This retrospective cohort study was conducted at Fuwai Central China Cardiovascular Hospital. Between January 2018 and January 2026, a total of 157 consecutive patients diagnosed with FM who were hospitalized at our center were included in this study.

The diagnosis of FM was established according to the Chinese Society of Cardiology guidelines on the diagnosis and treatment of adult FM^[15]. Patients were eligible for inclusion if they had a confirmed diagnosis of FM and available major early admission clinical data required for subsequent analyses.

Patients were excluded if critical baseline variables or outcome data were missing, or if they had pre-existing end-stage renal disease requiring chronic renal replacement therapy prior to CCU admission.

The study protocol was reviewed and approved by the Ethics Committee for Scientific Research, New Technologies and New Services of Fuwai Central China Cardiovascular Hospital (Approval No. 2026-Lun-23) and was conducted in accordance with the Declaration of Helsinki. Given the retrospective nature of the study, the use of anonymized clinical data, and the absence of any additional patient contact or intervention, the requirement for informed consent was waived by the ethics committee.

New-onset AKI was defined according to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria^[16]. Patients were classified as having new-onset AKI if they initially fulfilled KDIGO AKI criteria beyond the first 24 h after CCU admission, based on any of the following: (1) increase in serum creatinine (SCr) of ≥ 0.3 mg/dL within 48 h; (2) increase in SCr to ≥ 1.5 times baseline within 7 days; or (3) urine output < 0.5 mL/kg/h for at least 6 h^[16]. Patients meeting AKI criteria within the first 24 h post-CCU admission were not considered as having new-onset AKI for the primary outcome. The primary outcome was defined as the development of new-onset AKI occurring after the initial 24 h of CCU admission.

Baseline characteristics were selected based on routine availability during the early CCU period, clinical relevance to AKI risk, and availability before AKI onset. All selected variables were obtained from clinical records within the first 24 h after CCU admission, before the assessment window for new-onset AKI, to support early individualized risk assessment. Demographic and anthropometric variables included age, sex, and body mass index (BMI). Presenting symptoms were documented and coded as symptoms (ranging from 0-6). Vital signs recorded at admission comprised heart rate, systolic blood pressure (SBP), diastolic blood pressure (DBP), respiratory rate, and body temperature.

Laboratory measurements included hematological parameters (white blood cell [WBC], red blood cell [RBC], platelet count, neutrophils, lymphocytes, red cell distribution width [RDW], hemoglobin, hematocrit), inflammatory markers (C-reactive protein [CRP]), electrolytes and metabolic indices (sodium, potassium, calcium, chloride, glucose, anion gap, lactate), liver-associated biomarkers (alanine aminotransferase [ALT], aspartate aminotransferase [AST], total bilirubin), lipid profile (triglycerides, total cholesterol, LDL-cholesterol [LDL-C], HDL-cholesterol [HDL-C]), cardiac biomarkers (creatine kinase-MB [CK-MB], B-type natriuretic peptide [BNP]), and baseline renal function markers, defined as urea and SCr measured within the first 24 h after CCU admission and before the assessment window for new-onset AKI. Cardiac systolic function was assessed by ejection fraction (EF) at admission.

Comorbidities and clinical conditions documented within the first 24 h after CCU admission included hypertension, diabetes, coronary heart disease, and pulmonary infection. Pulmonary infection was defined according to standardized ICD-10-coded clinical records and routine clinical documentation available within this early 24-h window. Treatment-related variables included the use of vasoactive agents, mechanical circulatory support, pacemaker insertion, and mechanical ventilation, all of which were extracted only if documented within the first 24 h after CCU admission. Mechanical ventilation referred to documented use of invasive or non-invasive ventilatory support during this early 24-h window, rather than during the entire CCU stay.

To facilitate modeling, several categorical variables were derived and standardized. First, the variable symptoms was converted into a composite variable (symptom_code), categorized into five symptom groups: fever (1), chest tightness or chest pain (2), syncope or loss of consciousness (3), poor appetite or fatigue (4), and diarrhea (5). If two or more symptom groups were present, symptom_code was assigned a value of 6; if no symptoms were present, it was set to 0.

Second, vasoactive therapy exposure was summarized as vasoactive_count (ranging from 0-3), indicating the number of vasoactive agents administered among norepinephrine, dopamine, and metaraminol (0 = none; 1 = one agent; 2 = two agents; 3 = three agents).

Third, mechanical circulatory support was recoded based on the provision of intra-aortic balloon pump (IABP) and/or extracorporeal membrane oxygenation (ECMO) as follows: 0 (neither), 1 (IABP only), 2 (ECMO only), and 3 (both).

