This study enrolled patients who had MHD treatment for at least 3 months and were aged 18 years or older at the hemodialysis unit of the Affiliated Hospital of Southwest Medical University between January 1, 2020, and January 1, 2022. The exclusion criteria encompassed the following conditions: a history of peritoneal dialysis or renal transplantation; underlying malignancy; severe infection; severe hepatic insufficiency; or active tuberculosis. Additional exclusions included multiple myeloma, bone tumors, and other disorders affecting calcium and phosphorus metabolism in vivo; parathyroidectomy; and an active phase of autoimmune disease requiring high-dose glucocorticoids or immunosuppressive therapy. This study was approved by the Ethics Committee of the Affiliated Hospital of Southwest Medical University (Approval No.: KY2024300), as an exempt study with a waiver of informed consent, permitting a retrospective review of medical records.

The study employed a comprehensive dataset comprising four key components: demographic characteristics (9 variables), comorbidities (6 variables), medication history (9 variables), and baseline laboratory values (63 variables). All predictor variables were extracted from electronic medical records, with specific details provided in Supplementary Table 1.

The endpoints of this study were the first occurrence or recurrence of CVE and all-cause mortality. A broad definition of CVE was adopted (24), which included stroke (including transient ischemic attacks), severe cardiac arrhythmias (such as ventricular fibrillation, ventricular tachycardia, atrial fibrillation, atrial flutter, severe bradycardia, and heart block), acute myocardial infarction, unstable angina pectoris, coronary artery revascularization, development of various types of heart failure (HF) requiring hospitalization, sudden cardiac death, and peripheral vascular disease necessitating intervention or amputation. The follow-up period concluded on June 30, 2024.

Data were stored and managed using Excel 2016, while statistical analyses were conducted using the R language (version 4.4.1). Variables with missing rates exceeding 30% were excluded from the analysis. For variables with missing rates ≤ 30%, the following imputation methods were employed: median imputation, mean imputation, or mode imputation, depending on the variable's trend. For count data, random interpolation was performed based on the proportion of available positive and negative data. Normality tests were conducted on continuous variables. Normally distributed variables were presented as mean ± standard deviation, while non-normally distributed variables were presented as median (P25, P75). Categorical variables were expressed as proportions. Comparisons of variable distributions between groups were performed using ANOVA or the Kruskal–Wallis H test, as appropriate. All statistical tests were two-sided, with a significance level set at P < 0.05.

The Cox regression model was constructed using the “coxph” function in R. Initially, univariate Cox regression analyses were conducted to identify potential risk factors associated with CVE and all-cause mortality among MHD patients, with significance set at P < 0.05. Variables that were significant in the univariate analyses were subsequently included in the multivariate Cox regression model to determine independent predictors. The stability of the model was evaluated using 5-fold cross-validation. For each fold, risk scores were calculated, and the validation sets along with their predicted outcomes were integrated. The optimal cutoff value for risk stratification was determined using the ‘surv_cutpoint‘ function. Kaplan-Meier survival curves were generated using the ‘ggsurvplot‘ function, and a nomogram was created with the ‘nomogram‘ function. The predictive performance of the model was assessed by plotting time-dependent receiver operating characteristic (ROC) curves, which integrate specificity and sensitivity. A model with an area under the curve (AUC) greater than 0.70 was considered to have good discrimination.

This step aims to identify a subset of features from the original dataset that maximizes the outcome benefit, thereby reducing model complexity and enhancing generalizability. For feature selection, we employed the Sequential Feature Selector (SFS) method in conjunction with a Random Forest regressor. SFS is a greedy algorithm that iteratively adds or removes features to optimize model performance. The model-building process began with an empty feature set, and features were added incrementally in steps of 2. This iterative process continued until either a predefined number of features was reached or further improvements in model performance plateaued.

ML models were developed using Python software (version 3.10.0). Six classical ML algorithms were employed to predict the risk of CVE and all-cause mortality in MHD patients. These algorithms included logistic regression (LR), support vector machine (SVM), random forest (RF), decision tree (DT), extreme gradient boosting (XGBoost), and Naive Bayesian model (NBM). The primary functions of these algorithms were defined, and the models were iteratively trained with varying numbers of features to generate corresponding prediction results and performance reports. Five-fold cross-validation was used for internal validation, and the average values of these validations were accepted as the final results to mitigate the risk of overfitting.

