitus, sourced from the Coronary Heart Disease Specialized Database of the Beijing Anzhen Hospital affiliated with the Capital Medical University between 2022 and 2023, was retrospectively included. A total of 3171 cases were exported. After cleaning, which involved removing cases that did not meet the inclusion criteria and those with more than 30 missing features, 3153 cases remained. These cases were then categorized based on whether the patients had undergone treatment (percutaneous interventional or coronary artery bypass grafting). AI, artificial intelligence.

Fig. 2.

Evaluation metrics for five machine learning algorithms. (A) Calibration curve of the KNN. (B) Calibration curve of the LRs. (C) Calibration curve of the XGBoost. (D) Calibration curve of the RF. (E) Calibration curve of the SVM. (F) Performance of five machine learning algorithms. (G) Confusion matrix for KNN. (H) Confusion matrix for LR. (I) Confusion matrix for XGBoost. (J) Confusion matrix for RF. (K) Confusion matrix for SVM. (L) Comparison of subject work characteristics (ROC) curves for the five machine learning models. ROC, receiver operating characteristic curve.

Fig. 3.

Scatterplot of the relationship between continuous and ending variables in the features. (A) Hospitalized activated partial thromboplastin time test results vs surgery or not. (B) Internal diameter of pulmonary artery trunk vs surgery or not. (C) Glomerular filtration rate vs surgery or not. (D) Alanine aminotransferase assay value vs surgery or not. (E) Free thyroxine (FT4) test results vs surgery or not. (F) Maximum aortic flow velocity vs surgery or not. (G) Age vs surgery or not. (H) Random blood glucose test results vs surgery or not. (I) Hemoglobin test results vs surgery or not. (J) Glycated hemoglobin test results vs surgery or not. (K) HDL cholesterol test results vs surgery or not. (L) LV diastolic E-wave flow velocity max vs surgery or not. (M) Results of coefficient of variation of erythrocyte distribution width vs surgery or not.

Fig. 4.

Correlation and distribution between features. (A) Comparison of baseline characteristics between the two data sets. (B) Spearman correlation analysis between features. (C) Pearson correlation coefficient between continuous variables in characteristics.

Fig. 5.

SHAP value. (A) Feature importance plot in XGBoost model. (B) SHAP summary plot in XGBoost model. SHAP, SHapley Additive exPlanation.