A total of 47 baseline variables were considered as candidate predictors for model development. The incident AKI outcome had no missing data, and missingness among the candidate predictors was minimal, with all variables having missing rates below 0.7%. Missing values in candidate predictors were handled using multiple imputation before model development.

The initial dataset contained 47 features. To minimize multicollinearity, pairwise correlation analyses were conducted, and variables exhibiting correlation coefficients exceeding 0.90 were removed.

Patients were randomly divided into a training set and an internal test set at a 7:3 ratio. Continuous variables were standardized using z-score normalization, and categorical variables underwent one-hot encoding.

To address class imbalance, the synthetic minority oversampling technique (SMOTE) was applied exclusively to the training set. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO). All preprocessing steps, including correlation-based filtering, normalization, encoding, SMOTE, and LASSO selection, were initially applied and fitted on the training set, and subsequently applied to the internal test set to avoid information leakage.

Baseline characteristics were compared between the AKI and non-AKI groups. Continuous variables with approximately normal distributions are presented as mean ± standard deviation (SD), whereas non-normally distributed continuous variables are presented as median (interquartile range [IQR]). Categorical variables are presented as n (%). Between-group comparisons were performed using one-way analysis of variance (ANOVA) for approximately normally distributed continuous variables, the Kruskal-Wallis H test for non-normally distributed continuous variables, and the chi-square test or Fisher’s exact test for categorical variables, as appropriate.

For machine-learning model development, 14 algorithms were implemented and benchmarked within the mlr3 framework: AdaBoostM1, Bayesian network (Bayes_net), CatBoost, ExtraTrees, gradient boosting machine (GBM), k-nearest neighbors (KKNN), LightGBM, logistic regression (logistic), naive Bayes (Naive_bayes), neural network (NNET), random forest (Ranger), classification and regression tree (RPART), support vector machine (SVM), and extreme gradient boosting (XGBoost)^[17-25]. A standardized preprocessing and evaluation workflow was maintained consistently across training and internal test sets. Model performance metrics reported included area under the curve (AUC), accuracy (ACC), Brier score, F-beta score (Fbeta), classification error (CE), precision-recall AUC (PRAUC), sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV).

Within the training set, 10-fold cross-validation was utilized to derive resampling-based performance estimates for benchmarking the candidate algorithms, and differences in performance metrics across folds were assessed using ANOVA or the Kruskal-Wallis H test, as appropriate. Therefore, the AUC reported during the machine-learning benchmarking stage represented the 10-fold cross-validation AUC within the training set, rather than an independent external testing AUC or a bootstrap optimism-corrected AUC.

For the model demonstrating optimal and robust internal validation performance, SHAP-based interpretability analyses were conducted to quantify feature contributions and provide both global and patient-specific explanations. After selection of the final logistic regression-based predictive model, additional internal validation was performed using 1,000 bootstrap resampling iterations. In this bootstrap procedure, the final model specification with the selected predictors was fixed, while model coefficients were refitted within each bootstrap sample. The complete machine-learning algorithm selection and model selection procedures were not repeated within each bootstrap sample. For optimism correction, the AUC of the refitted model was calculated both in the bootstrap sample and in the original full cohort, and the difference between these two AUC values was used to estimate optimism for each bootstrap iteration. The average optimism across 1,000 bootstrap resamples was then subtracted from the apparent AUC of the final model fitted in the full cohort to obtain the optimism-corrected AUC. The reported confidence intervals were derived from the bootstrap distribution of the internally validated AUC estimates.

Bootstrap-corrected model performance was assessed in terms of discrimination using the optimism-corrected AUC, calibration using calibration curves as well as calibration intercept and slope, and clinical utility using decision curve analysis. To evaluate the robustness of model performance with respect to potentially less objective clinical predictors, sensitivity analyses were further conducted by excluding pulmonary infection and mechanical ventilation individually and simultaneously from the logistic regression model, followed by repeated bootstrap internal validation and assessment of discrimination, calibration, and decision curve performance. All statistical analyses and ML model training were performed using Free Statistics (version 2.4), an R-based platform integrating the mlr3 ecosystem for benchmarking and interpretability analyses^[26].

