Abstract
Background
Elderly heart failure (HF) patients have an increased risk of acute kidney injury (AKI) during hospitalization. Identifying risk factors and developing a prediction model are clinically important.
Study Method
We retrospectively enrolled 7076 elderly HF patients hospitalized from January 2019 to December 2023. All predictors were restricted to data available within 24 hours of admission and prior to AKI occurrence. Candidate predictors were first screened by univariate analysis (P<0.05) and then further refined by LASSO regression, and the selected variables were ultimately entered into multivariable logistic regression. Thirteen machine learning models were constructed and compared to identify the optimal model. Model performance was evaluated using the area under the ROC curve, calibration curves, and decision curve analysis.
Results
AKI occurred in 796 patients (11.2%). The final model included diuretic use, hypertension, diabetes, valvular heart disease, COPD, CKD, admission cystatin C, and LVEF. Among the 13 machine learning models, logistic regression performed best in the validation set (AUC = 0.804, 95% CI: 0.777–0.830). Calibration was acceptable in both the training and validation sets, with Hosmer–Lemeshow P values of 0.561 and 0.605, calibration slopes of 0.896 and 0.835, and Brier scores of 0.078 and 0.082, respectively. Decision curve analysis confirmed its clinical utility. A nomogram was constructed for model visualization.
Conclusion
In this single-center study with internal validation only, the prediction model demonstrated favorable performance for in-hospital AKI in elderly HF patients and may serve as a candidate tool for early risk stratification. External validation in independent cohorts is required before routine clinical implementation.
Keywords: heart failure, acute kidney injury, elderly patients, risk model
Background
Heart failure (HF), as the terminal stage of cardiovascular disease, is among the leading causes of hospitalization and death among elderly people worldwide, and its high morbidity and mortality rates have become global public health challenges. The incidence rate of heart failure among people aged 70 and above in developed countries is as high as 10%.1 When a patient experiences heart failure, reduced cardiac output, venous congestion, and activation of the fluid response can cause acute kidney injury (AKI). Some medications used to treat heart failure may also trigger AKI.2 Previous studies have shown that approximately 10 to 20% of patients with acute heart failure develop acute kidney injury3,4 and once this is combined with AKI, the mortality rate increases significantly, and the medical burden also increases greatly. Elderly patients with heart failure often have multiple chronic diseases, such as hypertension, diabetes, and anaemia.5 Moreover, compared with those of normal patients, the kidney reserve and repair function of elderly patients are reduced, the probability of acute renal failure is greater, the prognosis is poorer, the likelihood of developing chronic kidney disease (CKD) is greater, and the dependence on dialysis is significantly greater. According to statistics from the American Kidney Data System in 2018, the mortality rate among hospitalized AKI patients older than 66 years was as high as 8.2%.6
The glomerular filtration rate decreases with age, and the muscle mass and nutritional status of elderly individuals also affect serum creatinine levels. Relying solely on serum creatinine levels to determine changes in renal function in elderly patients is not accurate enough. If the risk factors related to AKI are identified early and prevented in time, 20% to 30% of AKI cases can be predicted and avoided.7 For populations with a significantly increased risk of AKI in different clinical scenarios, several prediction models have been developed in recent years. For instance, Wang and Zhao et al constructed a prediction model for patients with acute decompensated heart failure (ADHF) and identified six predictive factors, namely, advanced age, diabetes, previous renal dysfunction, elevated baseline creatinine, elevated B-type natriuretic peptide, and low albuminaemia. However, this model is applicable only to ADHF patients and has limited discriminatory ability.8 A study establishing a machine learning prediction model for patients undergoing coronary intervention revealed that advanced age, eGFR, preoperative haemoglobin level, ST-segment elevation myocardial infarction, symptoms of heart failure, and cardiogenic shock are important predictors of postoperative AKI.9 A prospective study established an AKI risk prediction model for ICU patients, which included risk factors such as hypertension, chronic kidney disease, acute pancreatitis, heart failure, shock, a pH ≤ 7.30, a CK > 1000 U/L, hypoproteinaemia, exposure to nephrotoxic drugs, and male sex.10 Previous models were mostly constructed on the basis of patients with heart failure, those who underwent cardiac surgery, or those in the ICU. However, this study focused specifically on elderly patients aged 65 years or older who had heart failure. This group simultaneously experienced renal reserve deficiency due to ageing and renal perfusion deficiency due to heart dysfunction. The susceptibility to AKI is significantly increased in this group. However, dedicated risk prediction tools for this specific population remain limited.
To address these gaps, the present study specifically targeted elderly hospitalized patients with heart failure, a population that remains understudied in existing AKI prediction research. Methodologically, we employ a rigorous variable selection strategy, including univariate screening, multivariable logistic regression, and least absolute shrinkage and selection operator (LASSO) regression, and systematically compare thirteen machine learning models to identify the optimal model. Furthermore, we prioritize clinically accessible predictors that can be readily obtained within 24 hours of admission, ensuring practical utility at the bedside. Therefore, this study aims to develop and validate a risk prediction model for AKI in hospitalized elderly patients with heart failure, providing a simple, effective, and interpretable tool for early risk stratification and individualized intervention in clinical practice.
Materials and Methods
Research Design
This was a retrospective study. We included 7842 elderly patients with heart failure who were hospitalized at the General Hospital of the Northern Theater Command from January 1, 2019, to December 31, 2023. The inclusion criteria included heart failure occurring during hospitalization (including acute and chronic heart failure), age ≥65 years, length of hospital stay ≥48 hours, and complete data. The exclusion criteria were as follows: patients with end-stage renal disease, kidney transplantation, and dialysis treatment; patients with acute kidney injury at admission; patients with contrast-induced nephropathy caused by interventional therapy; and patients with a history of malignant tumours. Patients with missing data for any predictor or outcome variable were excluded. In accordance with the inclusion and exclusion criteria, 7076 patients were ultimately included in this study (Figure 1). The outcome was the occurrence of AKI during hospitalization.
Figure 1.
This flowchart illustrates the process of patient selection for inclusion in the final analysis. A total of 7842 patients were initially screened. After the exclusion criteria were applied, 7076 patients were included in the final analysis.
This study was conducted in accordance with the Declaration of Helsinki and approved by the Clinical Research Ethics Committee of the General Hospital of the Northern Theater Command (Approval No. Y2025-364). Informed consent was waived because of the retrospective nature of the study. The detailed ethical and confidentiality statements are provided in the Ethics Statement section at the end of the manuscript.
Definitions
AKI is diagnosed on the basis of the criteria of the Organization for Improving Global Prognosis of Kidney Disease (KDIGO),11 and it can be diagnosed if any of the following conditions are met: 1) an elevated serum creatinine concentration (Δ Cr) > 26.5 µmol/L within 48 hours; 2) an increase of ≥1.5 times in serum creatinine concentration within 7 days; 3) a urine output < 0.5 mL/kg/h for more than 6 hours. Baseline values were set for the patient’s serum creatinine and urea nitrogen levels at admission.
The diagnosis of heart failure was based on the 2023 ESC diagnostic criteria.12 After admission, the patient’s clinical symptoms, laboratory indicators, and auxiliary examination results were confirmed by a cardiologist.
Prediction Time Point
The prediction time point (time zero) was defined as hospital admission. All predictors were restricted to data available within 24 hours of admission and prior to AKI occurrence. Laboratory measurements were obtained from the first blood draw after admission, and all examination results, including laboratory and ancillary examination findings, were also restricted to those obtained before AKI occurrence. The medications recorded were limited to those taken chronically before admission. This ensured that all predictors preceded the outcome, eliminating the risk of temporal outcome leakage.
Data Collection
We collected data on age; sex; weight; comorbidities (hypertension, diabetes, arrhythmia, cerebral infarction, valvular heart disease, and chronic obstructive pulmonary disease); cardiac function classification; and the use of angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers, beta-blockers, calcium channel blockers, vasoactive drugs, antiplatelet drugs, anticoagulants, and diuretics. Laboratory data were obtained from the hospital’s Rimet laboratory management system, and admission laboratory parameters were collected. The laboratory parameters included WBC count, RBC count, PLT count, haemoglobin, serum albumin, aspartate aminotransferase, alanine aminotransferase, alkaline phosphatase, total bilirubin, blood uric acid, C-reactive protein, high-sensitivity troponin T, N-terminal pro-B-type natriuretic peptide, total cholesterol, triglycerides, high-density lipoprotein, low-density lipoprotein, D-dimer, blood urea nitrogen, serum creatinine, cystatin C, blood glucose, and urine protein. Auxiliary examination parameters, such as the left ventricular ejection fraction (LVEF) and the presence of pleural effusion, were obtained from the hospital’s electronic medical records system.
