Introduction

Diabetes mellitus (DM) represents the most prevalent chronic disorder [1, 2]. In December 2021, the International Diabetes Federation (IDF) released the most recent Global Diabetes Map. On the basis of the latest report, approximately 537 million adults (aged 20–79) worldwide were afflicted with DM in 2021 (one in ten individuals was a DM patient). It is projected that this number will increase to 643 million by 2030 and reach 783 million by 2045 [3]. DM complicated with coronary heart disease (CHD) is a coronary atherosclerotic heart disease caused by metabolic disorders in DM patients [4,5,6]. CHD refers to a cardiovascular disease caused by atherosclerosis in the coronary arteries, which leads to stenosis or occlusion of the lumen, resulting in myocardial ischemia, hypoxia or necrosis. It is clinically divided into four mutually exclusive subtypes: angina pectoris, acute myocardial infarction, ischemic cardiomyopathy, and silent myocardial infarction [7]. The clinical symptoms of CHD can range from asymptomatic presentations to severe complications, such as arrhythmias, acute myocardial infarction, heart failure, and even sudden death [8,9,10,11]. Currently, coronary angiography remains the gold standard for diagnosing CHD [12]. However, it is an invasive procedure and may not be suitable for certain patients, such as those with contrast agent allergies or claustrophobia [13]. In addition, patients in the silent CHD stage frequently exhibit minimal or no clinical symptoms. Therefore, improving the understanding of CHD risk factors in DM patients and establishing predictive methods to guide personalized diagnostics, treatment plans, and follow-up management are critically needed.

Clinical big data is characterized by its enormous size, diversity and high value, and is a valuable resource for researchers [14, 15]. Combining medical datasets with artificial intelligence (AI) technology can create various computational models to predict health risks [16,17,18,19,20]. Although comorbidity models linking CHD and DM have been reported in some studies [21,22,23,24,25], these studies have significant limitations. First, some models were developed using the data from a single-center population, with a relatively limited sample size. Second, a majority of existing studies have incorporated traditional laboratory indicators as risk factors, however, these analyses are not exhaustive. Notably, several emerging biomarkers related to lipid metabolism and inflammation have yet to be integrated into models for comprehensive evaluation. Furthermore, machine learning (ML) techniques have not been widely adopted in many studies, and existing models often lack interpretability and intuitiveness. Therefore, an intuitive and interpretable model for predicting the risk of CHD in DM patients still needs to be developed.

In this study, we aim to develop and validate an interpretable machine learning model for CHD risk prediction in patients with DM using dual-center cohorts. We will include routine laboratory indicators and derived novel inflammation and lipid metabolism markers as candidate predictors, apply multiple variable screening methods to identify core features, and construct the model with the LGBM algorithm. We will further use SHAP to clarify model interpretability, and build a web-based prediction tool to facilitate clinical translation.

Methods

The entire experimental plan strictly adheres to the strengthening the reporting of observational studies in epidemiology (STROBE) [26] and transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) [27].

Survey description

This study is a retrospective cohort study. This study protocol complies with the guidelines of the Declaration of Helsinki and was approved by the Ethics Committee of the Sixth Affiliated Hospital of Kunming Medical University (China, trial registration number: 2022-kmykdx6f-90) and Yan’an Hospital Affiliated to Kunming Medical University (China, trial registration number: 2022–026–01). All participants provided written consent after being fully informed.

Participant enrollment

We collected both the basic information and medical records of patients with DM (both Type I and Type II) who were admitted to the Sixth Affiliated Hospital of Kunming Medical University (n = 5002) and Yan’an Hospital of Kunming City (n = 7122) between January 2022 and October 2024. The data of the two centers were collected simultaneously. The inclusion criteria were as follows: (1) the participants were diagnosed with DM and their detailed medical history and glycemic control history were recorded; (2) the participants underwent coronary angiography during their hospitalization and the angiography records were complete; (3) age ≥ 40. The exclusion criteria were as follows: (1) the participants with acute complications of DM (Diabetic ketoacidosis and hyperosmolar hyperglycemic state), gestational diabetes mellitus or DM diagnosed within past six months; (2) the participants with autoimmune diseases, including rheumatic heart disease and systemic lupus erythematosus; (3) the participants with cancer; (4) the participants previously diagnosed with CHD; (5) the participants with severe organ failure; (6) the participants with systemic infection.

