Development and validation of a model for predicting acute kidney injury after cardiac surgery in patients of advanced age

Abstract Objective To develop a clinical model for predicting postoperative acute kidney injury (AKI) in patients of advanced age undergoing cardiac surgery. Methods A total of 848 patients (aged ≥ 60 years) undergoing cardiac surgery were consecutively enrolled. Among them, 597 were randomly selected for the development set and the remaining 251 for the validation set. AKI was the primary outcome. To develop a model for predicting AKI, visualized as a nomogram, we performed logistic regression with variables selected by Lasso regression analysis. The discrimination, calibration, and clinical usefulness of the new model were assessed and compared with those of Cleveland Clinic score and Simplified Renal Index (SRI) score in the validation set. Results The incidence of AKI was 61.8% in the development set. The new model included seven variables including preoperative serum creatinine, hypertension, preoperative uric acid, New York Heart Association classification ≥ 3, cardiopulmonary bypass time > 120 min, intraoperative red blood cell transfusion, and postoperative prolonged mechanical ventilation. In the validation set, the areas under the receiver operating characteristic curves for assessing discrimination of the new model, Cleveland Clinic score, and SRI score were 0.801, 0.670, and 0.627, respectively. Compared with the other two scores, the new model presented excellent calibration according to the calibration curves. Decision curve analysis presented the new model was more clinically useful than the other two scores. Conclusions We developed and validated a new model for predicting AKI after cardiac surgery in patients of advanced age, which may help clinicians assess patients' risk for AKI.


| INTRODUCTION
With the continued expansion of the aging population and the higher prevalence of cardiovascular disorders in the advanced age population, the number of advanced age patients who need cardiac surgery has been rising in recent years. 1,2 Currently, cardiac surgeries offered to patients of advanced age include coronary artery bypass graft (CABG) surgery and valve surgery.
Given the higher rate of postoperative complications, 3 the quality of life of patients of advanced age after cardiac surgery is more likely to be compromised than that of younger patients. 4 Acute kidney injury (AKI) is one of the major complications associated with cardiac surgery. Based on the type of cardiac surgery and the definition of AKI, the AKI incidence after cardiac surgery ranges from 20% to 70%. 5,6 AKI may compromise patient's quality of life and is linked to high mortality, all of which pose a heavy financial burden on society and families. 7 An easily calculable clinical risk model based on the basic characteristics of patients and clinical data that are easily available during the perioperative period may facilitate clinical decision-making, patient counseling, and medical optimization.
To date, several models [8][9][10][11][12] have been developed, such as Simplified Renal Index (SRI) score 9 and Cleveland Clinic score, 11 which have been frequently validated in European and American patients. However, most existing models were designed to predict AKI requiring renal replacement therapy. Compared with mild AKI, which does not require dialysis, the incidence of renal replacement therapy is low and occurs late in clinical practice, which limits the application of these models. Because a mild increase in serum creatinine is also related to a poor prognosis, 13 it is clinically imperative to develop a model to predict all stages of AKI. In addition, recent studies have uncovered several new independent risk factors for AKI after cardiac surgery, including uric acid level 14 and red blood cell (RBC) transfusion, 15 which were not included in the development of the aforementioned predictive models. Furthermore, cardiac surgery is booming in developing countries due to the development of medical technology. The proportions of population race, comorbidities, and valve surgery are quite different between developing countries, such as China, and those of the existing model derivation cohorts. 16   This study complied with the Declaration of Helsinki and was approved by the Ethics Committee of Guangdong Provincial People's Hospital without the need for signed informed consent from the participants (No. GDREC2018416H). Also, as this was a retrospective study, all subject identification information was removed before analysis.

| Data collection
The data of all participants were collected retrospectively through electronic health records established in our hospital. The potential variables used to develop this new prediction model were selected based on the well-recognized AKI risk factors in the field. 19 F I G U R E 1 Flowchart outlining participant selection. ESRD, end-stage renal disease; RRT, renal replacement therapy HU ET AL.

| Outcomes
The primary endpoint was AKI after cardiac surgery, which was defined based on the Kidney Disease Improving Global Outcomes (KDIGO) criteria, 21 which is an elevation in serum creatinine of ≥0.3 mg/dl (≥26.5 µmol/L) within 48 h after surgery or an elevation in serum creatinine by 50% from baseline within 7 days after surgery.
The last preoperative serum creatinine was used as the baseline.

| Sample size
According to the rule of thumb that a minimum of five events is required for every predictor variable in a logistic model, 22 we estimated that at least 545 patients were required in the development set for 60 candidate predictor variables, with an assumed event rate of 55%.