Data Analysis
Continuous variables conforming to a normal distribution are presented as the mean ± standard deviation and were compared using independent samples t-tests; those with a severely skewed distribution are expressed as the median (interquartile range) and were compared using the Mann–Whitney U-test. Categorical variables are described as numbers (percentages) and were compared using the chi‑square test.
In this study, data from 7076 elderly patients with heart failure were randomly divided into training and validation cohorts at a 7:3 ratio. To ensure the independence of the model evaluation, the validation set was used solely for assessing model performance and was not involved in variable selection, model training, or parameter optimization. Candidate variables were subjected to univariate analysis followed by LASSO regression for variable selection. First, on the basis of the literature and clinical experience, 43 potential factors associated with AKI development in elderly heart failure patients were included. Variables with P < 0.05 in univariate analysis were selected. Next, LASSO regression was applied to compress the variables. The optimal penalty parameter λ was determined by tenfold cross-validation, and the absolute values of the standardized regression coefficients were used as a quantitative measure of variable importance. Variables with absolute standardized regression coefficients greater than 20% were retained, whereas those below this threshold were excluded because of their low marginal contribution. Finally, the variables selected by LASSO were incorporated into a multivariable logistic regression model to further evaluate their independent predictive value.
During the model construction phase, 13 machine learning models were built separately on the training dataset, including random forest, gradient boosting, support vector machine (SVM_Kernel), CATBoost, K-nearest neighbours (NeighbourMethod), partial least squares (PLSModel), boosting, neural network (NeuralNet), naive Bayes (BayesMethod), discriminant analysis (DiscriminantModel), AdaBoost (AdaptiveBoosting), logistic regression (LogisticModel), and LightGBM. Each model employed tenfold cross-validation for hyperparameter optimization to identify the best parameter combinations.
The predictive performance of the 13 different machine learning models was comprehensively compared in both the training and validation datasets. The performance metrics included the specificity, accuracy, positive predictive value, negative predictive value, F1 score, and Youden index. Discriminative ability was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC), with 95% confidence intervals calculated for the AUC. Model calibration was assessed using calibration curves, calibration slopes, calibration intercepts, and Brier scores. Clinical utility was evaluated through decision curve analysis (DCA). On the basis of these indicators, the model with the highest AUC in the validation set, along with good calibration and clinical usefulness, was selected as the final prediction model.
Statistical analysis was performed using R language version 4.5.1. P < 0.05 was considered to indicate statistical significance.
Results
Baseline Features
This study included a total of 7076 elderly patients with heart failure who were hospitalized at the General Hospital of the Northern Theater Command. They were randomly divided into a training set (4954 cases) and a validation set (2122 cases) at a 7:3 ratio. There was no statistically significant difference in demographic characteristics such as age (69.76 ± 6.43 years vs 69.57 ± 6.33 years, P = 0.252), sex (64.9% male vs 66.9%, P = 0.111), or weight (67.71 ± 10.75 kg vs 68.03 ± 10.51 kg, P = 0.237) between the training set and the validation set. There were no significant differences in the prevalence of comorbidities such as chronic kidney disease (CKD) (8.4% vs 8.9%, P = 0.593), hypertension (54.7% vs 55.0%, P = 0.825), or COPD (7.9% vs 7.5%, P = 0.579) between the two groups. Overall, the distribution of baseline characteristics in the training set and the validation set was balanced, suggesting a good randomization effect and comparability between the two groups. The baseline characteristics of the two groups are shown in Table 1.
Table 1.
Baseline
| Variable | Overall (n=7076) | Training (n=4954) | Validation (n=2122) | P |
|---|---|---|---|---|
| Sex (Male, n %) | 4636 (65.50) | 3216 (64.90) | 1420 (66.90) | 0.111 |
| Weight (kg) | 67.80 (10.68) | 67.71 (10.75) | 68.03 (10.51) | 0.237 |
| Age (years) | 69.70 (6.40) | 69.76 (6.43) | 69.57 (6.33) | 0.252 |
| CKD (yes, n %) | 606 (8.6) | 418 (8.4) | 188 (8.9) | 0.593 |
| Hypertension (yes, n%) | 3879 (54.8) | 2711 (54.7) | 1168 (55.0) | 0.825 |
| Diabetes (yes, n %) | 1942 (27.4) | 1395 (28.2) | 547 (25.8) | 0.043 |
| CVD (yes, n %) | 2928 (41.4) | 2064 (41.7) | 864 (40.7) | 0.475 |
| COPD (yes, n %) | 551 (7.8) | 392 (7.9) | 159 (7.5) | 0.579 |
| Arrhythmia (yes, n %) | 1712 (24.2) | 1216 (24.5) | 496 (23.4) | 0.306 |
| Stroke (yes, n %) | 1088 (15.4) | 741 (15.0) | 347 (16.4) | 0.146 |
| VHD (yes, n %) | 1541 (21.8) | 1090 (22.0) | 451 (21.3) | 0.504 |
| Diuretic yes, (n %) | 4381 (61.9) | 3094 (62.5) | 1287 (60.7) | 0.16 |
| Vasoactive drugs (yes, n %) | 2247 (31.8) | 1592 (32.1) | 655 (30.9) | 0.307 |
| ACEIs/ARBs (yes, n %) | 3676 (52.0) | 2526 (51.0) | 1150 (54.2) | 0.014 |
| Anticoagulants (yes, n %) | 2941 (41.6) | 2066 (41.7) | 875 (41.2) | 0.734 |
| Antiplatelet drugs (yes, n %) | 4472 (63.2) | 3106 (62.7) | 1366 (64.4) | 0.189 |
| Urine protein (yes, n%) | 1136 (16.1) | 805 (16.2) | 331 (15.6) | 0.517 |
| LVEF (%) | 0.52 ± 0.12 | 0.52 ± 0.12 | 0.52 ± 0.12 | 0.609 |
| WBC (×109/L) | 7.55 ± 3.03 | 7.57 ± 3.08 | 7.52 ± 2.92 | 0.51 |
| RBC (×1012/L) | 4.27 ± 0.66 | 4.27 ± 0.67 | 4.28 ± 0.65 | 0.751 |
| Hb (g/L) | 131.86 ± 22.10 | 131.78 ± 22.43 | 132.03 ± 21.31 | 0.665 |
| Plt (×109/L) | 197.41 ± 55.71 | 196.71 ± 55.98 | 199.06 ± 55.07 | 0.103 |
| CRP (mg/L) | 4.35 (1.40, 20.40) | 4.40 (1.40, 20.48) | 4.30 (1.37, 20.37) | 0.625 |
| Total bilirubin (μmol/L) | 9.40 (5.90, 13.90) | 9.50 (5.90, 13.90) | 9.40 (5.90, 13.80) | 0.554 |
| ALT (U/L) | 19.48 (15.00, 27.36) | 19.61 (15.00, 27.20) | 19.20 (15.00, 27.64) | 0.964 |
| AST (U/L) | 20.87 (14.07, 33.95) | 20.77 (14.01, 34.12) | 21.00 (14.29, 33.61) | 0.643 |
| AKP (U/L) | 65.32 ± 25.69 | 64.99 ± 25.73 | 66.08 ± 25.58 | 0.104 |
| Alb (g/L) | 36.68 ± 4.96 | 36.72 ± 5.00 | 36.59 ± 4.85 | 0.337 |
| BUN (mmol/L) | 6.95 (5.55, 9.03) | 6.98 (5.57, 9.05) | 6.88 (5.52, 8.95) | 0.524 |
| SCr (μmol/L) | 77.90 (63.90, 88.40) | 77.89 (63.60, 88.20) | 78.00 (64.70, 88.68) | 0.219 |
| Cysc (mg/L) | 1.04 (0.87, 1.33) | 1.04 (0.87, 1.34) | 1.05 (0.87, 1.32) | 0.473 |
| Uric acid (μmol/L) | 368.00 (294.00, 456.00) | 369.00 (293.00, 453.75) | 366.00 (297.00, 462.00) | 0.388 |
| Glu (mmol/L) | 7.15 ± 3.12 | 7.16 ± 3.10 | 7.13 ± 3.15 | 0.742 |
| CtnT (ng/L) | 0.88 (0.01, 24.00) | 1.00 (0.01, 24.00) | 0.43 (0.01, 24.00) | 0.561 |
| NT-proBNP (pg/mL) | 2098.00 (656.08, 3832.28) | 2075.50 (664.52, 3832.28) | 2175.50 (641.60, 3832.28) | 0.578 |
| T chol (mmol/L) | 3.85 (3.18, 4.62) | 3.85 (3.17, 4.62) | 3.84 (3.19, 4.63) | 0.495 |
| TGs (mmol/L) | 1.22 (0.90, 1.69) | 1.23 (0.91, 1.69) | 1.20 (0.90, 1.70) | 0.582 |
| HDL (mmol/L) | 1.05 (0.88, 1.25) | 1.05 (0.88, 1.25) | 1.05 (0.88, 1.26) | 0.504 |
| LDL (mmol/L) | 2.18 (1.68, 2.79) | 2.18 (1.68, 2.79) | 2.17 (1.70, 2.79) | 0.892 |
| D dimer (mg/L) | 0.45 (0.26, 1.13) | 0.45 (0.26, 1.16) | 0.44 (0.27, 1.10) | 0.301 |
| β-blocker (yes, n %) | 4506 (63.7) | 3174 (64.1) | 1332 (62.8) | 0.311 |
| CCB (yes, n %) | 1760 (24.9) | 1243 (25.1) | 517 (24.4) | 0.536 |
| Cardiac function grade, n (%) | 2.57 (0.91) | 2.57 (0.91) | 2.57 (0.90) | 0.93 |
| I | 753 (10.6) | 542 (10.9) | 211 (9.9) | 0.385 |
| II | 2843 (40.2) | 1963 (39.6) | 880 (41.5) | |
| III | 2199 (31.1) | 1553 (31.3) | 646 (30.4) | |
| IV | 1281 (18.1) | 896 (18.1) | 385 (18.1) |
Notes: Data are presented as n (%), mean ± SD, or median (IQR), as appropriate. The numbers in each cell represent subgroup counts within the training or validation set, not the total sample size of that cohort.