As our ML models were binary classifiers, their performance was evaluated using several key metrics: accuracy, recall, precision, F1-score, and AUC. These metrics were derived from the four possible outcomes of binary classification: true positive (TP), true negative (TN), false positive (FP), and false negative (FN) (Table 1). Accuracy measures the proportion of correct predictions (both TP and TN) among all subjects. Recall (also known as sensitivity or the “TP rate”) represents the proportion of actual non-surviving patients that are correctly identified as non-surviving by the classifier. Precision indicates the proportion of TP results among all positive predictions, reflecting the classifier's ability to avoid FP results. F1-score is the harmonic mean of precision and recall, providing a balanced measure of the two. Specificity (also known as the “TN rate”) measures the proportion of actual surviving patients that are correctly predicted to survive. AUC was computed by plotting sensitivity against 1-specificity across all possible cutoff points. It serves as an overall measure of the model's discrimination ability, with higher AUC values indicating better performance.

To enhance the interpretability of ML models, Shapley Additive exPlanations (SHAP) were employed. SHAP leverages the concept of SHAP values, which are grounded in game theory, to quantify the importance of each feature in the model. By calculating the SHAP value for each feature, it assesses the contribution of that feature to the prediction outcome. This approach generates both visual and quantitative interpretations, enabling users to understand the decision-making process of the ML model more transparently and thereby enhancing the model's credibility.

The flowchart of patient selection for this study is presented in Figure 1. A total of 373 patients undergoing MHD were identified at our hospital between January 1, 2020, and January 1, 2022. Patients with a history of parathyroidectomy (n = 1), peritoneal dialysis (n = 13), renal transplantation (n = 15), or malignancy (n = 16) were excluded. Ultimately, 328 patients were included in the study and followed up until June 30, 2024. During this period, 26 patients were lost to follow-up, and 27 were transferred to other dialysis centers. Consequently, 275 participants were included in the final analysis. The median age of the participants was 56.0 years [interquartile range (IQR) 48.0–67.0], and the median dialysis vintage was 64.0 months (IQR 41.0–92.0). The cohort comprised 62.2% males. The primary underlying renal diseases were chronic glomerulonephritis (39.3%), diabetic nephropathy (28.7%), and hypertensive nephropathy (18.9%). The remaining 13.1% of patients had other renal diseases, including polycystic kidney disease, obstructive nephropathy, gouty nephropathy, etc. The overall rate of missing data was 0.51%, and these missing values were imputed using the median method, as detailed in Supplementary Table 1.

The median follow-up period was 50.0 months (IQR 34.5–53.0). During this period, a total of 119 patients (43.3%) experienced CVE. Among these patients, 80 were men (incidence rate of 67.2%) and 39 were women (incidence rate of 32.8%). The specific types of CVE included: HF in 49 cases (41.18%), cerebral hemorrhage in 27 cases (22.69%), cerebral infarction in 13 cases (10.92%), unstable angina pectoris in 11 cases (9.24%), cardiac arrhythmia in 9 cases (7.56%), myocardial infarction in 4 cases (3.36%), peripheral vascular disease in 4 cases (3.36%), and transient cerebral ischemic attack in 2 cases (1.68%). Baseline characteristics were compared between the CVE group and the non-CVE group, with detailed data presented in Supplementary Table 2.

During the follow-up period, a total of 75 patients (27.3%) died. Of these, 52 (69.3%) were male and 23 (30.7%) were female. CVE were the cause of death in 42 patients (56%). The remaining 33 patients (44%) died from non-CVE causes, including respiratory failure, sepsis, poisoning, gastrointestinal hemorrhage, uremic encephalopathy, suicide, and an unknown cause. Baseline characteristics were compared between the death and survival groups, with detailed data presented in Supplementary Table 3.