A total of 157 patients with FM admitted to the CCU were included; 58 (36.9%) developed new-onset AKI, while 99 (63.1%) did not [Table 1]. Patients who developed AKI exhibited higher heart rates and greater inflammatory and metabolic disturbances, characterized by higher WBC counts, glucose, and anion gap levels, and lower platelet counts, serum calcium, chloride, and HDL-C levels [Table 1]. In addition, the median respiratory rate was statistically lower in the AKI group than in the non-AKI group, although the absolute difference was modest (18.00 [13.25, 22.00] vs. 20.00 [18.00, 22.00] breaths/min, P = 0.009; Table 1). Cardiac function at admission was more impaired in the AKI group, as reflected by a lower EF. Furthermore, these patients more frequently required intensive interventions, including vasoactive agents, mechanical support, and mechanical ventilation [Table 1]. AKI was also associated with a higher incidence of pulmonary infection [Table 1]. Markers of hypoperfusion and multi-organ injury were significantly elevated among patients with AKI, including higher lactate levels, liver enzymes (ALT and AST), total bilirubin, cardiac biomarkers (CK-MB and BNP), and renal parameters (urea and creatinine) at admission [Table 1]. In contrast, age, sex, BMI, major comorbidities, including hypertension, diabetes, and coronary heart disease, and several hematologic indices, including hemoglobin, hematocrit, and RDW, did not differ significantly between groups [Table 1].

Prior to model development, distributions and pairwise correlations among candidate variables were evaluated. The initial dataset included 47 features. To mitigate multicollinearity, pairwise correlations were assessed, and variables exhibiting correlation coefficients greater than 0.90 were excluded. The remaining predictors were standardized (z-score normalization for continuous variables; one-hot encoding for categorical variables) exclusively within the training set. Due to class imbalance, SMOTE oversampling was applied only to the training set. Subsequently, feature selection was performed using LASSO regression, retaining six predictors for modeling: anion gap, chloride, mechanical ventilation, pulmonary infection, total bilirubin, and urea. All preprocessing steps, including correlation-based filtering, scaling, encoding, SMOTE, and LASSO, were fitted to the training set and then consistently applied to the internal test set to prevent information leakage.

Figure 1 illustrates the correlation matrix among the six selected predictors. Pairwise correlations were generally low to moderate, indicating minimal multicollinearity among retained variables after filtering and feature selection.

In the baseline benchmark for the binary outcome, multiple models (AdaBoostM1, BayesNet, CatBoost, ExtraTrees, GBM, KKNN, LightGBM, Logistic regression, NaiveBayes, NNET, Ranger, RPART, SVM, and XGBoost) were evaluated on the same dataset [Table 2]. Discrimination performance (AUC) ranged from 0.73 to 0.87. Logistic regression achieved top-tier discrimination (AUC = 0.87, equal to GBM and NaiveBayes) and demonstrated robust precision-recall performance (PRAUC = 0.80), alongside consistent overall accuracy (ACC = 0.77; Fbeta = 0.68). Regarding calibration, logistic regression displayed favorable performance (Brier score = 0.16; CE = 0.23), indicating reliable probability predictions. Additionally, logistic regression provided a balanced sensitivity-specificity trade-off (sensitivity = 0.67; specificity = 0.83). Given its competitive discrimination, strong calibration, and practical interpretability, logistic regression was selected as the final model for subsequent analysis.

To further assess the robustness and clinical utility of the logistic regression-based predictive model, we performed internal validation using 1,000 bootstrap resampling iterations. As shown in Figure 2A, the logistic regression model demonstrated good discrimination. The apparent AUC was 0.944, and the mean estimated optimism derived from 1,000 bootstrap resamples was 0.013, resulting in an optimism-corrected AUC of 0.931 (95%CI: 0.892-0.969). The calibration curve showed good agreement between the predicted probabilities generated by the logistic regression model and the observed outcomes [Figure 2B]. After 1,000 bootstrap resamples, the optimism-corrected calibration intercept was -0.040 and the calibration slope was 0.842. Decision curve analysis demonstrated that the logistic regression model provided a higher net benefit than both the treat-all and treat-none strategies at threshold probabilities ranging approximately from 0.02 to 0.99 [Figure 2C].

To further evaluate the stability of the model, three sensitivity analyses were performed based on the full logistic regression model, and internal validation was conducted using 1,000 bootstrap resampling iterations for each analysis. Overall, the results of the sensitivity analyses showed trends consistent with those of the full model in terms of discrimination, calibration, and clinical net benefit.

After simultaneous exclusion of pulmonary infection and mechanical ventilation, the model showed an optimism-corrected AUC of 0.827 (95%CI: 0.756-0.898; Supplementary Figure 1A). After 1,000 bootstrap resamples, the optimism-corrected calibration intercept was -0.056 and the calibration slope was 0.909, indicating generally good calibration after internal validation [Supplementary Figure 1B]. Decision curve analysis showed that the model provided a higher net benefit than both the treat-all and treat-none strategies at threshold probabilities of approximately 0.02-0.99 [Supplementary Figure 1C].