Abbreviations: AKI, acute kidney injury; ACEIs/ARBs, angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers; LVEF, left ventricular ejection fraction; WBC, white blood cell count; RBC, red blood cell count; Hb, haemoglobin; Plt, platelet count; CRP, C-reactive protein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; AKP, alkaline phosphatase; Alb, albumin; BUN, blood urea nitrogen; SCr, serum creatinine; Cysc, cystatin C; Glu, glucose; CtnT, cardiac troponin T; NT-proBNP, N-terminal pro-B-type natriuretic peptide; T chol, total cholesterol; TG, triglycerides; HDL, high-density lipoprotein cholesterol; LDL, low-density lipoprotein cholesterol.
Variable Selection
Results of Univariate Analysis
Univariate analysis revealed that 38 variables were significantly associated with AKI (P < 0.05). The details of all variables are presented in Table 2.
Table 2.
Univariate Analysis Results
| Variable | OR | CI_lower | CI_upper | P |
|---|---|---|---|---|
| Weight | 1.00 | 1.00 | 1.01 | < 0.001 |
| Sex | 0.84 | 0.72 | 0.99 | 0.036 |
| Age | 1.05 | 1.04 | 1.06 | < 0.001 |
| CKD | 3.29 | 2.70 | 4.01 | < 0.001 |
| Hypertension | 1.24 | 1.06 | 1.44 | 0.006 |
| Diabetes | 1.51 | 1.29 | 1.76 | < 0.001 |
| CVD | 0.45 | 0.38 | 0.53 | < 0.001 |
| COPD | 4.61 | 3.79 | 5.61 | < 0.001 |
| Arrhythmia | 1.09 | 0.92 | 1.29 | 0.316 |
| Stroke | 0.52 | 0.40 | 0.66 | < 0.001 |
| Valvular heart disease | 4.41 | 3.79 | 5.14 | < 0.001 |
| Diuretic | 2.81 | 2.34 | 3.37 | < 0.001 |
| Vasoactive drugs | 1.43 | 1.23 | 1.67 | < 0.001 |
| Cardiac function grade | 1.75 | 1.60 | 1.90 | < 0.001 |
| ACEI/ARB | 0.79 | 0.69 | 0.92 | < 0.001 |
| β-blockers | 0.70 | 0.60 | 0.81 | < 0.001 |
| Calcium channel blockers | 1.10 | 0.93 | 1.30 | 0.258 |
| Anticoagulants | 1.16 | 1.00 | 1.34 | 0.055 |
| Antiplatelet drugs | 0.67 | 0.58 | 0.78 | < 0.001 |
| LVEF | 0.08 | 0.04 | 0.14 | < 0.001 |
| WBC | 1.13 | 1.11 | 1.15 | < 0.001 |
| RBC | 0.63 | 0.57 | 0.70 | < 0.001 |
| Hb | 0.98 | 0.98 | 0.99 | < 0.001 |
| Plt | 1.00 | 1.00 | 1.00 | 0.023 |
| CRP | 1.01 | 1.00 | 1.01 | < 0.001 |
| Total bilirubin | 0.95 | 0.94 | 0.97 | < 0.001 |
| ALT | 1.00 | 1.00 | 1.00 | < 0.001 |
| AST | 1.01 | 1.00 | 1.01 | < 0.001 |
| AKP | 0.93 | 0.92 | 0.94 | < 0.001 |
| Alb | 0.95 | 0.93 | 0.96 | < 0.001 |
| BUN | 1.23 | 1.20 | 1.25 | < 0.001 |
| SCr | 1.05 | 1.05 | 1.06 | < 0.001 |
| Cysc | 4.40 | 3.85 | 5.03 | < 0.001 |
| Uric acid | 1.00 | 1.00 | 1.00 | < 0.001 |
| Glu | 1.08 | 1.06 | 1.10 | < 0.001 |
| CtnT | 1.00 | 1.00 | 1.00 | < 0.001 |
| NT-proBNP | 1.00 | 1.00 | 1.00 | < 0.001 |
| T chol | 1.08 | 1.01 | 1.15 | 0.021 |
| TG | 1.06 | 0.98 | 1.14 | 0.147 |
| HDL | 0.68 | 0.53 | 0.88 | 0.003 |
| LDL | 1.10 | 1.00 | 1.19 | 0.037 |
| D dimer | 1.07 | 1.05 | 1.08 | < 0.001 |
| Urine protein | 2.89 | 2.45 | 3.41 | < 0.001 |
Abbreviations: AKI, acute kidney injury; ACEIs/ARBs, angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers; LVEF, left ventricular ejection fraction; WBC, white blood cell count; RBC, red blood cell count; Hb, haemoglobin; Plt, platelet count; CRP, C-reactive protein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; AKP, alkaline phosphatase; Alb, albumin; BUN, blood urea nitrogen; SCr, serum creatinine; Cysc, cystatin C; Glu, glucose; CtnT, cardiac troponin T; NT-proBNP, N-terminal pro-B-type natriuretic peptide; T chol, total cholesterol; TG, triglycerides; HDL, high-density lipoprotein cholesterol; LDL, low-density lipoprotein cholesterol.
LASSO Regression Results
Using LASSO regression for feature selection, variables with an importance greater than 20% were set as the variable retention criteria. Among the 43 candidate variables, 8 with an importance exceeding 20% were selected for the LASSO regression, including diuretic use, history of hypertension, history of diabetes, history of heart valve disease, history of chronic obstructive pulmonary disease, history of chronic kidney disease, admission cystatin C level, and LVEF. These 8 variables were determined to be the candidate variables for final inclusion in the prediction model. This threshold (>20) is an empirical cut-off point set by this study and is based on the data distribution characteristics rather than the inherent default output of the LASSO algorithm. The cut-off value was set at 20%, which fully retained the core variables with the greatest contribution while excluding marginal variables whose coefficients fell below 20% and whose standard errors were larger (Figure 2).
Figure 2.
Lasso regression for variable selection. (a) Coefficient paths of the 43 candidate predictors and the cross-validation error curve. (b) Eight predictors above the study-defined importance threshold of >20% were retained and entered into the multivariable logistic regression model.
Final Variable Selection Through Multivariate Logistic Regression Analysis
The 8 variables selected by LASSO were included in the multivariable logistic regression model, and the results are shown in Table 3. Cystatin C at admission, COPD, valvular heart disease, CKD, diabetes, diuretic use, and hypertension were independent risk factors for AKI, whereas LVEF was an independent protective factor. The detailed ORs and 95% CIs are presented in Table 3.