Diagnostic criteria

According to the American Diabetes Association (2024) diabetes diagnostic criteria [28], DM was defined as an fasting plasma glucose (FPG) ≥7.0 mmol/L or glycated hemoglobin (HbA1c) ≥6.5%. The diagnostic criteria for CHD and hypertension (HTN) were based on the guidelines of the American Heart Association (2024) and American College of Cardiology (2024) [29]. According to the International Classification of Diseases 11th Revision of the World Health Organization (https://icd.who.int/zh), CHD can be classified into five categories, namely silent myocardial infarction (SMI), angina pectoris (AP), myocardial infarction (MI), ischemic cardiomyopathy (ICM) and sudden cardiogenic death (SCD). The diagnostic criteria for hyperlipidemia (HLD) was based on guidelines of the American Heart Association (2018) [30]. The diagnostic results were independently evaluated by an endocrinologist and a cardiologist.

Clinical features

The data in this study were included the basic patient information, laboratory indicators and clear diagnostic information during their hospitalization. The biochemical, lipid, metabolic and inflammatory indicators measured using the Roche cobas 8000 automated biochemical analyzer (Roche Diagnostics, Mannheim, Germany), including the levels of total bilirubin (TBIL), total protein (TP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), gamma-glutamyl transferase (GGT), total bile acid (TBA), blood urea nitrogen (BUN), serum creatinine (Scr), uric acid (UA), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), non-high-density lipoprotein cholesterol (non-HDL-C, calculated as TC minus HDL-C), remnant cholesterol (RC, calculated as TC minus HDL-C minus LDL-C), apolipoprotein A1 (ApoA1), apolipoprotein B (ApoB), lipoprotein(a) [Lp(a)], lactic acid (LA), homocysteine (HCY), glucose (Glu), hypersensitive C-reactive protein (hsCRP). The hematological indicators measured using the Mindray BC-6800 automated hematology analyzer (Mindray Medical International, Shenzhen, China), including the levels of white blood cell count (WBC), neutrophil count (NEUT), lymphocyte count (LYM), monocyte count (MONO), red blood cell count (RBC), hemoglobin (Hb), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC) and platelet count (PLT). Based on routine clinical laboratory parameters, we calculated two sets of candidate predictors, namely inflammation-related indices and novel lipid metabolism-related indices. The inflammation-related indices included PLR (platelet-to-lymphocyte ratio), NLR (neutrophil-to-lymphocyte ratio), LMR (lymphocyte-to-monocyte ratio), NPR (neutrophil-to-platelet ratio), HHR (hsCRP-to-HDL-C ratio), MHR (monocyte-to-HDL-C ratio), LHR (lymphocyte-to-HDL-C ratio), NHR (neutrophil-to-HDL-C ratio), PHR (platelet-to-HDL-C ratio), SII (systemic immune-inflammation index, platelet count×neutrophil count/lymphocyte count) and SIRI (systemic inflammation response index, white blood cell count×hematocrit/hs-CRP). The novel lipid metabolism-related indices included TyG (triglyceride-glucose index, ln [TG×FBG/2]), TyG-BMI (TyG×BMI) and AIP (atherogenic index of plasma, log [TG/HDL-C]). Outlier detection and processing were performed for all variables, and outlier-containing samples were excluded from the final cohorts.

Variable screening and importance ranking

Univariate and multivariate logistic regression analyses were used to preliminarily screen important factors for the risk of CHD in 12,124 DM patients on the based of the odds ratios (ORs) and P values. We further used the least absolute shrinkage and selection operator (LASSO) [31], Boruta [32] and recursive feature elimination (RFE) [33] methods for variable selection to improve model prediction performance. The specific methods were as follows: a) To select features with nonzero coefficients, we utilized LASSO regression. This method incorporates an L1 regularization term into the cost function and performs 5-fold cross-validation, with a maximum of 1000 iterations allowed for convergence; b) The Boruta method, which relies on repeated random forest analyses, is utilized to evaluate whether the results are independent of random variations; c) RFE is a feature selection method that iteratively trains a model, assesses the importance of each feature, and removes the least important ones until a specified number of features or an optimal model performance is achieved. In this study, the RFE with XGBoost was used to select variables (10-fold CV) and evaluate the model performance. The intersection of the variables selected by LASSO, Boruta, and XGBoost-RFE was determined, resulting in the identification of 12 variables that are most closely associated with the risk of CHD in DM patients. Finally, we employed eight ML models to rank the importance of the 12 selected variables.