After exclusion of pulmonary infection alone, the optimism-corrected AUC was 0.931 (95%CI: 0.893-0.968; Supplementary Figure 2A). After 1,000 bootstrap resamples, the optimism-corrected calibration intercept was -0.029 and the calibration slope was 0.868, indicating generally good calibration after internal validation [Supplementary Figure 2B]. The corresponding decision curve analysis showed that the model provided a higher net benefit than both the treat-all and treat-none strategies at threshold probabilities of approximately 0.02-0.99 [Supplementary Figure 2C].

After exclusion of mechanical ventilation alone, the model showed an optimism-corrected AUC of 0.836 (95%CI: 0.766-0.906; Supplementary Figure 3A). After 1,000 bootstrap resamples, the optimism-corrected calibration intercept was -0.068 and the calibration slope was 0.886, indicating generally good calibration after internal validation [Supplementary Figure 3B]. Decision curve analysis further showed that the model provided a higher net benefit than both the treat-all and treat-none strategies at threshold probabilities of approximately 0.02-0.99 [Supplementary Figure 3C].

Together, these sensitivity analyses showed that after excluding pulmonary infection and/or mechanical ventilation, the model maintained good calibration and favorable clinical net benefit. Excluding pulmonary infection alone produced results similar to the full model. Excluding mechanical ventilation, either alone or together with pulmonary infection, resulted in lower AUC values; however, the optimism-corrected AUCs remained above 0.80.

A nomogram was developed based on the final logistic regression model to enable individualized prediction of new-onset AKI after CCU admission in FM [Figure 5]. The nomogram incorporates six predictors: mechanical ventilation, anion gap, chloride, total bilirubin, pulmonary infection, and urea. For each patient, points are assigned according to the value of each predictor and summed to generate a total score, which corresponds to the linear predictor and the estimated probability of AKI. This graphical tool translates the final model into a clinically applicable instrument for bedside risk assessment.

Several limitations should be acknowledged. First, this was a single-center study with a relatively small sample size, which may limit the generalizability of the findings to other CCU populations or institutions with different patient characteristics, treatment strategies, and clinical workflows. Therefore, although the model showed favorable performance in the present cohort, its applicability to broader clinical settings should be interpreted with caution.

Second, this study was observational in nature, and the associations between candidate predictors and new-onset AKI may have been influenced by unmeasured or residual confounding. Although we included routinely available demographic, clinical, laboratory, comorbidity, and treatment-related variables, some potentially relevant factors, such as overall disease severity, multi-organ dysfunction, detailed hemodynamic parameters, nephrotoxic medication exposure, fluid balance, contrast use, and dynamic changes in renal function, were not fully incorporated into the model. Therefore, the observed associations may partly reflect underlying illness severity rather than direct causal effects.

Third, the model was developed using baseline variables obtained within the first 24 h after CCU admission. This design was intended to support early risk stratification before the assessment window for new-onset AKI; however, it may not capture subsequent clinical deterioration or treatment changes during hospitalization. Future studies incorporating time-updated variables may further improve predictive accuracy.

Fourth, although we performed 7:3 internal validation, 10-fold cross-validation, and additional bootstrap internal validation with optimism-corrected performance estimates, no independent external validation cohort was available in the present study. External validation in multicenter cohorts is therefore necessary to confirm the robustness, calibration, and clinical utility of the model before routine clinical implementation.

Fifth, because this was a retrospective observational study, the model should be interpreted as a predictive and associative tool rather than a causal model. The identified predictors should not be regarded as direct causal determinants of new-onset AKI. In particular, treatment- and complication-related variables, such as mechanical ventilation and pulmonary infection, may partly reflect underlying disease severity, systemic inflammatory burden, or clinical deterioration rather than acting as independent causal drivers of AKI. Therefore, the model is intended to support early risk stratification, but not to determine causal pathways or justify interventions targeting individual predictors.

Despite these limitations, the model was based on early and easily accessible clinical variables, making it potentially useful as an interpretable tool for early identification of patients at high risk of new-onset AKI in the CCU setting.

A parsimonious prediction model incorporating six readily available bedside variables was developed for predicting new-onset AKI in adult patients with FM upon CCU admission. The model showed robust discrimination and good calibration, offering a practical approach for early risk stratification and kidney-protective management in the CCU setting.