Table 3.
Multivariate Logistics Analysis
| Variable | β | OR | 95%CI | P |
|---|---|---|---|---|
| LVEF (%) | −1.50 | 0.224 | 0.102–0.480 | <0.001 |
| Diabetes | 0.641 | 1.899 | 1.523–2.368 | <0.001 |
| CKD | 0.762 | 2.143 | 1.629–2.818 | <0.001 |
| Diuretic | 0.531 | 1.701 | 1.328–2.178 | <0.001 |
| COPD | 1.08 | 2.954 | 2.280–3.828 | <0.001 |
| Cys c(mg/L) | 1.10 | 3.013 | 2.526–3.592 | <0.001 |
| Hypertension | 0.34 | 1.405 | 1.139–1.733 | 0.0014 |
| VHD | 1.04 | 2.827 | 2.293–3.488 | <0.001 |
Abbreviations: LVEF, left ventricular ejection fraction; diabetes, history of diabetes mellitus; CKD, chronic kidney disease; diuretic, use of diuretic medication; COPD, chronic obstructive pulmonary disease; Cys c, cystatin C; hypertension, history of hypertension; VHD, history of valvular heart disease; 95%CI, 95% Confidence Interval.
Evaluation of Machine Learning Models and Selection of the Optimal Model
In the training set, 13 machine learning models, namely, RandomForest, GradientBoosting, SVM_Kernel, CATBoost, NeighborMethod, PLSModel, BoostingMethod, NeuralNet, BayesMethod, DiscriminantModel, AdaptiveBoosting, LogisticModel, and LightGBM, were constructed. The predictive performance of each model in the validation set and training set is shown in Table 4.
Table 4.
AUC and Predictive Performance of Machine Learning Models
| Model | Cohort | AUC (95% CI) | Sensitivity | Specificity | Accuracy | PPV | NPV | F1 Score | Youden’s Index |
|---|---|---|---|---|---|---|---|---|---|
| RandomForest | Training | 1.000 (1.000–1.000) | 1.000 | 0.999 | 0.999 | 0.989 | 1.000 | 0.995 | 0.999 |
| Validation | 0.788 (0.759–0.817) | 0.187 | 0.969 | 0.885 | 0.424 | 0.908 | 0.259 | 0.156 | |
| GradientBoosting | Training | 1.000 (1.000–1.000) | 1.000 | 0.996 | 0.996 | 0.968 | 1.000 | 0.984 | 0.996 |
| Validation | 0.734 (0.703–0.764) | 0.218 | 0.936 | 0.859 | 0.292 | 0.908 | 0.249 | 0.154 | |
| SVM_Kernel | Training | 1.000 (1.000–1.000) | 0.989 | 0.999 | 0.998 | 0.989 | 0.999 | 0.989 | 0.988 |
| Validation | 0.742 (0.707–0.778) | 0.182 | 0.970 | 0.885 | 0.423 | 0.908 | 0.255 | 0.152 | |
| CATBoost | Training | 1.000 (0.999–1.000) | 1.000 | 0.999 | 0.999 | 0.989 | 1.000 | 0.995 | 0.999 |
| Validation | 0.609 (0.570–0.649) | 0.209 | 0.943 | 0.864 | 0.307 | 0.908 | 0.249 | 0.152 | |
| NeighborMethod | Training | 0.999 (0.998–1.000) | 1.000 | 0.997 | 0.998 | 0.979 | 1.000 | 0.989 | 0.997 |
| Validation | 0.603 (0.574–0.632) | 0.218 | 0.967 | 0.886 | 0.445 | 0.911 | 0.293 | 0.185 | |
| PLSModel | Training | 1.000 (1.000–1.000) | 0.989 | 0.997 | 0.996 | 0.978 | 0.999 | 0.984 | 0.987 |
| Validation | 0.744 (0.711–0.777) | 0.231 | 0.969 | 0.889 | 0.473 | 0.912 | 0.310 | 0.200 | |
| BoostingMethod | Training | 1.000 (1.000–1.000) | 1.000 | 0.997 | 0.998 | 0.979 | 1.000 | 0.989 | 0.997 |
| Validation | 0.633 (0.591–0.676) | 0.218 | 0.941 | 0.863 | 0.310 | 0.909 | 0.256 | 0.159 | |
| NeuralNet | Training | 0.999 (0.997–1.000) | 1.000 | 0.999 | 0.999 | 0.989 | 1.000 | 0.995 | 0.999 |
| Validation | 0.583 (0.557–0.610) | 0.196 | 0.978 | 0.894 | 0.518 | 0.910 | 0.284 | 0.174 | |
| BayesMethod | Training | 0.999 (0.996–1.000) | 0.478 | 0.999 | 0.943 | 0.978 | 0.941 | 0.642 | 0.477 |
| Validation | 0.751 (0.718–0.785) | 0.142 | 0.988 | 0.897 | 0.582 | 0.905 | 0.229 | 0.130 | |
| DiscriminantModel | Training | 0.998 (0.996–1.000) | 1.000 | 0.996 | 0.996 | 0.968 | 1.000 | 0.984 | 0.996 |
| Validation | 0.636 (0.589–0.684) | 0.218 | 0.933 | 0.856 | 0.283 | 0.908 | 0.246 | 0.151 | |
| AdaptiveBoosting | Training | 0.999 (0.997–1.000) | 1.000 | 0.999 | 0.999 | 0.989 | 1.000 | 0.995 | 0.999 |
| Validation | 0.650 (0.607–0.693) | 0.196 | 0.960 | 0.878 | 0.373 | 0.908 | 0.257 | 0.156 | |
| LogisticModel | Training | 1.000 (1.000–1.000) | 1.000 | 0.999 | 0.999 | 0.989 | 1.000 | 0.995 | 0.999 |
| Validation | 0.804 (0.777–0.830) | 0.218 | 0.939 | 0.861 | 0.301 | 0.909 | 0.253 | 0.157 | |
| LightGBM | Training | 0.969 (0.954–0.984) | 0.000 | 1.000 | 0.893 | N/A | 0.893 | N/A | 0.000 |
| Validation | 0.713 (0.678–0.749) | 0.000 | 1.000 | 0.892 | N/A | 0.892 | N/A | 0.000 |
The discriminative ability of each model was evaluated using ROC curves and AUC (Figure 3). All variable selection and model training were performed exclusively in the training set using 10-fold cross-validation, and all included predictors were ascertained before AKI occurrence. The validation set was used only for model evaluation. In the training set, except for LightGBM (AUC = 0.969), the AUCs of the other 12 models were ≥0.998, with the logistic model reaching 1.000 (95% CI: 1.000–1.000). This unusually high training-set performance suggests substantial overfitting rather than good generalization. Therefore, the validation set was the core basis for evaluating generalization ability. In the validation set, the logistic model performed best, with an AUC of 0.804 (95% CI: 0.777–0.830), which was superior to that of the random forest (0.788), Bayes method (0.751), and other models. Although the AUC of each model in the validation set was lower than that in the training set, the substantial drop in the logistic model from 1.000 to 0.804 suggests limited generalizability. Given the large sample size and the relatively low incidence of AKI (11.2%), the resulting class imbalance limited the sensitivity of all models in the validation set, which in turn contributed to the relatively low Youden index and F1 scores. In terms of the other validation set performance metrics, the PLSModel achieved the highest Youden index (0.200) and the highest sensitivity (0.231), although its AUC (0.744) was inferior to that of the logistic model. The Bayes method achieved the highest specificity (0.988) and PPV (0.582), but its sensitivity was notably low (0.142), indicating a tendency to predict the majority of cases as negative. Overall, while the logistic model did not rank first in every individual metric, it provided the highest validation-set AUC and was therefore selected as the final model.
Figure 3.
Receiver operating characteristic (ROC) curves for the 13 machine learning models in the training set (a) and validation set (b). In the training set, most models achieved near-perfect AUC (>0.998), indicating good fit but with a risk of overfitting. In the validation set, logistic regression achieved the highest AUC of 0.804 (95% CI: 0.777–0.830) and was selected as the optimal model.
