Application of an interpretable machine learning model based on optimal feature selection for predicting triple-vessel coronary disease: a multicenter retrospective study

Study population

This retrospective observational study enrolled patients who were hospitalized and underwent coronary angiography between January 1 and December 31, 2024, at two medical centers in China: Enshi Tujia and Miao Autonomous Prefecture Central Hospital and Minda Hospital of Hubei Minzu University. Eligible participants were those diagnosed for the first time with TVD, defined as ≥50% stenosis in all three major coronary arteries—the left anterior descending, left circumflex, and right coronary arteries—based on coronary angiographic findings (Li et al., 2024). Patients were excluded if they had known structural or functional cardiac disorders, malignancies, or incomplete key biochemical data, particularly levels of high-sensitivity hsCRP and the full lipid profile (total cholesterol, triglycerides, HDL-C, LDL-C). The study was conducted in accordance with the Declaration of Helsinki and approved by the institutional ethics committees of Enshi Tujia and Miao Autonomous Prefecture Central Hospital (Approval No. 202500201). As a retrospective study, the requirement for informed consent was waived, and all patient data were anonymized to protect privacy. A total of 2,911 patients meeting the inclusion criteria were ultimately included analyzed (Fig. 1). Patient data were accessed and extracted from the hospital information systems between March 1 and March 15, 2025.

Data collection

Clinical and laboratory data were systematically collected from electronic medical records. Demographic information included age and sex. Medical history included hypertension, type 2 diabetes mellitus (T2DM), hyperlipidemia, hyperuricemia, atrial fibrillation, and stroke. Lifestyle data included smoking status. Vital signs included diastolic blood pressure (DBP) and heart rate. Laboratory parameters encompassed: leukocytes, lymphocytes, monocytes, hemoglobin, platelets, hsCRP, aspartate aminotransferase (AST), total bilirubin, direct bilirubin, albumin, total cholesterol, triglycerides, LDL-C, HDL-C, apolipoprotein A1 (LP_A1), apolipoprotein B (LP_B), serum creatinine, fasting glucose, uric acid, and serum potassium. In addition, the RCII was calculated to represent the composite burden of lipid metabolism and inflammation. RC was estimated as: RC = total cholesterol − HDL-C − LDL-C, and RCII was defined as: RCII = (RC × hsCRP)/10 (Chen et al., 2025).

Feature selection

To enhance model performance and reduce noise from redundant variables, two complementary approaches were used for feature selection. First, the least absolute shrinkage and selection operator (LASSO) regression was applied to identify predictive variables. Subsequently, univariate logistic regression was performed, and variables with a significance level of p < 0.05 were further analyzed using multivariate logistic regression to identify independent predictors. Finally, variables selected from both LASSO and multivariate models were compared and integrated to define a stable feature set, which was then used for machine learning model development. Only overlapping predictors were retained to ensure the robustness and interpretability of the ML models.

Model construction, evaluation, and interpretability

Based on the selected features, six machine learning (ML) algorithms were employed to develop prediction models: multilayer perceptron (MLP), decision tree (DT), random forest (RF), k-nearest neighbor (KNN), light gradient boosting machine (LightGBM), and extreme gradient boosting (XGBoost). All data were randomly divided into training and testing sets in a 7:3 ratio. Hyperparameter tuning for all machine learning models was performed within the training set using Bayesian optimization. For each model, relevant hyperparameter search spaces were defined, and five-fold cross-validation was applied during tuning. Model performance metrics were collected at each iteration to identify the optimal hyperparameter combination. Model performance was assessed using the testing set, based on the area under the receiver operating characteristic curve (AUC), overall accuracy, sensitivity, specificity, F1 score, and positive and negative predictive values. The best-performing model was selected for interpretability analysis. To interpret the model outputs, the SHapley Additive exPlanations (SHAP) framework was utilized. SHAP summary plots (scatter and bar plots) were used to evaluate the global contribution and directional effect of each feature on model predictions. SHAP dependence plots were employed to visualize how SHAP values change with feature values, reflecting their marginal effects. Moreover, SHAP interaction plots were generated to investigate potential synergistic effects between RCII and other key variables, capturing complex nonlinear relationships affecting TVD risk. For reproducibility, SHAP analyses were performed with 1,000 background samples and 100 iterations per sample, and interaction values were normalized. All ML modeling and SHAP analyses were conducted using Python (version 3.9) with scikit-learn (version 1.2.2), XGBoost (version 1.7.6), LightGBM (version 3.3.5), and shap (version 0.47.2) packages.