Model development

A total of 5002 samples from the Sixth Affiliated Hospital of Kunming Medical University were subjected to 10-fold cross-validation as internal validation for model development. To predict the risk of CHD in DM patients, eight ML models were employed: logistic regression (LR), decision tree (DT), random forest (RF), K-nearest neighbor (KNN), eXtreme gradient boosting (XGB), light gradient boosting machine (LGBM), gradient boosting decision tree (GBDT) and adaptive boosting (AdaBoost). To enhance the performance of the prediction models, the optimal hyperparameters for each model were obtained using grid search by the “caret” package of the R language. Ultimately, the model was retrained using the optimal feature subset and hyperparameters to establish the final model.

Model performance comparison and sensitivity analysis

To evaluate the performance of the models, we utilized several common evaluation indicators including the area under the receiver operating characteristic (ROC) [34] curve (AUC), precision, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, recall and F1 score. We construct the PR (precision-recall) curve on the basis of precision values and recall values. Calibration curves [35] were constructed to illustrate the alignment between the predicted probabilities and actual outcomes. Decision curve analysis (DCA) [36] was conducted to evaluate the clinical net benefit of the models at various threshold probabilities. The Delong test was applied to determine if there were significant differences in the AUC values between the two models. We also performed net reclassification improvement (NRI) [37] and integrated discrimination improvement (IDI) [38] to compare the predictive performance of models established by any two ML methods. By comparing these evaluation indicators, we identified the optimal prediction model. In addition, we employed sensitivity analysis to explore the stability of the predictive performance of various ML methods under different CHD classification.

Model explanation

Given that clinicians are reluctant to accept predictive models lacking direct interpretability, we utilized the SHAP method to elucidate the final model’s output by quantifying each variable’s contribution to the prediction [39]. SHAP is a post hoc model explanation method and can explain “black box” models from both global and local perspectives.

Online prediction model

To enhance the practicality of the model in clinical environments, the final prediction model was incorporated into a web-based application platform using the Django framework. After the values of relevant features are input, this application can provide the possibility of CHD occurrence for DM patients.

Statistical analysis

Statistical Product and Service Solutions (version: 29.0), Python (version: 3.12.6) and R language (version: 4.4.1) were used for all analyses. In the present study, the basic and clinical characteristics of research participants with and without CHD were described. Continuous variables are expressed as the means ± standard deviations (SDs) and were compared via the unpaired t test and Mann‒Whitney U test. Categorical variables are presented as percentages and were compared via the chi-square test or Fisher’s exact test. With missing data in only 26 patients (0.21%) from routine variables, we performed complete case analysis after confirming no significant differences in baseline, predictors, and primary outcome between excluded and included samples (all p > 0.05), with no selection bias introduced. Univariate and multivariate logistic regression analyses were used to preliminarily screen important factors for the risk of CHD in DM patients on the basis of the OR values and P values (Python: statsmodels 0.11.1). We further used the LASSO CV (Python: scikit-learn 1.1.3), Boruta (R: Boruta packages) and XGBoost-RFE (Python: scikit-learn 1.1.3 and XGBoost 2.0.1) methods for data dimensionality reduction. The association between the dose-response of continuous variables and the risk of CHD in DM patients was determined via a restricted cubic spline (R: rms package). Eight ML models were developed using 10-fold CV and there predictive performance was evaluated via ROC curves, DCA, PR curves,calibration curves, etc. (Python: scikit-learn 1.1.3). On the Basis of the AUC value of the ROC curve, the DeLong test was employed to determine whether a significant difference existed in the AUC values between the two ML models (Python: scikit-learn 1.1.3). IDI and NRI were utilized to assess the enhancement in prediction performance of the new model relative to the baseline model (R: nricens packages). A p-value less than 0.05 was considered to indicate statistical significance.

Results

Baseline characteristics of the study participants

The baseline characteristics of the participants are shown in Table 1 and Table S1 which include the basic patient information, laboratory indicators and diagnostic information. A total of 12,124 participants were included in this study, among which 5,002 were for internal validation and 7,122 for external validation. The average age of the entire sample was 61.89 ± 10.23 years old, with 51.82% being male and 48.18% female. Among them, 1,750 men and 1,455 women had diabetes combined with CHD, accounting for 14.43% and 12.00% of the total population, respectively. The data was divided into a DM without CHD group (DM-CHD) and a DM with CHD group (DM+CHD), and the disparities between the two groups were compared with respect to each feature. As shown in Table 1 and Table S1, the participants in the DM+CHD group were older (p < 0.05) and had a higher BMI (p < 0.05), than those in the DM-CHD group. Furthermore, there were significant differences in multiple laboratory indicators and calculated index between the two groups. For example, the SCR, LDL-C, LA, Glu, SIRI and TyG in the DM+CHD group were significantly greater than those in the DM-CHD group, whereas the HDL-C level in the DM+CHD group was significantly lower than that in the DM-CHD group, with p-values all < 0.001. The specific follow-up study design is shown in Fig. 1.