Model calibration was also assessed. The calibration curve was used to visually evaluate calibration (Figure 4). In the validation set, the Hosmer–Lemeshow test P value was 0.605 (>0.05), the calibration intercept was 0.013, and the calibration slope was 0.835. The Brier score was 0.082 in the validation set and 0.078 in the training set, with a difference of only 0.004. These quantitative metrics indicated that the overall calibration of the model was acceptable. The calibration curve in the validation set showed a mild staircase pattern, with a tendency to overestimate risk in the lower predicted probability range. This is likely related to the low incidence of AKI (11.2%) and the resulting class imbalance, which leads to sparse positive events within probability strata and unstable observed event rates, a phenomenon commonly seen in prediction models for low-event-rate outcomes. This model is intended for risk stratification rather than precise estimation of individual probabilities, and its overall discriminative ability and calibration metrics support its use as a candidate risk-stratification tool.
Figure 4.
Calibration curves of the logistic regression model. (a) Calibration curve in the training set. (b) Calibration curve in the validation set. In the validation set, the Hosmer–Lemeshow test P‑value was 0.605, the calibration intercept was 0.013, the calibration slope was 0.835, and the Brier score was 0.082, indicating acceptable calibration.
The clinical practicality of each model was evaluated through decision curve analysis (DCA) (Figure 5). In the training set, the net benefit of the logistic model remained stable at approximately 0.100 across a wide threshold range and was consistently superior to the “treat none” strategy. In the validation set, the net benefit of the logistic model remained positive at thresholds below approximately 0.28 but fell below zero at higher thresholds. This threshold-dependent pattern indicates that the model’s clinical utility is limited at higher thresholds requiring high specificity. This pattern is closely related to the low incidence of AKI (11.2%) in our cohort: under low event rates, class imbalance leads to overestimation of risk in the lower predicted probability range, which is consistent with the staircase pattern observed in the calibration curve and jointly explains the negative net benefit at higher thresholds. Therefore, the DCA results suggest that the model should not be regarded as an early risk screening tool. Instead, it may serve as a candidate tool for risk stratification in elderly patients with heart failure, helping to identify high-risk individuals who may benefit from targeted preventive interventions, pending further external validation.
Figure 5.
Decision curve analysis (DCA) of the 13 machine learning models. (a) DCA in the training set. (b) DCA in the validation set. In the validation set, logistic regression showed a positive net benefit at thresholds below approximately 0.28, indicating potential clinical utility mainly in low-threshold screening scenarios; at higher thresholds, the net benefit fell below zero.
Nomogram Construction
On the basis of the determination that logistic regression was the optimal model, this study further constructed a nomogram to predict the risk of AKI in elderly patients with heart failure (Figure 6). The visualization of the prediction model was achieved by converting the regression coefficients of the logistic regression model into an intuitive scoring system. This nomogram included 8 predictors in the final model, namely, diuretic use, hypertension, diabetes, valvular heart disease, chronic obstructive pulmonary disease(COPD), chronic kidney disease, admission cystatin C level, and LVEF. Each predictor was assigned a corresponding score on the basis of its regression coefficient, and the individualized AKI risk prediction value of the patient could be intuitively read by calculating the total score, providing a simple and intuitive bedside risk stratification tool for clinical use. Notably, as a protective factor, the LVEF had a lower correlation with a higher score, indicating an increase in AKI risk, which was completely consistent with the negative correlation between the LVEF and AKI occurrence in the logistic regression model.
Figure 6.
Nomogram for predicting AKI risk in hospitalized elderly patients with heart failure. The nomogram includes eight predictors: diuretic use, hypertension, diabetes, valvular heart disease, COPD, CKD, admission cystatin C, and LVEF. With respect to the LVEF, lower values corresponded to higher points and thus higher AKI risk, which is consistent with its protective effect in the logistic regression model.
Score System Construction
To further improve the clinical practicality of the prediction model, an individualized risk scoring system based on a multivariate logistic regression model was constructed (Table 5). This scoring system achieved visualization and simplified application of the model by converting the regression coefficients of each predictor into integer scores. Cystatin C was divided into three groups on the basis of the clinical reference range and commonly used cut-off points in previous studies: <1.0, 1.0–1.5, and ≥1.5 mg/L. These groups were assigned 0, 4, and 7 points, respectively.13 The LVEF was divided into three groups according to the 2021 ESC guidelines: <40%, 40%–49%, and ≥50%, and patients with a <40% LVEF were assigned 7 points. The specific scoring for the other categorical variables is detailed in Table 5. The total risk score was obtained by adding the scores of each factor, and a higher score indicated a greater AKI risk, providing a simple bedside risk stratification tool for clinical use.
Table 5.
Risk Assessment Scale
| Risk factor | Addition to risk score | Risk score |
|---|---|---|
| Cys C(mg/L) | <1.0 1.0–1.5 ≥1.5 0 4 7 |
|
| LVEF | ≥0.5 0.4–0.5 <0.4 0 4 7 |
|
| CKD | No 0 Yes 3 | |
| COPD | No 0 Yes 3 | |
| Diuretic Use | No 0 Yes 2 | |
| Hypertension | No 0 Yes 2 | |
| VHD | No 0 Yes 1 | |
| Diabetes | No 0 Yes 1 | |
| Total risk score= |
Abbreviations: LVEF, left ventricular ejection fraction; diabetes, history of diabetes mellitus; CKD, chronic kidney disease; diuretic, use of diuretic medication; COPD, chronic obstructive pulmonary disease; Cys c, cystatin C; hypertension, history of hypertension; VHD, history of valvular heart disease.
Discussion
This study was based on the clinical data of 7076 elderly patients with heart failure who were hospitalized at the General Hospital of the Northern Theater Command. All predictors were restricted to data available within 24 hours of admission and prior to AKI occurrence. The risk factors for AKI in elderly patients with heart failure during hospitalization were systematically analysed, and a risk prediction model was constructed. Single-factor analysis, LASSO regression, and multivariate logistic regression analyses were used to determine the independent risk factors for AKI in elderly patients with heart failure during hospitalization. Through multiple screening strategies, the predictive variables were streamlined to the greatest extent possible, improving the accuracy of the prediction and avoiding overfitting of the prediction model. After multiple screenings, a total of 8 characteristic factors were included in the final multivariate logistic regression model and subsequent nomograms and scoring systems, including diuretic use, history of hypertension, history of diabetes, history of valvular heart disease, history of COPD, history of chronic kidney disease, admission cystatin C level, and LVEF. In this study, 13 machine learning models were constructed, and the predictive value of each model for elderly heart failure AKI patients was compared. On the basis of the comprehensive prediction performance and stability, the LR model had an AUC of 1.000 in the training set and 0.804 in the validation set, with good model stability, and was determined to be the best risk prediction model for AKI in elderly hospitalized patients with heart failure. Nomograms and risk score systems constructed based on the LR model can be used as candidate tools for the early identification of high-risk patients in clinical practice.
Multiple studies have shown that AKI typically occurs in the early stages of acute cardiac events, and early identification of AKI risk is crucial for predicting patient outcomes.14 In hospitalized elderly patients, concomitant heart failure is among the major risk factors for AKI.15,16 Compared with the general population, elderly patients with heart failure have a greater risk of developing AKI.17 Heart failure combined with acute kidney injury significantly increases hospital stay and the risk of death within one year.18,19
In recent years, some scholars have focused on developing predictive models to identify high-risk populations for AKI, with related studies covering diverse disease backgrounds and clinical settings. Yuqi Yang and colleagues developed a model to predict AKI in a cohort of heart failure patients and reported that older age, poorer baseline renal function, and more severe heart failure were associated with a greater probability of moderate to severe AKI.20 Recently, several researchers have introduced deep learning technology and have used long short-term memory networks to establish a dynamic prediction model for AKI in patients with heart failure. The results revealed that urine volume, SOFA score and systolic blood pressure play dominant roles in short-term prediction, whereas troponin T levels and a history of cardiovascular surgery are more important for long-term prediction.21 Xi Peng et al utilized the MIMIC-III database to develop a prediction model for AKI in patients with congestive heart failure. The model, centred on predictors such as serum creatinine, blood urea nitrogen, and urine output, demonstrated strong discriminatory performance.22 In the field of cardiac surgery, for patients undergoing coronary artery bypass grafting with reduced left ventricular ejection fraction, researchers have developed a perioperative AKI prediction nomogram that includes variables such as female sex, preoperative elevated creatinine levels, an LVEF < 35%, previous myocardial infarction, hypertension, the use of extracorporeal circulation, and perioperative blood transfusion.23 Li Zhao et al established a predictive model for AKI in elderly patients in the ICU. On the basis of the results of multivariate logistic analysis, age, weight, urine volume, SOFA score, and SAPS II score were included as predictive factors, as these factors had good predictive effects.24 With respect to elderly inpatients, Kayatas et al reported that advanced age, low diastolic blood pressure, high CRP level and low haemoglobin level were independent risk factors for AKI.25 Furthermore, some studies have employed an ensemble learning strategy that integrates multiple machine learning algorithms. Heart failure, coronary heart disease, sepsis, and respiratory failure are important predictors of AKI in elderly hospitalized patients.26 Although these models have achieved good predictive performance, they mainly target general heart failure patients or ICU inpatients. However, this model is specifically designed for elderly patients aged 65 years or older with heart failure. This group faces the dual risks of decreased renal function reserve due to ageing and heart dysfunction. This population has not been adequately covered in previous studies. Methodologically, this research initially screened variables through LASSO regression and then further adjusted for confounding factors using multivariate logistic stepwise regression, which not only controlled for overfitting risks but also enhanced the clinical interpretability of the predictive factors. In terms of the composition of predictive factors, this model incorporates variable combinations that are closer to the clinical characteristics of elderly heart failure patients, which can supplement previous models.
