This study was reviewed and approved by the Ethics Committee of the participating hospital (Approval No. YJ202570) and was conducted in accordance with the Declaration of Helsinki and relevant institutional requirements. Because this retrospective observational study used routinely collected and de-identified clinical data, the requirement for written informed consent was waived in accordance with the decision of the ethics committee. All data were accessed only by authorized investigators, and strict measures were implemented to protect patient privacy and confidentiality.

This was a single-center model-development and temporal-validation study conducted at a tertiary hospital in Jingzhou, Hubei Province, China. Patients treated from January 2021 through December 2023 constituted the development-period cohort, whereas patients treated from January 2024 through December 2025 constituted the temporal validation cohort. The development-period cohort was used for model training and same-period holdout evaluation, and all fitted models were subsequently applied without refitting or retuning to the later-period cohort. Because the later-period data originated from the same hospital, this cohort represents temporal validation rather than external validation in an independent setting. All observations represented actual clinical records; no synthetic data generation, bootstrap-based sample expansion, artificial perturbation of clinical variables, model-guided outcome assignment, or performance-targeted cohort construction was performed.

Demographic characteristics, comorbidities, vital signs, anthropometric measurements, and laboratory findings were obtained from the hospital electronic medical record system, laboratory information system, and hemodialysis management database. Consecutive adult patients receiving maintenance hemodialysis during the study periods were screened. Patients were eligible if they were aged 18 years or older, had received maintenance hemodialysis for at least 3 months, had available BNP and hs-cTnI measurements during the index clinical assessment, and had sufficient information for outcome definition, feature selection, and model development. Patients receiving temporary dialysis for acute kidney injury, those without a definable primary outcome, those lacking variables required for the final model, and those with duplicate or internally inconsistent records that could not be resolved were excluded. When more than one eligible assessment was available for the same patient during a study period, only the first qualifying assessment was retained. Patient identifiers were cross-checked between study periods to prevent the same patient from contributing observations to both cohorts. After application of the eligibility criteria, 600 patients treated from January 2021 through December 2023 were included in the development-period cohort, and 360 patients treated from January 2024 through December 2025 were included in the temporal validation cohort. Each patient contributed one patient-level analytical record.

Clinical information was extracted according to a prespecified data dictionary, and identical variable definitions, coding procedures, measurement units, and extraction rules were applied to the development and temporal validation cohorts. Predictor values and index cardiac-biomarker measurements were obtained from the same clinical assessment episode. When multiple predictor measurements were available, the result obtained closest to and before, or on the same day as, the index BNP and hs-cTnI assessment was retained. Blood samples were obtained as part of routine clinical care, but the retrospective source systems did not consistently encode whether each specimen was collected before dialysis, during dialysis, or after dialysis, nor the immediately preceding ultrafiltration volume. Sampling time relative to dialysis therefore could not be standardized retrospectively and may have influenced BNP and hs-cTnI concentrations. Body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared. The prognostic nutritional index (PNI) was calculated as serum albumin in g/L plus five times the absolute lymphocyte count in 10^9/L, and the neutrophil-to-lymphocyte ratio (NLR) was calculated as the absolute neutrophil count divided by the absolute lymphocyte count. D-dimer concentrations were natural-log transformed because of their right-skewed distribution. The primary outcome was a study-defined contemporaneous composite elevation of cardiac biomarkers. A patient was classified as having the outcome when BNP was at least 1045.13 pg/mL or hs-cTnI was at least 47.62 ng/L. These thresholds corresponded to the 75th percentiles of the observed BNP and hs-cTnI distributions in the complete development-period cohort and were calculated using the type 8 quantile estimator before the cohort was randomly partitioned. The outcome definition was subsequently applied unchanged to the training set, same-period holdout set, and temporal validation cohort. BNP and hs-cTnI were used exclusively to define the outcome and were excluded from feature selection and model development. Because this was a study-specific contemporaneous biomarker endpoint, these thresholds should not be interpreted independently as diagnostic cutoffs for heart failure or acute myocardial infarction.