Table 1 Basic and clinical characteristics between the groups of DM patients with or without CHD in internal validation cohort and external validation cohort
Full size table
Fig. 1
Fig. 1
Full size image

Flow chart of the study design. Abbreviations: DM: diabetes mellitus; CHD: coronary heart disease; LASSO: least absolute shrinkage and selection operator; REF: recursive feature elimination; ML: machine learning; LR: logistic regression; DT: decision tree; GBDT: gradient boosting decision tree; AdaBoost: adaptive boosting; LGBM: light gradient boosting machine; KNN: K-nearest neighbor; RF: random forest; XGBoost: eXtreme gradient boosting; CV: cross validation; ROC: receiver operating characteristic; PR: precision-recall; NRI: net reclassification improvement; IDI: integrated discrimination improvement; RCS:Restricted cubic spline; SHAP: SHapley additive exPlanations

Screening of variables related to the risk of CHD in DM patients

We screened the independent risk factors for CHD in DM patients in the entire cohort (n = 12124). As shown in Table S2, a total of 38 potential risk factors related to CHD in DM patients (p < 0.05) were detected via univariate logistic regression analysis. Following multivariate logistic regression analysis, 20 key factors significantly associated with the risk of CHD in DM patients were finally determined (p < 0.05), namely, sex, HTN (yes or no), HLD (yes or no), age, DBP, AST, ALP, BUN, SCR, UA, HDL-C, LDL-C, Lp (a), LA, Glu, HCT, MCV, MCHC, SIRI and TyG. To improve predictive ability of the model, the 20 factors previously identified were subjected to data dimensionality reduction using the LASSO, Boruta and XGB-REF algorithms. The optimal penalty term coefficient λ was identified via the 5-fold CV of the LASSO regression model. At log (λ-1se) = −4.48, a total of 11 potential predictors were selected (Fig. S1A, B), and their relative importance was subsequently ranked (Fig. 2A). A total of 19 variables were identified as important features, and 1 feature was identified as a rejected feature using Boruta algorithm (Fig. 2B). A total of 18 potential predictors were selected using the XGBoost-RFE method (Fig. 2C). Through taking the intersection of the LASSO, Boruta and XGBoost-RFE results, 11 factors associated with the risk of CHD in DM patients were finally identified, namely, HTN, HLD, age, DBP, SCR, LA, HDL-C, LDL-C, Glu, SIRI and TyG (Fig. 2D). The importance score of each variable for the 8 ML models was calculated and presented in the form of a bar chart (Fig. 2E–L) and summarized the all results (Fig. 2M).

Fig. 2
Fig. 2
Full size image

Screening of variables related to the risk of CHD and importance ranking. A. the importance of the selected features was evaluated using a LASSO regression model with 5-fold cross-validation. B: the importance of the variables is depicted as Z score boxplots ranked by the Boruta algorithm. Blue represents the average, minimum and maximum values of the shadow variables (random variables). Green represents passed important features, and red represents the rejected feature. C. Recursive feature elimination with XGBoost (XGBoost RFE) was used to find the ideal number of features. A high cross validation score can be achieved by using only 18 features. D. Taking the intersection of the results of the LASSO, Boruta and XGBoost RFE methods, 11 factors associated with the risk of CHD in DM patients were finally identified, which are presented using a venn diagram. E-L. Eight ML algorithms were used to rank the importance of each independent factor in predicting the risk of CHD in DM patients. M. Summary of the importance rankings of the variables in the eight ML algorithms