In predicting CVE, multivariate Cox regression analysis identified several independent risk factors: a history of CVD [hazard ratio (HR): 1.984, 95% confidence interval (CI): 1.282–3.070], creatine kinase isoenzyme (CK-MB) (HR: 1.098, 95% CI: 1.001–1.204), red cell distribution width-coefficient of variation (RDW-CV) (HR: 1.007, 95% CI: 1.001–1.012), and mean corpuscular hemoglobin (MCH) (HR: 0.935, 95% CI: 0.875–0.998) (P < 0.05 for all). Based on a cutoff value of 0.61, subjects were stratified into high-risk (cutoff > 0.61) and low-risk (cutoff ≤ 0.61) groups for CVE, comprising 180 and 95 cases, respectively. The Kaplan-Meier survival plot demonstrated a significant difference in survival between the high-risk and low-risk groups (P < 0.001) (Figure 2A). The time-dependent ROC plot showed AUC values for predicting CVE at 1, 2, 3, and 4 years were 0.681, 0.709, 0.728, and 0.743, respectively (Figure 2B).

In predicting all-cause mortality, multivariate Cox regression analysis identified several independent risk factors: age (HR: 1.030, 95% CI: 1.003–1.058), direct bilirubin (DBIL) (HR: 1.235, 95% CI: 1.003–1.520), myohemoglobin (MYO) (HR: 1.0023, 95% CI: 1.000–1.004), dialysis vintage (HR: 0.95, 95% CI: 0.935–0.966), and the use of roxadustat (HR: 0.395, 95% CI: 0.193–0.810). Based on a cutoff value of 1.87, subjects were stratified into high-risk (cutoff > 1.87) and low-risk (cutoff ≤ 1.87) groups for all-cause mortality, comprising 61 and 214 cases, respectively. The Kaplan-Meier survival plot demonstrated a significant difference in survival between the high-risk and low-risk groups (P < 0.001) (Figure 2C). The time-dependent ROC plot showed AUC values for predicting all-cause mortality at 1, 2, 3, and 4 years were 0.902, 0.927, 0.864, and 0.858, respectively (Figure 2D). Finally, the nomograms for predicting CVE (Figure 3A) and all-cause mortality (Figure 3B) were constructed based on the selected independent factors. Each variable was first scored on its corresponding subscale. Subsequently, the scores of all variables were summed to obtain a total score, which corresponded to the risk of CVE or all-cause mortality occurrence.

In predicting CVE, the average overall AUC of the six ML models was 0.788. Among these models, RF achieved the highest AUC value of 0.757, followed by XGBoost, SVM, LR, DT, and NBM (Figure 4A). The time-dependent ROC curves of these models demonstrate that prediction performance improves gradually as the prediction time extends (Figure 5). Detailed performance evaluations of the six ML models are presented in Table 2. Notably, both RF and XGBoost required the fewest predictors to achieve optimal prediction performance. Although RF had the highest AUC value, XGBoost outperformed it in terms of accuracy, recall, precision, and F1 score. Therefore, XGBoost was identified as the optimal prediction model for CVE. The SHAP plot ranked the variables according to their contribution to the XGBoost model's output. The most important feature variable was the history of CVD, followed by total iron-binding capacity (TIBC), body mass index (BMI), red blood cell (RBC) count, MCH, and serum magnesium (Mg) (Figure 4C).

In predicting all-cause mortality, the average overall AUC of the six ML models was 0.837. Among these models, RF and XGBoost achieved the highest AUC value of 0.828, followed by SVM, LR, NBM, and DT (Figure 4B). The time-dependent ROC curves indicate that the models exhibited the greatest superiority in predicting all-cause mortality at 2 years. However, as the prediction time extended beyond this period, the prediction performance decreased somewhat (Figure 6). As shown in Table 3, RF required the fewest predictors and demonstrated the best performance in terms of accuracy, precision, and AUC value. Therefore, RF was determined to be the best prediction model for all-cause mortality. The SHAP plot reveals that dialysis vintage is the most significant feature, followed by post-dialysis β2-microglobulin (β2-MG), B-Carboxy-Terminal Peptide of Type I Collagen (β-CTX), total bilirubin (TBA), lymphocytes (LYM), lactate dehydrogenase (LDH), mean corpuscular hemoglobin concentration (MCHC), and the use of roxadustat (Figure 4D).

As illustrated in Tables 4, 5, the XGBoost model demonstrates superior overall predictive performance compared to the Cox regression model in predicting CVE. While the Cox regression model exhibits slightly better performance in the first and second years, it still falls short of XGBoost in predicting the third and fourth years. In predicting all-cause mortality, RF consistently outperforms the Cox regression model both in overall performance and at each individual time point.