The completeness and consistency of the data were assessed before model development. Because no missing values were present in the variables included in the analytical datasets, neither single nor multiple imputation was performed. Data-quality checks were conducted to identify duplicate observations, impossible values, inconsistent measurement units, and data-entry errors. No patient was excluded solely because of an extreme but clinically plausible value, and no Winsorization or model-driven outlier removal was performed. After the outcome had been defined, the development-period cohort was divided once into a training set and a same-period holdout set at a ratio of 7:3 using outcome-stratified random sampling and random seed 123. Because the biomarker percentile thresholds had been estimated from the complete development-period cohort before this split, the holdout set was not fully independent of outcome-definition development and is described as same-period holdout evaluation rather than strict internal validation. Least absolute shrinkage and selection operator (LASSO) logistic regression was conducted exclusively in the training set to select predictors for model development. The penalty parameter lambda was determined through 10-fold cross-validation using the one-standard-error rule (lambda.1se), and predictors with nonzero coefficients at the selected lambda value were retained. The selected predictor set was subsequently fixed and used in all base models and the stacking ensemble, with no further feature selection performed in the holdout or temporal validation cohorts. Spearman correlation coefficients were calculated among the retained continuous predictors, and predictor pairs with an absolute correlation coefficient greater than 0.50 were flagged for review rather than automatically excluded.

Six base models were developed using the predictors retained by LASSO: logistic regression, classification and regression tree (CART), elastic net, random forest, radial support vector machine, and gradient boosting machine. Outcome-stratified 10-fold cross-validation was conducted within the training set, and identical fold assignments were used for all base models. Hyperparameters were selected by maximizing the mean cross-validated area under the receiver operating characteristic curve (ROC-AUC), and out-of-fold predicted probabilities were retained for every patient in the training set. Centering and scaling were incorporated into the modeling pipelines for logistic regression, elastic net, and the radial support vector machine, whereas tree-based models used predictors on their retained analytical scales. Logistic regression was fitted without a tuning grid. For elastic net, alpha values of 0, 0.5, and 1 and 10 logarithmically spaced lambda values ranging from 0.001 to 1 were evaluated. CART evaluated 10 complexity-parameter values ranging from 0.001 to 0.05. Random forest used 600 trees and evaluated mtry values of 3, 5, 8, and 11 and minimum node sizes of 5 and 12 under Gini splitting. For the radial support vector machine, automatic tuning was performed through the caret training interface using a tuning length of six. Gradient boosting evaluated 75, 125, and 200 trees; interaction depths of 1, 2, and 3; shrinkage values of 0.03 and 0.06; and a minimum of 10 observations per terminal node.

The three base models with the highest mean cross-validated ROC-AUC values in the training set were selected for inclusion in the stacking ensemble. Their out-of-fold predicted probabilities were used as inputs to an L1-penalized logistic meta-model. To obtain training-set stacking predictions, each outer fold was predicted using a meta-model fitted on the remaining folds, with the penalty parameter selected through an additional five-fold cross-validation procedure. A final meta-model was subsequently fitted using the complete out-of-fold prediction matrix, with its penalty parameter selected through 10-fold cross-validation. The final meta-model was used to combine the corresponding base-model predictions in the same-period holdout and temporal validation cohorts. No feature selection, preprocessing parameter, model coefficient, hyperparameter, stacking weight, or classification threshold was re-estimated using the temporal validation cohort.

ROC-AUC was the primary measure of discrimination and was reported separately for the training out-of-fold, same-period holdout, and temporal validation data sets. Precision-recall area under the curve was calculated by trapezoidal integration of precision over recall. Additional classification measures included accuracy, balanced accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Matthews correlation coefficient, Jaccard index, and detection prevalence. For each model, the classification threshold was selected by maximizing the Youden index using its training-set out-of-fold predictions and was then applied unchanged to the same-period holdout and temporal validation cohorts. Calibration was described by Brier score and by grouping predicted probabilities into 10 approximately equal-sized groups and comparing mean predicted probability with observed event proportion. Calibration-in-the-large/intercepts, calibration slopes, and confidence intervals were not estimated in the original analysis; therefore, the reported calibration results should be interpreted as descriptive rather than as a complete assessment of systematic miscalibration. Likewise, ROC-AUC, precision-recall AUC, Brier score, and threshold-based metrics are reported as point estimates without confidence intervals, so small differences between models should not be interpreted as evidence of superiority. Decision curve analysis was conducted over threshold probabilities from 0.01 to 0.80 in increments of 0.01. Net benefit was calculated as the true-positive proportion minus the false-positive proportion weighted by the odds of the threshold probability and was compared with the treat-all and treat-none strategies. Because no clinical intervention was assigned according to model predictions, decision curve analysis was interpreted as an assessment of potential decision-analytic value rather than direct evidence of improved clinical outcomes.