Dose-response relationship

On the basis of the feature selection results, we further explored the relationships among the continuous variables and the risk of CHD in DM patients. RCS curves are adept at modeling nonlinear associations between a predictor variable and an outcome. Moreover, RCS curves can also effectively capture and reflect threshold effects. The RCS curves revealed nonlinear relationships of age, LDL-C, Glu and TyG with the risk of CHD (P for overall < 0.001, P for nonlinear < 0.05, Fig. 3F–I), whereas no significant nonlinear relationships were observed between SCR, SIRI, DBP, LA or HDL-C and the risk of CHD in DM patients (P for overall < 0.001, P for for nonlinear > 0.05, Fig. 3A–E). The risk of CHD in DM patients increased rapidly when meeting any of the following thresholds: age > 58, SIRI > 1.7, SCR > 62.18 μmol/L, DBP > 70.10 mm Hg, LDL-C > 2.38 mmol/L, LA > 4.31 mmol/L, Glu > 7.48 mmol/L, TyG > 9.04 or HDL-C < 1.05 mmol/L. The results of the risk threshold analysis indicated that the ORs exceeded 1 when any one of the following conditions was met: age > 62, SIRI > 1.7, SCR > 62.18 μmol/L, DBP > 70.10 mm Hg, LDL-C > 2.91 mmol/L, LA > 4.31 mmol/L, Glu > 7.48 mmol/L, TyG > 9.81 or HDL-C < 1.05 mmol/L.

Fig. 3
Fig. 3
Full size image

The RCS illustrates the associations between different continuous variables and the risk of CHD in DM patients. The specific continuous variables included SCR (A), SIRI (B), LA (C), DBP (D), HDL-C (E), LDL-C (F), age (G), Glu (H) and TyG (I). Each variable’s overall significance and nonlinearity of the associations with the risk of CHD in DM patients were measured by p-values. p < 0.05 was defined as statistically significant or indicating a nonlinear relationship. The solid gold lines represent the odds ratio of CHD, and the blue region corresponds to the 95% CI. The dashed black line indicates the reference value. The red solid line represents the risk threshold or inflection point of the variables. DBP: diastolic blood pressure; SCR: serum creatinine; HDL-C: high density lipoprotein cholesterol; LDL-C: low density lipoprotein cholesterol; LA: lactic acid; Glu: Glucose; SIRI: systemic inflammation response index; TyG: triglyceride-glucose index; 95% CI: 95% confidence interval

Model development and performance comparison in internal validation

We carried out 10-fold CV to construct 8 ML models with optimal hyperparameters (Table S3) using the 11 selected factors to predict the risk of CHD in DM patients (Fig. S2A-H). In the test set, the LGBM model performed the best in terms of predictive performance with an ROC-AUC of 0.87 ± 0.02, a brier value of 0.12 ± 0.03 and a PR-AUC of 0.91 ± 0.01 (Fig. 4A–C). Compared with other seven ML models, the LGBM model had the highest accuracy (0.78±0.02), PPV (0.83±0.02), specificity (0.68±0.05), NPV (0.72±0.01) and F1 score (0.83±0.02) (Table S4). The Delong test revealed that the LGBM model achieved a higher AUC than the other models did (p < 0.05, p < 0.001)(Table S5). Compared to the other models, the LGBM model presented significantly higher NRI and IDI values (p < 0.001, Table S6), indicating its excellent ability to classify and distinguish. DCA indicated that the LGBM model exhibited superior clinical net benefit throughout the entire threshold range of 0.1 to 0.9 (Fig. 4D). The LGBM model exhibited superior performance in 10-fold cross-validation, consequently being designated as the optimal predictive model for predicting CHD risk in DM patients.

Fig. 4
Fig. 4
Full size image

Performance of the ML models in predicting the risk of CHD in DM patients in the internal validation and external validation. A. the internal validation test set (10-fold CV) was conducted via eight ML algorithms, and ROC analysis was performed on the results. The different colors of the lines represent the different ML models and are presented as the AUC±SD. B. Calibration curve of eight ML models demonstrating the extent to which the predicted probabilities are in accordance with the actual probabilities. The 45° black dashed line represents the ideal calibration. The different colors of the lines represent the different ML models and are presented as brier±SD. C. the PR curve was used to observe the changes in precision and recall of the ML models at different thresholds. The different colors of the lines represent the different ML models and are presented as the AUC±SD. D. DCA of eight ML models. The different colors of the lines represent the net benefit of the ML models within a range of threshold probabilities. The black line represents the “treat-all” strategy. The horizontal dashed line represents the “treat-none” strategy. E-H. the ROC curve, PR curve, calibration curve and DCA of the LGBM model in the external validation. Abbreviations: ROC: receiver operating characteristic; PR: precision-recall; DCA: decision curve analysis

Sensitivity analysis for predicting different classifications of CHD in internal validation

We employed sensitivity analysis to explore the predictive performance of various ML methods under different CHD classifications. Owing to the limited sample size of sudden cardiogenic death cases (n = 25), no sensitivity analysis was conducted. The LGBM model performed well in terms of the risk prediction for the other four types of CHD (Table S7). Compared with seven other ML prediction models, the LGBM model achieved superior performance in each classification prediction of CHD. In addition, compared with ability to predict total CHD, the LGBM model demonstrated consistent performance in predicting CHD classifications without significant fluctuations.

