Development and Validation of a Cardiovascular Disease Risk Prediction Model for Patients with Non-Dialysis-Dependent Chronic Kidney Diseases Based on the Nomogram

The alarming number of patients worldwide diagnosed with chronic kidney disease (CKD) is a global health issue. The most recent statistical data show that nearly 700 million patients suffer from this condition, and due to this high incidence rate, medical resources are heavily skewed toward these patients. This may result in substantial economic challenges for families and health policy at a state level in different countries [1]. Patients with CKD are also at higher risk of cardiovascular (CV) complications [2] as noted in the 2016 EAS/ECS guidelines [3]. Among them, stage 3 patients are classified as the highest risk group, whereas stage 4 patients are classified under a specific high-risk category. Given the increased prevalence of CV complications in patients with CKD, the medical community has started to advocate for the implementation and development of precise prediction techniques to help identify CV risk in these patients. In view of this current need, an accurate prediction model can assist nephrologists in making clinical decisions and help avoid unnecessary expenditure of medical resources allocated for low-risk patients.

Recently, the nomogram method has been widely applied to assess the prognosis of complex diseases [4, 5]. The method has good clinical applicability based upon its utility in predicting poor outcomes among patients using numerous laboratory characteristics and relevant physiological indices. A previous study [6] demonstrated that patients with CKD are more likely to have CV complications than to undergo dialysis. However, to date, scholars [7, 8] have been focusing on predicting CV events in dialysis patients. The development of accurate prediction models of patients with non-dialysis-dependent chronic kidney disease (NDD-CKD) is of equal importance. Aiming to fill this research gap, this study presents a new nomogram that assesses the risk of CV events in patients with NDD-CKD, which may help nephrologists predict and treat CV-related events in patients with NDD-CKD.

The data for enrolled patients were extracted from the Chronic Kidney Disease Research of Outcomes in Treatment and Epidemiology (CKD-ROUTE) study. The CKD-ROUTE study is a prospective study that aims to investigate the prognosis of NDD-CKD patients. The Tokyo Medical and Dental University Hospital and its 15 associated facilities recruited about 1,000 individuals to participate in this research which was approved by the Ethical Committees at the institution. The observation took place between October 2010 and December 2011, and only patients who met the following criteria were included: (1) they were over 20 years old; (2) they had just visited or been referred to the participant nephrology centers; (3) they have been diagnosed with CKD stages 2–5 according to the Kidney Disease Improving Global Outcomes (KDIGO). Patients undergoing dialysis, experiencing active gastrointestinal bleeding, or with a history of renal transplantation, and those who did not provide written informed consent were excluded [9].

After enrollment, the team conducted follow-up of all the participants for 3 years. End of follow-up was defined as initiation of renal replacement therapy, death, withdrawal from the trial, or revocation of informed consent. Informed consent was signed by all participants.

The main variables included in the study were gender, age, body max index, systolic blood pressure, hemoglobin (Hb), estimated glomerular filtration rate (eGFR) (according to Modification of Diet in Renal Disease equation: eGFR = 194 × serum creatinine −1.094 × age − 0.287 [if female, × 0.739]), albumin (alb), urine protein to creatinine ratio (UPCR) (normal: 0–0.5 g, microalbuminuria: 0.5–3 g, macroalbuminuria: >3 g), hypertension, diabetes, cardiovascular disease (CVD) history, etiology of CKD (determined by the physician based on the patient’s medical history, clinical characteristics, and outcomes and histological findings in kidney specimens biopsied at the time of enrollment), use of renin-angiotensin-aldosterone system inhibitors (at least 4 weeks), use of diuretic substance, use of calcium channel blocker, CKD stage (CKD stage: stage 3a: eGFR 45–59 mL/min per 1.73 m²; stage 3b: eGFR 30–44 mL/min per 1.73 m²; stage 4: eGFR 15–29 mL/min per 1.73 m²; stage 5: eGFR 0–14 mL/min per 1.73 m²). All the data were derived from the Dryad database, a free public data storage platform.