Multiple studies have shown that in patients with heart failure, both baseline renal impairment and the occurrence and development of AKI are closely associated with poor prognosis. A history of chronic kidney disease, hypertension, advanced age, and the use of diuretics are considered important risk factors, which is consistent with our findings.18 Butler J et al reported that risk factors for AKI in patients with acute heart failure include the use of diuretics and serum creatinine levels at admission.27 Kevin D et al’s secondary analysis of 57 clinical studies further emphasized that chronic kidney disease, advanced age, and diuretic use are key predictors of AKI in patients with acute heart failure and that once AKI occurs, the mortality rate of patients increases significantly.28 Early renal function impairment is an independent predictor of all-cause mortality in patients, and the use of serum creatinine alone to predict early AKI is not accurate.29 Cystatin C is not affected by muscle mass and can more accurately reflect the eGFR. Its predictive value for early AKI in elderly patients is greater than that of serum creatinine.30 A large study of 80,000 elderly patients aged 65 and above revealed that estimates of the glomerular filtration rate based on cystatin C were more accurate at predicting the risk of adverse outcomes in elderly patients than those based on creatinine were.31
Diuretics are important for reducing volume overload in patients with heart failure and are among the standard treatments for heart failure at our hospital. However, diuretic use may lead to decreased renal perfusion pressure, activate the renin angiotensin aldosterone system (RAAS) and the sympathetic nervous system, worsen renal ischaemia, and negatively affect kidney function.32 For patients with diuretic resistance or impaired renal function, blindly increasing the dose of diuretics may aggravate renal damage. In such cases, ultrafiltration therapy can be considered. The ESC guidelines recommend loop diuretics as the first-line treatment. The dosage should be adjusted dynamically on the basis of the patient’s condition. When loop diuretic resistance occurs, thiazide drugs can be used in combination. If the congestion does not subside after diuretic treatment and if there is a progressive deterioration of renal function or severe electrolyte imbalance, ultrafiltration or renal replacement therapy should be considered.33 When diuretics are used in elderly patients with heart failure, serum creatinine, urea nitrogen and electrolytes should be closely monitored, and treatment plans should be adjusted promptly on the basis of the monitoring results to prevent and manage acute kidney injury.34
Reduced cardiac output in patients with heart failure can lead to insufficient renal perfusion, and venous congestion can easily cause renal oedema, resulting in renal dysfunction and activation of the neuroendocrine system.29 Studies have shown that patients with a low left ventricular ejection fraction (LVEF) have a greater risk of developing AKI after cardiac surgery.35 Low LVEF values are an independent risk factor for AKI in patients with acute myocardial infarction who are undergoing coronary intervention.36 This study also revealed that a lower LVEF was significantly associated with the occurrence of AKI during hospitalization in elderly patients with heart failure. These findings further confirmed the important role of decreased LVEF in the onset of AKI. Valvular heart disease such as aortic stenosis or mitral regurgitation can directly affect cardiac output and renal perfusion pressure. Aortic valve stenosis leads to an increase in the afterload of the left ventricle and a decrease in cardiac output, which in turn causes a reduction in renal perfusion and an increase in renin secretion.37 Mitral valve regurgitation leads to an increase in the volume load on the left heart, and elevated venous pressure can further impair renal perfusion, thereby creating a vicious cycle in which both cardiac and renal functions deteriorate.38 Moreover, AKI remains a common and serious perioperative complication for patients who require radical treatments such as surgical heart valve replacement or transcatheter aortic valve implantation.39
In this study, the New York Cardiac Function Classification (NYHA) and NT-proBNP did not have independent predictive value for acute kidney injury, which may be related to the fact that all included patients were elderly patients with heart failure. In such patients, the mechanism of AKI is influenced by mechanisms such as renal perfusion levels, venous congestion, and other comorbidities.40 Therefore, the use of the NYHA classification, as an indicator of subjective symptoms and clinical assessment, has limited predictive power in this study population, whereas the use of the LVEF, as an indicator of cardiac pumping function and renal perfusion, has greater predictive value for renal function changes. Previous studies have shown that the basal NT-proBNP levels in patients with acute exacerbation of chronic heart failure are inherently very high. At high baseline levels, the value of NT-proBNP in differentiating certain complications or prognosis decreases.41
In elderly patients with heart failure, the relationship between arrhythmia and AKI is complex and mostly bidirectional.42 A systematic review involving 736,210 patients revealed that the risk of new-onset atrial fibrillation in AKI patients increased by 94% (RR=1.94) in the short term. Arrhythmia may be more a manifestation of the intensification of heart–kidney interactions after the occurrence of AKI than the original driving factor for AKI.43 In addition, effective methods such as catheter ablation for controlling heart rhythm can improve haemodynamics without worsening renal function, indicating that when the rhythm is stable or effectively controlled, arrhythmia itself may not significantly increase the risk of AKI.44 Therefore, although arrhythmia may be involved in the occurrence of AKI in a situation of severe haemodynamic instability, in this study, in elderly patients with heart failure due to relatively stable rhythms during the disease course or who received effective control, arrhythmia was not an independent predictor of AKI.
Hypertension and diabetes, as common chronic comorbidities in elderly patients, have become common risk factors for acute kidney injury in elderly patients with heart failure.5 Hypertension leads to increased vascular resistance, reduced renal blood flow, and poor vascular elasticity in elderly patients, making them more prone to renal function impairment.45 Treatment plans for hypertension in elderly patients should be chosen with caution, as excessive intensification of blood pressure reduction is also likely to induce AKI.46 Diabetic nephropathy is the most common microvascular lesion of diabetes and affects mainly the glomerular filtration rate and reduces the reserve function and self-regulation ability of the kidneys. Therefore, patients with diabetic nephropathy are more prone to acute kidney injury.47 For elderly diabetic patients, drugs with heart–kidney protective effects, such as sodium‒glucose cotransporter 2 (SGLT2) inhibitors, mineralocorticoid receptor antagonists (MRAs), and glucagon-like peptide-1 (GLP-1) inhibitors, are recommended.40 Chronic obstructive pulmonary disease (COPD) is a common disease among elderly patients and is characterized by irreversible airflow limitation associated with chronic inflammation of the lung parenchyma and surrounding airways.48 In this study, COPD was identified as a risk factor for AKI in elderly patients with heart failure. This finding is consistent with the findings of previous studies in which the incidence of AKI in COPD patients was significantly greater. Further analysis indicated that renal function impairment in COPD patients might be related to factors such as an inflammatory state, pulmonary hypertension, and hypercapnia.49,50
The predictive model constructed in this study provides clinicians with a simple and intuitive tool that can quickly stratify the risk of AKI in elderly patients with heart failure upon admission. All the predictive factors in the model are routine clinical data available within 24 hours of admission. Without the need for additional tests or complex calculations, clinicians can map individual patient indicators onto a total score using the nomogram, thereby intuitively determining the AKI risk level. The core clinical value of this model lies in identifying high-risk populations and guiding early intervention. For patients with a score indicating high risk, targeted preventive measures can be taken in the early stage of admission. Strictly adhere to the indications for the use of diuretics to avoid excessive dehydration leading to insufficient renal perfusion. Urine volume and changes in renal function should be dynamically monitored. Prudently adjust antihypertensive regimens to avoid excessive intensification of antihypertensive treatment, which can induce AKI.46 For patients with underlying diseases such as diabetes and CKD, renal function monitoring should be increased, and hypoglycaemic drugs with cardiorenal protective effects, such as SGLT2 inhibitors, MRAs, and GLP-1 inhibitors, should be preferentially selected.51 For elderly patients with heart failure who have also been diagnosed with COPD, early identification and active treatment of pulmonary infections, control of systemic inflammatory responses, and improvement of oxygenation status should be carried out. This will help reduce the risk of AKI.49 In addition, for patients with low-risk scores, the monitoring frequency for AKI can be appropriately relaxed to optimize the allocation of medical resources and avoid unnecessary excessive monitoring. By combining patient risk scores with comprehensive clinical assessments, doctors can formulate individualized AKI prevention strategies, thereby minimizing the risk of AKI occurrence and its adverse impact on prognosis.52 Moreover, the use of a nomogram can aid in doctor–patient communication, explain the condition to patients and their families through intuitive risk visualization, and improve treatment compliance.