External validation of the predictive performance

A total of 7122 samples from Yan’an Hospital of Kunming City were used for external validation of the LGBM model with an ROC-AUC of 0.85 (95% CI: 0.82–0.87), a brier value of 0.13 (95% CI: 0.11–0.14) and a PR-AUC of 0.90 (95% CI: 0.89–0.91)(Fig. 4E–H). The other performance evaluation indicators of the LGBM model were as follows: accuracy = 0.75, recall = 0.83, F1 = 0.8, specificity = 0.71, PPR = 0.77 and NPR = 0.78. In summary, the LGBM model exhibited consistently robust predictive performance in external validation settings.

Model explanation

The SHAP summary dot plot offered a graphical illustration of how each feature impacted the model predictions, highlighting both the direction and extent of their influence (Fig. 5A). Patients with DM who had HTN or HLD, along with a greater SIRI, TyG and age, elevated Glu, LDL-C, DBP and SCR level and reduced levels of HDL-C, were at increased risk of developing CHD. In the SHAP summary bar chart (Fig. 5B), features were ranked by their importance to the model, based on mean SHAP values in descending order. The top five features influencing the prediction model were SIRI, HTN, HLD, Glu and LDL-C. In addition, the SHAP waterfall chart visualized the impact of each feature on the model’s CHD prediction for the cohort’s first patient (Fig. 6C). An HDL-C of 0.84 mmol/L and age of 72 made positive contributions of +0.412 and + 0.215 to the prediction results, respectively, whereas HLD (No), a SIRI of 0.89, HLD (No), an LDL-C of 2.84 mmol/L and an SCR of 64 μmol/L made negative contributions of −0.299, −0.216, −0.182, −0.152, and −0.109 to the prediction results, respectively. Other features, such as TyG, LA, DBP and Glu also had a certain degree of influence. This graphical depiction enhances our understanding of the model’s decision-making process. In addition, the SHAP method has inherent limitations: it only explains our model’s prediction logic at the correlational level, cannot elucidate diabetes-related CHD mechanisms or causal feature-CHD links, and cannot be extrapolated to causal pathological mechanisms.

Fig. 5
Fig. 5
Full size image

The importance of the SHapley additive exPlanation (SHAP) method based on the LGBM model. A. the SHAP summary dot plot shows the importance of each feature in the model and its contribution to the prediction of CHD. Each dot represents a patient’s SHAP value for a given feature, with blue indicating higher SHAP values and orange indicating lower SHAP values. Dots are stacked vertically to show density. B. SHAP summary bar chart. This chart indicates the contribution of each feature to the model by using the average SHAP values, which are arranged from highest values to lowest values. C. SHAP waterfall plot shows the contribution of each feature to the prediction of CHD in the first patient using the LGBM model. The red bars indicate features that contribute positively to the prediction result, such as HDL-C (0.84 mmol/L, +0.412), whereas the blue bars indicate negative contributions, such as the SIRI (0.89, −0.216). The overall contribution is −0.337, with a baseline contribution of 0

Fig. 6
Fig. 6
Full size image

The web-based calculator for predicting CHD in DM patients using this model. By simply inputting the clinical information of the patients: LA, HDL-C, SCR, TyG, SIRI, Glu, LDL-C, HLD, HTN, DBP and age, it is possible to predict the risk of CHD in DM patients

Online application of CHD prediction

As depicted in Fig. 6, the ultimate predictive model was incorporated into a web-based platform designed for clinical applications using the Django framework. By entering the real values of the 11 variable features, the platform can automatically estimate the incidence of CHD in patients with DM. This web-based tool is accessible online via the following link: https://www.xsmartanalysis.com/model/list/predict/model/html?mid=31986%26symbol=3Ox1773Gh6468Va546kZ. All clinical data input by users are calculated and processed only in the user’s local browser, with no data uploaded to the server, no patient information stored or recorded in the background. Meanwhile, We will carry out targeted adaptation and integration for the hospital’s Electronic Medical Record (EMR) system to achieve its application in real clinical scenarios. In addition, our web-based prediction platform is only a proof-of-concept prototype, with substantial steps remaining before clinical deployment, to be addressed in our subsequent clinical translation work.