| Statistical analysis
Data were collected using a standardized form for each operation in which the designated recorder was present. All data were entered in Epidata 3.1 (The EpiData Association, Odense, Denmark). Continuous variables are expressed as median (interquartile range) or mean ± standard deviation (SD), and Mann-Whitney U test or Student's t-test was used for statistical comparisons between groups.
Categorical variables are expressed as frequency (percentage), and Fisher exact test or χ 2 test was used for statistical comparisons between groups. The continuous predictors (platelet count, albumin level, natremia, calcium level, magnesemia, and phosphorus level) were truncated at the 1st and 99th percentiles to limit the influence of extreme values. For missing data, multiple imputations with chain equations and an iteration of 10 times was used to estimate the missing data and were merged according to Rubin's rules. 23 Then, multicollinearity between variables was tested with variance inflation factors. The Box-Tidwell method was used to test the linear correlations between continuous variables and the risk of AKI.
Least absolute shrinkage and selection operator (LASSO) regression was used to reduce the dimensions of the data and select variables in the development set. This approach avoids issues of multicollinearity and overfitting, even with a high number of potential predictors and a small sample size. Ten-fold cross-validation and the 1-SE rule were performed to control for overfitting. 24 The final variables were included in the logistic regression, and a new model was developed. To facilitate its clinical use, a nomogram was drawn based on the weight of each variable in the final multivariable regression model. The weighted point was calculated by the beta coefficient of each variable in the model. The variable with the highest beta coefficient was scored on a 100 points scale, and the remaining variables were scored according to their individual weighted effect. Finally, the total number of points was calculated. 25 We validated the new model using the bootstrap method 26 in the development set with an iteration of 1000 times. The performance of the new model, focusing on discrimination, calibration, and clinical usefulness, was also analyzed in the validation set. The performance of this new model to predict AKI was compared with that of the SRI score and Cleveland Clinic score in the validation set. The SRI score and Cleveland Clinic score were calculated based on the data included for the validation set. The definition of each variable used for scoring adopted the original standard 9,11 ( Table S1). The area under the receiver operating characteristic curve (AUC) was used to assess discrimination. A calibration curve was plotted to evaluate the calibration and was accompanied by the Hosmer-Lemeshow test. Decision curve analysis (DCA) was used to evaluate the clinical usefulness of the model by quantifying the net benefits at different threshold probabilities in the validation set. 27 The DeLong method was used to compare the AUC of each model. 28 All analyses and reports for the development and validation of this model followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis guidelines. All statistical analyses were carried out with IBM SPSS v.25.0 (SPSS IBM) and R software (version 3.6.1; https://www.r-project.org). A p < .05 was considered statistically significant.

| Demographic and baseline clinical data of participants
The demographic and baseline clinical data of both the groups are presented in Table 1. Both the development set and the validation set had an incidence of AKI of 61.8%.

| Feature selection and model construction
Several variables were correlated with a higher risk of AKI according to the univariate analysis in the development set (Table S2) Table 2. A nomogram was also drawn according to the logistic regression results ( Figure 3).

| DISCUSSION
In this study, we developed a new model for predicting AKI after cardiac surgery in patients of advanced age. Compared with the SRI score and Cleveland Clinic score, in terms of the discrimination, calibration, and clinical usefulness, our new model may be more sui-  The previous models [8][9][10][11][12] for predicting AKI were generally evaluated in terms of discrimination and calibration. Whether the decisions made based on these models could generate clinical benefits were not assessed. The DCA curve is a novel method for evaluating the clinical utility of a model by quantifying the net benefits, which may aid clinical decision-making. 27 We found that our new model exhibited better performance than the Cleveland Clinic score or SRI score. The reason for the improved performance of our new model is likely the inclusion of new important risk factors, including the preoperative uric acid level and intraoperative RBC transfusion, which were not considered in these models.
Recent studies have suggested that preoperative uric acid elevation is an independent risk factor for AKI in patients undergoing cardiac surgery. 14 Consistently, we also observed this. Uric acid causes kidney damage through multiple mechanisms, including inhibiting the proliferation and migration of endothelial cells, promoting the apoptosis of proximal renal tubules and vascular endothelial cells, activating the renin-angiotensin system to induce vasoconstriction, increasing reactive oxygen radical levels, and promoting the release of inflammatory mediators. 29 Some studies have shown that with modern economic development and lifestyle changes, the prevalence of hyperuricemia in developing countries has increased, especially in an advanced age. 30  The predictors in our model, such as baseline serum creatinine, NYHA class III or IV, CPB time above 120 min, and hypertension, have been identified as risk factors for AKI after cardiac surgery in previous studies. 8,10,11,17 Mechanical ventilation duration was another AKI risk factor. The pathogenic mechanism for mechanical ventilation may involve the reduction of cardiac output, induction of the release of inflammatory factors, and redistribution of renal blood flow, thus causing the organ's injury. 33 A duration of mechanical ventilation longer than 24 h increased the risk of AKI by three-fold. 34 In our study, prolonged mechanical ventilation (defined as mechanical ventilation for more than 24 h) was predictive of AKI.
The incidence of AKI in this study was 61.8%, which was higher than that reported in European and American populations. 35 was not available and thus was not included in our study. Lastly, we only used the serum creatinine recommended by KDIGO, but F I G U R E 5 Decision curve analyses for prediction models. The x-axis shows the threshold probability. The y-axis shows the net benefit. The dashed and solid black lines represent the hypothesis that no patients and all patients had AKI, respectively. The net benefit was computed by subtracting the proportion of false positives from the proportion of true positives in all patients, with weighting the relative harm driven by the false positive. The threshold probability was where the expected benefit of avoiding treatment is equal to the expected benefit of treatment. For each decision threshold, the net benefits of the new model, Cleveland score, and SRI score are presented. For a given threshold, the difference in net benefit between two scores was the additional number of AKI cases identified (per 1000) without increasing the number of false-positive classifications. Across the range of decision thresholds, the new model was consistently positive and had a larger net benefit than the SRI and Cleveland scores. AKI, acute kidney injury; SRI, Simplified Renal Index not urine volume, as the diagnostic criterion for AKI. Because urine volume is easily affected by a number of factors, such as diuretic use, fluid replacement, and urine collection, most studies on AKI only use serum creatinine as a diagnostic criterion. 36

| CONCLUSION
We developed a model for predicting AKI after cardiac surgery in patients of advanced age. Compared with the SRI score and Cleveland Clinic score, this new model exhibited better performance. In the future, we will integrate this model into our electronic medical records system to facilitate its clinical application.

CONFLICT OF INTERESTS
The authors declare that there are no conflict of interests.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.