SHapley Additive exPlanations (SHAP) were used to interpret the stacking ensemble. SHAP values were estimated on the predicted-probability scale using a Monte Carlo approximation with 60 repetitions and adjustment enabled. The training predictor matrix was used as the background dataset, and patients in the same-period holdout set were used as the cases to be explained. Global feature importance was defined as the mean absolute SHAP value for each predictor. SHAP summary plots displayed the magnitude and direction of predictor contributions, whereas dependence plots illustrated the relationships between predictor values and their corresponding SHAP contributions.

Continuous variables were summarized as medians and interquartile ranges (IQRs) and were compared between the development-period and temporal validation cohorts using the two-sided Wilcoxon rank-sum test. Categorical variables were summarized as frequencies and percentages and were compared using Pearson’s chi-square test or Fisher’s exact test, as appropriate. Baseline comparisons were descriptive and were not used for feature selection, model tuning, or modification of the fitted models. All statistical analyses were performed using R version 4.4.3. All tests were two-sided, and a P value < 0.05 was considered statistically significant.

A total of 960 patients undergoing maintenance hemodialysis were included: 600 in the January 2021-December 2023 development-period cohort and 360 in the January 2024-December 2025 temporal validation cohort. The development-period cohort was divided into a training set (n = 420) and a same-period holdout set (n = 180). Overall, the median age was 64.0 years (IQR 54.0 - 72.0), and 590 of 960 patients (61.5%) were male. High cardiac biomarker burden was present in 340 of 960 patients (35.4%), including 213 of 600 (35.5%) in the development-period cohort and 127 of 360 (35.3%) in the temporal validation cohort (P = 0.944; Table 1). Compared with the development-period cohort, the temporal validation cohort had higher red cell distribution width (median 14.59%, IQR 13.66 - 15.52 vs 14.30%, IQR 13.50 - 15.27; P = 0.006), neutrophil count (median 5.37, IQR 3.60 - 7.83 vs 4.80, IQR 3.36 - 6.62 × 10^9/L; P = 0.008), beta-2 microglobulin concentration (median 21.25, IQR 17.21 - 26.18 vs 19.25, IQR 15.84 - 23.95 mg/L; P < 0.001), and neutrophil-to-lymphocyte ratio (median 6.46, IQR 4.27 - 11.27 vs 5.77, IQR 3.79 - 9.67; P = 0.033). BNP and high-sensitivity cardiac troponin I distributions did not differ between cohorts (P = 0.971 and P = 0.322, respectively; Table 1).

Table 1. Baseline characteristics of the development-period and temporal validation cohorts.

Values are median [25th percentile; 75th percentile] or n (%). Continuous variables were compared using the two-sided Wilcoxon rank-sum test; categorical variables were compared using Pearson’s chi-square test. P values compare the January 2021-December 2023 development-period cohort with the January 2024-December 2025 same-hospital temporal validation cohort; this later-period cohort is not an external validation cohort. BMI, body mass index; SBP, systolic blood pressure; DBP, diastolic blood pressure; FBG, fasting blood glucose; WBC, white blood cell; PLT, platelet; HB, hemoglobin; MCV, mean corpuscular volume; RDW, red cell distribution width; MPV, mean platelet volume; NEU, neutrophil; LYM, lymphocyte; MONO, monocyte. TP, total protein; ALB, albumin; GLB, globulin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; Scr, serum creatinine; UA, uric acid; PTH, parathyroid hormone; B2M, beta-2 microglobulin; PNI, prognostic nutritional index; NLR, neutrophil-to-lymphocyte ratio. TC, total cholesterol; TG, triglyceride; LDL-C, low-density lipoprotein cholesterol; HDL-C, high-density lipoprotein cholesterol; hs-cTnI, high-sensitivity cardiac troponin I; BNP, B-type natriuretic peptide; PT, prothrombin time; INR, international normalized ratio; FIB, fibrinogen; APTT, activated partial thromboplastin time; TT, thrombin time; FDP, fibrin degradation products.