Discussion

Compared with non-DM patients with CHD, patients with DM complicated with CHD have a greater risk of myocardial infarction, cerebrovascular accidents, heart failure, etc., along with a higher incidence and mortality rate [40,41,42]. However, some patients with silent CHD may only show mild symptoms or no symptoms at all, which may lead to acute myocardial infarction when the disease suddenly occurs [43, 44]. Therefore, establishing a prediction model for CHD in DM patients can help patients focus on relevant indicators during their annual routine check-ups to achieve early management, diagnosis and treatment of the disease. This study aims to develop and validate an interpretable LGBM-based machine learning model for CHD risk prediction in DM patients from dual-center cohorts, using routine laboratory indicators and novel derived inflammation and lipid metabolism markers as candidates, with multi-method variable screening, SHAP-driven interpretability analysis, and a web-based tool for clinical translation.

Several studies have been published on forecasting CHD risk in DM patients. Xu et al. [24]. constructed a logistic regression model to predict the occurrence of CHD among DM patients, utilizing data from 1152 individuals in Henan Province, China. The model achieved an ROC-AUC score of 0.753. Seven key factors associated with CHD in DM patients were identified: sex, diabetes duration, LDL-C, SCR, HDL-C, HTN and heart rate. In a separate study, Xiao et al. [45]. explored the risk factors for CHD in patients with type 2 DM using the data from 560 participants. This study developed a predictive model that also yielded an ROC-AUC of 0.753. Their findings indicated that non-HDLC, apoA1, Lp(a) and AIP are potential risk factors. Fan et al. [25]. construct an RF ML model to predict CHD in type 2 DM patients, based on data from 1273 individuals. The conducted model had an ROC-AUC of 0.71. The top eight features contributing to the model’s predictions included age, LDL-C, DM duration, TC, heart rate, DBP, PLT and HTN. In addition, Chu et al. [46]. developed a deep neural network model for cardiovascular disease (CVD) risk prediction in a cohort of 834 T2DM patients, achieving a remarkable ROC-AUC of 0.91. The five most significant factors influencing CVD risk according to the DNN model were BMI, anxiety levels, depression status, TC and SBP.

Compared with previous studies, this study has made significant improvements and innovations in terms of research subjects and research methods, as follows: (1) With sufficient sample size support, a total of 12,124 samples were included in both internal and external validation. (2) In the combined analysis of traditional clinical indicators and new clinical indicators, 32 traditional laboratory indicators, 6 novel inflammation indicators and 10 novel lipid metabolic indicators were included in this study. (3) This project employed a combination of traditional methods and multiple ML methods to screen for key predictive factors and established a precise predictive model. Through univariate and multivariate logistic regression analyses, followed by LASSO, Brotua and XGBoos RFE dimension reduction, 11 factors related to the risk of CHD in DM patients were ultimately selected. We used eight kinds of ML algorithms (10-fold CV) to build different ML models for the risk of CHD in DM patients and compared their predictive performance of them via Delong, NRI and IDI tests. Finally, we adopted the LGBM algorithm to build a model to predict the risk of CHD with the highest ROC-AUC value of 0.87. The external validation of the LGBM model still revealed good predictive performance with an ROC-AUC value of 0.85. (4) We further employed sensitivity analysis to explore the predictive performance of various ML methods under different CHD classifications. The results showed that the LGBM model had stable predictive performance for four categories of CHD. (5) Although providing the both global and local explanations for each feature in the LGBM model, the SHAP method elucidated how the model functions, detailing the use of personalized input data to predict the incidence rate of CHD in individual DM patients. (6) The predictive models in previous studies often failed to present a clear incidence rate. This lack of intuitiveness may cause clinicians to hesitate when using these models. Taking these issues into consideration, we integrated the prediction model into a user-friendly online platform via the Django framework, making it accessible to both doctors and patients.