Statistical analysis

All statistical analyses were conducted using Python (version 3.9) and R (version 4.3.2). Continuous variables with a normal distribution were presented as mean ± standard deviation (SD) and compared using the independent sample t-test. Non-normally distributed variables were expressed as medians with interquartile ranges (IQR) and compared using the Mann–Whitney U test. Categorical variables were summarized as counts and percentages and compared using the chi-squared (χ²) test or Fisher’s exact test where appropriate. All statistical tests were two-tailed, and a p-value <0.05 was considered statistically significant.

Baseline characteristics

A comparative analysis of clinical characteristics between the TVD and non-TVD groups is presented in Table 1. The proportion of males was significantly higher in the TVD group. Additionally, the prevalence of hypertension, T2DM, hyperuricemia, and a history of smoking was markedly elevated among patients with TVD. Patients in the TVD group were older and had slightly higher diastolic blood pressure. In terms of inflammatory and hematological parameters, levels of leukocytes and monocytes were significantly increased in the TVD group. Regarding liver function, AST was notably higher in the TVD group. Among metabolic indices, fasting glucose, serum creatinine, uric acid, and potassium levels were all significantly elevated in the TVD group compared to the non-TVD group. Importantly, the RCII was significantly higher in patients with TVD. No significant differences were observed between the two groups in terms of heart rate, hyperlipidemia, atrial fibrillation, stroke history, lymphocyte count, hemoglobin, platelet count, triglycerides, apolipoprotein A1, apolipoprotein B, total bilirubin, direct bilirubin, or serum albumin.

Feature selection

To optimize model stability and reduce multicollinearity, LASSO regression was first employed. This regularization technique enabled efficient variable shrinkage and selection, ultimately retaining 16 candidate predictors: gender, hypertension, T2DM, hyperuricemia, smoking history, age, leukocytes, lymphocytes, monocytes, hemoglobin, platelets, AST, RCII, LP_B, creatinine, and glucose. In parallel, a traditional statistical approach involving univariate screening followed by multivariate logistic regression identified six independent predictors: Gender, Age, DBP, AST, RCII, and Potassium (Table 2). A comparison of both methods (Fig. 2) revealed four overlapping variables—Gender, Age, AST, and RCII—demonstrating consistent predictive performance across both techniques. These four features were thus selected for ML modeling to construct a more robust and interpretable TVD prediction framework. This conservative strategy was adopted to enhance model stability and interpretability by focusing on predictors consistently identified by both LASSO and logistic regression.

Model comparison and interpretability

The four selected features were used to train six machine learning models: MLP, DT, RF, KNN, LightGBM, and XGBoost. Models were evaluated on both training and test datasets. As shown in Table 3, the RF model achieved the highest performance in the training set, demonstrating superior accuracy and AUC. However, it demonstrated substantial overfitting, as evidenced by performance degradation in the test set. In contrast, the MLP model yielded the best generalization performance on the test set, followed closely by LightGBM. As illustrated in Figs. 3 and 4, the MLP model demonstrated both high stability and robust predictive capacity, and was therefore selected for subsequent interpretability analyses using the SHAP framework.

To elucidate the inner workings of the MLP model and the contribution of each input variable, SHAP analysis was performed. As shown in Fig. 5, SHAP summary bar and scatter plots revealed the relative importance of the four input features in predicting TVD. Notably, RCII ranked third among all features, with its SHAP value showing a strong positive correlation with its actual value. This indicates that elevated RCII significantly increases the predicted risk of TVD. Furthermore, older age, elevated AST levels (indicative of hepatic dysfunction or metabolic stress), and male sex were also associated with higher predicted TVD risk.

Figure 6 presents SHAP dependence plots, illustrating the quantitative relationships between feature values and model outputs. Both age and RCII exhibited a monotonic increase in SHAP values, particularly when RCII exceeded 0.2, reinforcing its role as a high-risk marker for advanced coronary atherosclerosis. The influence of AST and gender on SHAP values was also evident, with male patients demonstrating consistently higher risk contributions. Finally, Fig. 7 displays the SHAP-based interaction heatmap, highlighting synergistic effects among the four features. Significant interactions were observed between RCII and age, AST, and gender, suggesting that the predictive value of RCII is enhanced when combined with other clinical risk factors.

Limitations

Despite the promising findings, several limitations should be acknowledged. First, the retrospective design and reliance on previously recorded clinical data may introduce potential biases and residual confounding, including the influence of unmeasured factors such as medication use. Second, although the study included a multicenter cohort, it was still regionally restricted, and external validation in independent populations is warranted. Third, our feature selection strategy prioritized only overlapping predictors from LASSO and logistic regression, which enhances interpretability but may have excluded some clinically relevant variables, thereby potentially affecting model robustness. Furthermore, our findings are specific to patients with triple-vessel coronary disease, and the applicability of RCII to other forms or stages of coronary artery disease remains to be determined. Taken together, these considerations highlight the need for further studies to explore these aspects and to confirm the predictive utility of RCII across broader and more diverse CAD populations.