Although this study used 10-fold cross-validation, kept the training and validation sets completely independent, and restricted all included predictors to data ascertained before AKI occurrence, thereby reducing the risk of time-related outcome leakage, several issues still warrant caution. First, the unusually high AUCs in the training set, with the logistic model reaching 1.000, are more likely to reflect complete separation or overfitting than true generalizability. Therefore, the training-set performance should not be overinterpreted. In the validation set, the logistic model achieved an AUC of 0.804, and the calibration curve indicated acceptable but not excellent calibration. Decision curve analysis further showed a positive net benefit at lower thresholds but a net benefit below zero at higher thresholds, suggesting that the model is more suitable as a candidate risk-stratification tool for elderly patients with heart failure than as an early risk-screening tool. At present, the model is not yet ready for routine clinical implementation. Future studies should incorporate additional clinically relevant indicators. In addition, residual confounding by concomitant medications may exist, and further studies, including pharmacokinetic investigations, are needed to more accurately evaluate the relationship between drug exposure and AKI risk.
This study has several limitations. First, this was a single-centre retrospective study, and the patient population included may be specific, which may have led to selection bias in the research results. Second, this study focused only on elderly hospitalized patients with heart failure, and the predictive efficacy of the model in other populations needs to be further evaluated. Third, although the risk scoring system constructed in this study is based on a multivariate logistic regression model, the scoring criteria in different clinical scenarios still need to be further optimized and verified in a large sample population. Fourth, patients with missing data were excluded from the analysis, and no imputation was performed. This approach may have introduced some degree of selection bias, although its impact on the overall findings is likely to be limited.
Conclusion
In this study, an AKI risk prediction model for elderly heart failure patients was constructed using single-factor analysis, LASSO regression, and multivariate logistic regression, which included 8 predictive factors. Among the 13 machine learning models, logistic regression performed stably in the validation set. Despite some overfitting, the model demonstrated acceptable performance in the validation set and was selected as the optimal model for this study. The model showed acceptable discrimination and calibration and may serve as a candidate tool for early risk stratification in elderly heart failure patients, pending further external validation by multicenter and prospective cohort studies.
Acknowledgments
We thank all the doctors and patients who provided data to support this study. We sincerely thank the following individuals for their valuable contributions to data collection and study support: Panpan Fu, Chen Zhao, Yanping Zhang, Zhuo Ren, Qian Wang, and Kaiming Ren. Their efforts are greatly appreciated.
Funding Statement
This study was supported by the Liaoning Provincial Science and Technology Plan Joint Fund (2023-MSLH-350) and the Liaoning Provincial Science and Technology Plan Joint Fund (2023-MSLH-358).
Data Sharing Statement
All the data generated or analysed during this study are available from the corresponding author upon reasonable request. Requests should be directed to Ning Cao (bzxyjhk@126.com).
Statement of Ethics
The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the General Hospital of Northern Theater Command (Approval No. Y(2025)364). Informed consent was waived because of the retrospective nature of the study. Patient data were deidentified, stored on a password-protected server with restricted access, and reported in aggregate form in compliance with the Personal Information Protection Law of the People’s Republic of China and relevant confidentiality regulations.
Author Contributions
Mingyu Liu and Dongxia Qiu contributed equally to the work and should be regarded as co-first authors. All the authors made a significant contribution to the work reported, whether it is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.
Disclosure
No potential conflicts of interest were reported by the author(s).
References
- 1.Ponikowski P, Voors A A, Anker S D, et al. ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: the Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (Esc) developed with the special contribution of the heart failure Association (HFA) of the ESC. Eur Heart J. 2016;37(27):2129–18. [DOI] [PubMed] [Google Scholar]
- 2.Patschan D, Drubel K, Matyukhin I, et al. Kidney replacement therapy in cardiorenal syndromes. J Clin Med Res. 2022;14(7):264–272. doi: 10.14740/jocmr4676 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Chen JJ, Lee TH, Kuo G, et al. Acute kidney disease after acute decompensated heart failure. Kidney Int Rep. 2022;7(3):526–536. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Holgado JL, Lopez C, Fernandez A, et al. Acute kidney injury in heart failure: a population study. ESC Heart Fail. 2020;7(2):415–422. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Stille K, Kribben A, Herget-Rosenthal S. Incidence, severity, risk factors and outcomes of acute kidney injury in older adults: systematic review and meta-analysis. J Nephrol. 2022;35(9):2237–2250. doi: 10.1007/s40620-022-01381-2 [DOI] [PubMed] [Google Scholar]
- 6.Saran R, Robinson B, Abbott K C, et al. US renal data system 2018 annual data report: epidemiology of kidney disease in the United States. Am J Kidney Dis. 2019;73(3 Suppl 1):A7–a8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Kate RJ, Perez RM, Mazumdar D, et al. Prediction and detection models for acute kidney injury in hospitalized older adults. BMC Med Inf Decis Mak. 2016;16:39. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Wang L, Zhao Y T. Development and validation of a prediction model for acute kidney injury among patients with acute decompensated heart failure. Front Cardiovasc Med. 2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Kuno T, Mikami T, Sahashi Y, et al. Machine learning prediction model of acute kidney injury after percutaneous coronary intervention. Sci Rep. 2022;12(1):749. doi: 10.1038/s41598-021-04372-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Wang Q, Tang Y, Zhou J, et al. A prospective study of acute kidney injury in the intensive care unit: development and validation of a risk prediction model. J Transl Med. 2019;17:17. doi: 10.1186/s12967-019-2118-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Foundation N K. Kidney disease: improving global out comes. KDIGO clinical practice guideline foranaemia in chronic kidney disease. 2012:279–335. [Google Scholar]
- 12.McDonagh T A, Metra M, Adamo M, et al. Focused update of the 2021 ESC guidelines for the diagnosis and treatment of acute and chronic heart failure. Eur Heart J. 2023;44(37):3627–3639. [DOI] [PubMed] [Google Scholar]
- 13.Wang K, An Y, Zhou J, et al. A novel multi-level feature selection method for radiomics. Alex Eng J. 2022;66:993–999. doi: 10.1016/j.aej.2022.10.069 [DOI] [Google Scholar]
- 14.Marrani C, Zenjelaj T, Bartoli D, et al. Role of cystatin C as a biomarker of acute kidney injury and as an independent long-term predictor of cardiovascular events, mortality and functional outcome. Italian J Med. 2013. [Google Scholar]
- 15.Ismail Y, Kasmikha Z, Green H L, et al. Cardio-renal syndrome type 1: epidemiology, pathophysiology, and treatment. Semin Nephrol. 2012;32(1):18–25. [DOI] [PubMed] [Google Scholar]
- 16.Maciel CG, Vasconcelos E M R, Cavalcanti B, et al. Factors associated with acute kidney injury in hospitalized elderly: an integrative review. Cien Saude Colet. 2025;30(suppl 2):e14762023. [DOI] [PubMed] [Google Scholar]
- 17.Funahashi Y, Chowdhury S, Eiwaz M B, et al. Acute Cardiorenal Syndrome: models and Heart-Kidney Connectors. Nephron. 2020;144(12):629–633. doi: 10.1159/000509353 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zhang W, Wen D, Zou Y F, et al. One-year survival and renal function recovery of acute kidney injury patients with chronic heart failure. Cardiorenal Med. 2015;5(1):40–47. doi: 10.1159/000369834 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Damman K, Valente M A, Voors A A, et al. Renal impairment, worsening renal function, and outcome in patients with heart failure: an updated meta-analysis. Eur Heart J. 2014;35(7):455–469. [DOI] [PubMed] [Google Scholar]