It is worth noting that several features in our CHD model have been identified as having significant associations with CHD in prior studies, thereby providing a solid basis for their inclusion in our model. For example, Dong et al. [47]. reported that the TyG index, the SII index, and the SIRI index serve as independent risk factors for CHD in individuals with nonalcoholic fatty liver disease (NAFLD). These indices are strongly associated with both the prediction and severity of CHD. Song Vogue et al. [48]. reported that the impact of fasting serum glucose levels on the risk of CHD is greater in women than in men, particularly starting from prediabetic ranges (≥110 mg/dL). Turin et al. [49]. demonstrated that among Japanese men and women of all ages, the lifetime risk of CHD increases for those with HTN. Specifically, the risk is 14.12% for men with normal blood pressure, 26.95% for men with HTN, 6.21% for women with normal blood pressure, and 14.85% for women with HTN. Hu et al. [50]. reported that a positive correlation was observed between the LDL-C/HDL-C ratio and CHD, indicating that the LDL-C/HDL-C ratio could potentially serve as a marker for CHD. We incorporated these reported CHD risk factors into a new model for predicting CHD risk in DM patients and included several new features, such as SCR, LA, DBP and HLD. All included predictors have established biological relevance to CHD in T2DM patients: age, HTN and DBP drive vascular endothelial injury and atherosclerotic progression; hyperlipidemia, LDL-C and dysfunctional HDL-C are core mediators of atherogenesis; elevated Glu promotes diabetic vascular damage via glycation and oxidative stress; SCR and LA diameter reflect target organ damage (renal impairment and myocardial remodeling) linked to elevated CHD risk; while SIRI and TyG index quantify the two core pathological pathways of T2DM-related CHD: chronic low-grade inflammation and insulin resistance [51,52,53].

This study has important clinical implications for CHD prevention in DM patients. Our dual-center validated, interpretable machine learning model uses only routine clinical and easily derived markers, with full accessibility across care settings and no extra costs. SHAP analysis addresses the model “black box” to improve clinical uptake and guide personalized intervention via modifiable risk factor identification. The model enables accurate CHD risk stratification to target early intensive care for high-risk DM patients and avoid over-medicalization in low-risk groups, while the accompanying web-based tool supports rapid point-of-care assessment with strong clinical translation potential.

AI has emerged as a superior tool to conventional methods for cardiovascular risk prediction, and our study confirms prior findings while providing critical incremental value. Consistent with Saurabhi Samant et al. [54]. we verified machine learning (ML) models using routine clinical markers effectively predict coronary heart disease (CHD) risk in diabetes mellitus (DM) patients, with enhanced generalizability via dual-center external validation. In line with Shayan Shojaeiet al. [55]., we confirmed the independent predictive value of the TyG index and SIRI in this population, and further quantified their feature importance via SHAP analysis. Corroborating Shayan Shojaei et al. [56]. on the critical role of explainable AI in clinical translation, we addressed ML’s “black box” limitation with SHAP to improve clinical acceptability. Finally, extending the work of Ahsanullah Niazai et al. [57]. on ML-based risk stratification in high-risk cardiovascular populations, we developed an open-access web-based tool to facilitate real-world clinical application of our model.

We recognize several limitations in this study. First, the retrospective nature of this study limited the documentation of CAG indications, potentially biasing the patient cohort and restricting the model’s applicability to diabetic patients with similar clinical profiles. Second, the results of this study are derived from clinical data from China, and further investigation is necessary to confirm the applicability of the conclusions to other populations. Third, sudden cardiogenic death is the most severe type of CHD. Owing to the small sample size, no sensitivity analysis was conducted for this type of CHD. In addition, atrial fibrillation (AF) and chronic kidney disease (CKD) are core CHD risk factors in T2DM patients, but our laboratory-based CHD prediction model study did not systematically collect verified AF and CKD in this retrospective cohort. In subsequent studies, we will continue to collect such samples to complete the entire analysis. Fourth, all CHD cases and subtypes included in this study were confirmed by coronary angiography, which ensured the accuracy of the outcome labels, but also led to the enrolled population being limited to diabetic patients with clinical indications for coronary angiography. Therefore, extreme caution must be exercised when applying the model to other populations, and it cannot be directly generalized to the entire diabetic population. Fifth, this study did not incorporate imaging data. Both traditional radiomics features and those based on deep learning have the potential to increase the model’s accuracy and clinical applicability. We will include imaging data in model development in future research. Final, for future model refinement, we will apply well-validated probability calibration techniques, including Platt scaling and isotonic regression, to optimize the model’s predicted risk output for CHD subtypes in diabetic patients, improve the consistency between predicted and actual disease risks, and further enhance the clinical utility of the model.

Conclusions

In summary, we developed an interpretable ML model to predict CHD risk in DM patients. The LGBM-based model demonstrated strong predictive performance in internal and external validations. Future prospective randomized controlled trials are required to evaluate the efficacy of the final prediction model in reducing CHD incidence among patients with DM.