Ten-fold cross-validated least absolute shrinkage and selection operator logistic regression performed in the training set retained 13 predictors with nonzero coefficients at the lambda.1se value: age, male sex, body mass index, systolic blood pressure, diabetes mellitus, coronary heart disease, hemoglobin, red cell distribution width, prognostic nutritional index, neutrophil-to-lymphocyte ratio, log-transformed D-dimer, beta-2 microglobulin, and phosphate (Figure S1). These 13 predictors constituted the fixed predictor set for all subsequent model development and evaluation analyses. Spearman correlation analysis was subsequently performed among the 10 continuous predictors. No predictor pair exceeded the prespecified absolute correlation threshold of 0.50. The strongest inverse correlation was observed between the prognostic nutritional index and log-transformed D-dimer (rho = −0.42), whereas the strongest positive correlation was observed between beta-2 microglobulin and phosphate (rho = 0.34). Therefore, no predictor was removed after the correlation assessment, and all 13 LASSO-selected predictors were carried forward into the six base models and the stacking analysis (Figure S2).

Using the same 13 predictors retained by LASSO, 6 base models were developed: logistic regression, classification and regression tree, elastic net, random forest, radial support vector machine, and gradient boosting machine. Training-set out-of-fold evaluation identified elastic net and logistic regression as the strongest individual models for discrimination, with ROC-AUCs of 0.852 and 0.851 and precision-recall AUCs of 0.777 and 0.756, respectively. The radial support vector machine achieved a ROC-AUC of 0.835 and a precision-recall AUC of 0.744, followed by gradient boosting (0.828 and 0.749), random forest (0.821 and 0.732), and CART (0.744 and 0.646; Figures 1(a)-(e)). The unweighted overall performance scores ranged from 0.68 for CART to 0.77 for logistic regression (Figure 1(c)). Decision-curve analysis showed positive net benefit relative to the treat-none strategy across a broad range of low-to-moderate threshold probabilities, although the magnitude and persistence of benefit varied among models (Figure 1(f)). Grouped calibration plots provided a descriptive visual assessment of agreement between predicted and observed risks (Figure 1(g)). Cross-validation results were consistent with the out-of-fold estimates. Mean ROC-AUCs across the 10 folds were 0.86 (SD 0.06) for logistic regression, 0.86 (SD 0.07) for elastic net, 0.85 (SD 0.07) for the support vector machine, 0.84 (SD 0.06) for gradient boosting, 0.83 (SD 0.07) for random forest, and 0.75 (SD 0.05) for CART (Figure 1(h)). The corresponding precision-recall AUC results across the 10 cross-validation folds are shown in Figure 1(i). Logistic regression, elastic net, and the support vector machine consequently provided the 3 out-of-fold probability inputs for the stacking ensemble.

Figure 1. Development-period evaluation of six machine learning models. (a) Performance profiles across evaluation metrics; (b) Heat map of metric values; (c) Unweighted overall performance scores; (d) Training out-of-fold ROC curves; (e) Training out-of-fold precision-recall curves; (f) Decision-curve analysis; (g) Grouped calibration plots; (h) ROC-AUC across 10 cross-validation folds; (I) PR-AUC across 10 cross-validation folds.