- 20.Chetcuti S, Masengu A. Outcomes in acute kidney injury requiring haemodialysis - a retrospective cohort study. Ulster Med J. 2025;94(2):77–82. doi: 10.1186/cc7427 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Yang YQ, Da JJ, Nie S, et al. A prediction model for moderate to severe acute kidney injury in people with heart failure. Mil Med Res. 2024;11(1):57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Wang L, Zhou D, Xu P. A deep learning model for dynamic prediction of acute kidney injury in heart failure patients. Can J Cardiol. 2026;42(7):1605–1615. doi: 10.1016/j.cjca.2026.01.011 [DOI] [PubMed] [Google Scholar]
- 23.Peng X, Li L, Wang X, et al. A machine learning-based prediction model for acute kidney injury in patients with congestive heart failure. Front Cardiovasc Med. 2022;9:842873. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Lin H, Hou J, Tang H, et al. A novel nomogram to predict perioperative acute kidney injury following isolated coronary artery bypass grafting surgery with impaired left ventricular ejection fraction. BMC Cardiovasc Disord. 2020;20(1). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Zhao L, Li X, Zhao W, et al. Development and validation of a nomogram for predicting acute kidney injury in elderly patients in intensive care unit. Ren Fail. 2025;47(1):2499911. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Kayatas K, Sahin G, Tepe M, et al. Acute kidney injury in the elderly hospitalized patients. Ren Fail. 2014;36(8):1273–1277. [DOI] [PubMed] [Google Scholar]
- 27.Kate RJ, Perez RM, Mazumdar D, et al. Prediction and detection models for acute kidney injury in hospitalized older adults. Bmc Med Inf Decis. 2016;16(1):1–11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Butler J. Relationship between heart failure treatment and development of worsening renal function among hospitalized patients1, *1. Am Heart J. 2004;147(2):331–338. doi: 10.1016/S0002-8703(03)00708-7 [DOI] [PubMed] [Google Scholar]
- 29.Damman K, Valente M A E, Voors A A, O’Connor CM, van Veldhuisen DJ, Hillege HL. Renal impairment, worsening renal function, and outcome in patients with heart failure: an updated meta-analysis. Eur Heart J. 2014;35(7):455–469. doi: 10.1093/eurheartj/eht386 [DOI] [PubMed] [Google Scholar]
- 30.Presume J, Cunha G J L, Rocha B M L, et al. Acute kidney injury patterns in acute heart failure: the prognostic value of worsening renal function and its timing. Rev Port Cardiol. 2023;42(5):423–430. [DOI] [PubMed] [Google Scholar]
- 31.Wang N, Han F, Pan J, et al. Serum Cys C predicts acute kidney injury in patients with acute pancreatitis: a retrospective study. Arab J Gastroenterol. 2023;24(4):238–244. [DOI] [PubMed] [Google Scholar]
- 32.Fu EL, Carrero JJ, Sang Y, et al. Association of low glomerular filtration rate with adverse outcomes at older age in a large population with routinely measured cystatin C. Ann Intern Med. 2024;177(3):269–279. doi: 10.7326/M23-1138 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Jentzer JC, DeWald T A, Hernandez A F. Combination of loop diuretics with thiazide-type diuretics in heart failure. J Am Coll Cardiol. 2010;56(19):1527–1534. doi: 10.1016/j.jacc.2010.06.034 [DOI] [PubMed] [Google Scholar]
- 34.Mullens W, Damman K, Harjola V P, et al. The use of diuretics in heart failure with congestion - a position statement from the heart failure Association of the European Society of Cardiology. Eur J Heart Fail. 2019;21(2):137–155. [DOI] [PubMed] [Google Scholar]
- 35.McDonagh T A, Metra M, Adamo M, et al. Corrigendum to: 2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: developed by the Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC) with the special contribution of the heart failure Association (HFA) of the ESC. Eur Heart J. 2021;42(48):4901. [DOI] [PubMed] [Google Scholar]
- 36.Lin H, Hou J, Tang H, et al. A novel nomogram to predict perioperative acute kidney injury following isolated coronary artery bypass grafting surgery with impaired left ventricular ejection fraction. BMC Cardiovasc Disord. 2020;20(1):517. doi: 10.1186/s12872-020-01799-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Tan J, Zhang Y H, Si J, et al. Incidence, predictors and prognosis of acute kidney injury in acute ST-segment elevation myocardial infarction patients undergoing emergent coronary angiography/primary percutaneous coronary intervention. J Geriatr Cardiol. 2023;20(2):139–149. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Barton CH, Ni Z, Vaziri N D. Effect of severe aortic banding above the renal arteries on nitric oxide synthase isotype expression. Kidney Int. 2001;59(2):654–661. doi: 10.1046/j.1523-1755.2001.059002654.x [DOI] [PubMed] [Google Scholar]
- 39.Kuwata S, Izumo M, Okuno T, et al. Impact of renal congestion in patients with secondary Mitral regurgitation after Mitral transcatheter Edge-to-Edge repair. Circ J. 2024;88(4):510–516. doi: 10.1253/circj.CJ-23-0240 [DOI] [PubMed] [Google Scholar]
- 40.Ram P, Mezue K, Pressman G, et al. Acute kidney injury post-transcatheter aortic valve replacement. Clin Cardiol. 2017;40(12):1357–1362. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Rangaswami J, Bhalla V, Blair J E A, et al. Cardiorenal Syndrome: classification, pathophysiology, diagnosis, and treatment strategies: a Scientific statement from the American Heart Association. Circulation. 2019;139(16):e840–e878. doi: 10.1161/CIR.0000000000000664 [DOI] [PubMed] [Google Scholar]
- 42.Anwaruddin S, Lloyd-Jones D M, Baggish A, et al. Renal function, congestive heart failure, and amino-terminal pro-brain natriuretic peptide measurement: results from the ProBNP investigation of Dyspnea in the emergency Department (PRIDE) study. J Am Coll Cardiol. 2006;47(1):91–97. [DOI] [PubMed] [Google Scholar]
- 43.Genovesi S, Regolisti G, Burlacu A, et al. The conundrum of the complex relationship between acute kidney injury and cardiac arrhythmias. Nephrol Dial Transpl. 2023;38(5):1097–1112. [DOI] [PubMed] [Google Scholar]
- 44.De Clercq L, Ailliet T, Schaubroeck H, et al. Acute and chronic cardiovascular consequences of acute kidney injury: a systematic review and meta-analysis. Cardiorenal Med. 2023;13(1):26–33. doi: 10.1159/000527198 [DOI] [PubMed] [Google Scholar]
- 45.Marrouche NF, Brachmann J, Andresen D, et al. Catheter ablation for atrial fibrillation with heart failure. N Engl J Med. 2018;378(5):417–427. doi: 10.1056/NEJMoa1707855 [DOI] [PubMed] [Google Scholar]
- 46.Chang-Panesso M. Acute kidney injury and aging. Pediatr Nephrol. 2021;36(10):2997–3006. doi: 10.1007/s00467-020-04849-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Drawz PE, Lenoir KM, Rai NK, et al. Effect of intensive blood pressure control on kidney outcomes: long-term electronic health record-based post-trial follow-up of sprint. Clin J Am Soc Nephrol. 2024;19(2):213–223. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48.Wu G, Ye Y, Xu M, et al. Influencing factors of acute kidney injury in elderly patients with diabetic nephropathy and establishment of nomogram model. Front Endocrinol. 2024;15:1431873. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 49.Barnes P J. Inflammatory mechanisms in patients with chronic obstructive pulmonary disease. J Allergy Clin Immunol. 2016;138(1):16–27. doi: 10.1016/j.jaci.2016.05.011 [DOI] [PubMed] [Google Scholar]
- 50.Jiang S, Xu L, Wang X, et al. Risk prediction for acute kidney disease and adverse outcomes in patients with chronic obstructive pulmonary disease: an interpretable machine learning approach. Ren Fail. 2025;47(1):2485475. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Lee SA, Cozzi M, Bush E L, et al. Distant organ dysfunction in acute kidney injury: a review. Am J Kidney Dis. 2018;72(6):846–856. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 52.Rangaswami J, Bhalla V, Blair JE, et al. Cardiorenal Syndrome: classification, pathophysiology, diagnosis, and treatment strategies: a Scientific statement from the American Heart Association. Circulation. 2019;139. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
All the data generated or analysed during this study are available from the corresponding author upon reasonable request. Requests should be directed to Ning Cao (bzxyjhk@126.com).