In the same-period holdout set, logistic regression and elastic net had the highest discrimination point estimates among the individual models, with ROC-AUCs of 0.87 and 0.86, respectively. The support vector machine and gradient boosting each achieved a ROC-AUC of 0.84, compared with 0.80 for random forest and 0.77 for CART. In the temporal validation cohort, ROC-AUC point estimates were 0.83 for logistic regression, elastic net, and gradient boosting; 0.82 for the support vector machine; 0.81 for random forest; and 0.74 for CART (Figure 2(a)). The corresponding precision-recall curves for the same-period holdout and temporal validation cohorts are shown in Figure 2(b). Decision curves indicated that most models provided positive net benefit over treat-none across low-to-moderate thresholds in both evaluation cohorts, with net benefit generally diminishing at higher thresholds (Figure 2(c)). Descriptive calibration performance varied by model and validation period. Logistic regression had the lowest reported Brier score in same-period holdout evaluation (0.149), whereas gradient boosting had the lowest Brier score in temporal validation (0.162). The corresponding holdout and temporal validation Brier scores were 0.185 and 0.221 for CART, 0.189 and 0.190 for elastic net, 0.177 and 0.170 for random forest, and 0.161 and 0.173 for the support vector machine (Figure 2(d)). Because uncertainty intervals were not estimated, the small between-model differences in these point estimates should not be interpreted as evidence of superiority.

Figure 2. Performance of six machine learning models in the same-period holdout and temporal validation cohorts. (a) ROC curves; (b) Precision-recall curves; (c) Decision-curve analysis; (d) Grouped calibration plots with Brier scores.

The three base models with the highest mean cross-validated ROC-AUCs were logistic regression, elastic net, and the radial support vector machine; these models were selected as components of the stacking ensemble. Each component model was developed using the same 13 LASSO-selected predictors. Their out-of-fold predicted probabilities, rather than the original predictor values, were used as inputs to an L1-penalized logistic meta-model. The ensemble achieved ROC-AUC point estimates of 0.848 in training out-of-fold evaluation, 0.863 in same-period holdout evaluation, and 0.835 in temporal validation. The corresponding precision-recall AUC point estimates were 0.77, 0.76, and 0.70, and the Brier scores were 0.15, 0.15, and 0.16, respectively (Figure 3(a)). Thus, the ensemble remained comparable by point estimate with the strongest individual models; the small observed differences cannot establish algorithmic superiority without uncertainty intervals.

Figure 3. Performance and interpretation of the stacking ensemble. (a) Performance profiles and heat maps for the base models and stacking ensemble across training out-of-fold evaluation, same-period holdout evaluation, and temporal validation; (b) ROC curves for the stacking ensemble across the three data sets; (c) Mean absolute SHAP values; (d) SHAP summary and categorical-feature dependence plots; (e) Continuous-feature SHAP dependence plots.

At the prespecified classification threshold, the stacking model had sensitivity and specificity point estimates of 0.68 and 0.85 in training out-of-fold evaluation, 0.60 and 0.86 in same-period holdout evaluation, and 0.75 and 0.76 in temporal validation. Accuracy point estimates were 0.79, 0.77, and 0.75 across these data sets, respectively (Figure 3(a)). The ROC curves for the stacking ensemble across the three data sets are shown in Figure 3(b).

In the same-period holdout set, global SHAP analysis quantified the contributions of the 13 LASSO-selected predictors to the stacking ensemble. In descending order of mean absolute SHAP value, the predictors were prognostic nutritional index, log-transformed D-dimer, age, red cell distribution width, beta-2 microglobulin, coronary heart disease, male sex, body mass index, neutrophil-to-lymphocyte ratio, hemoglobin, diabetes mellitus, systolic blood pressure, and phosphate (Figure 3(c) and Figure 3(d)). Lower prognostic nutritional index values and higher values of log-transformed D-dimer, age, red cell distribution width, beta-2 microglobulin, and body mass index were associated with larger positive contributions to predicted risk across their observed ranges (Figure 3(e)). Coronary heart disease, male sex, and diabetes mellitus also contributed positively to model predictions, whereas systolic blood pressure and phosphate had comparatively small mean absolute SHAP values.

The single-patient explanation illustrated how these population-level patterns combined within an individual prediction. For the example patient, the predicted probability was 0.450 compared with an adjusted background expectation of 0.342. Red cell distribution width of 17.90%, log-transformed D-dimer of 1.92, male sex, body mass index of 24.01 kg/m², and a neutrophil-to-lymphocyte ratio of 8.19 increased the prediction, whereas a prognostic nutritional index of 44.65, age of 43 years, absence of coronary heart disease, beta-2 microglobulin of 16.86 mg/L, and hemoglobin of 83.00 g/L decreased it (Figure S3).