CV events were considered primary outcomes, i.e., coronary artery disease (such as angina pectoris, myocardial infarction, or coronary revascularization surgery), congestive heart failure (determined by physicians based on biological indexes, echocardiography, and clinical symptoms), peripheral artery disease, or stroke (cerebral infarction, transient ischemic attack, cerebral hemorrhage, or subarachnoid hemorrhage).

All enrolled patients were randomly divided into training and validation cohorts (7:3). The three-knot cubic spline (10, 50, and 90%) was employed to account for potential nonlinear effects of continuous data. The Shapiro-Wilk method was utilized to examine the normal distribution of continuous data, and all continuous variables with normal distributions were compared using an independent samples t-test. Results were presented as mean ± standard deviation.

In addition, the Mann-Whitney U test was applied to determine the non-normally distributed variables, and they were presented as the median (1st–3rd quartile). χ² tests were also performed to compare categorical variables. If the theoretical frequency was less than 10, Fisher’s exact test was preferred. The independent prognostic factors were screened by the least absolute shrinkage and selection operator (Lasso) regression [10].The predictors selected by Lasso regression were then incorporated into a Cox regression to build the nomogram. At this stage, the receiver operating characteristic curves [11] were used to test the sensitivity and specialty of the model.

After performing all these steps, the discrimination performance of the nomogram was measured by the area under the receiver operating characteristic curve (AUC). We used the calibration curve after 1,000 bootstrap resamples to test the accuracy of the prediction model [12]. To further determine the discrimination of the nomogram, we calculated each patient’s total scores, stratified them into high-risk and low-risk groups based on the median scores, and performed survival analysis via the Kaplan-Meier (KM) method. Subsequently, decision curve analysis [13] was conducted to determine the clinical usefulness of the nomogram by quantifying the net benefits at different threshold probabilities. For external validation, we calculated the total points of each patient according to an established nomogram and used it as a factor to perform the Cox regression.

We estimated the C-index and the calibration curve of the validation cohort. To assess the predictive performance of our constructed nomogram, we also built a model with all variables included for sensitivity analysis. p < 0.05 was considered statistically significant. All statistical analyses were performed using R software version 4.05 (https://www.rproject.org/). The study’s design and statistical analyses were carried out in compliance with TRIPOD guidelines [14].

A total of 1,138 NDD-CKD patients were included in this study (assessed over a follow-up period of about 35 months). In the training and validation cohorts, approximately 10.15% and 10.78% of NDD-CKD patients experienced new CV events, respectively. Patients in both cohorts were all 70 years old on average, and the median eGFRs were 31.58 mL/min per 1.73 m² and 27.65 mL/min per 1.73 m², respectively. The two cohorts had different baseline values of eGFR, Hb, alb, UPCR, diabetes, use of renin-angiotensin-aldosterone system inhibitors, and use of calcium channel blocker, as illustrated in detail in Table 1.

We used 10-fold cross-validation to select the risk factors and the deviation coefficient percentage and λ value to evaluate the selection process. When the deviation percentage between the two λ values was not significant, the operation was automatically interrupted. Variables with coefficient values other than zero were retained (in model 1: eGFR, alb, CVD history, etiology of CKD, and age; in model 2: incorporating gender, age, body mass index, systolic blood pressure, Hb, eGFR, UPCR, alb, hypertension, diabetes, CVD history, and etiology of CKD). To simplify the model as much as possible, we selected the screening results with fewer variables (model 1) for subsequent Cox regression modeling. Figure 1 illustrates the selection process of risk factors.

A total of five variables (eGFR: HR 0.98 [95% CI 0.97–0.99], alb: HR 0.56 [95% CI 0.37–0.86], CVD history: HR 1.94 [95% CI 1.33–2.83], etiology of CKD [reference: diabetes], glomerulonephritis: HR 0.46 [95% CI 0.26–0.83]; nephrosclerosis: HR 0.42 [95% CI 0.28–0.67]; other: HR 0.32 [95% CI 0.15–0.66]; age: HR 1.02 [95% CI 1.01–1.04]) were included in the Cox regression (Table 2), and a nomogram (Fig. 2) has been constructed according to these factors. The AUCs (Fig. 3) of 1 year, 2 years, and 3 years were 0.81 (95% CI 0.75–0.87), 0.80 (95% CI 0.75–0.86), and 0.80 (95% CI 0.73–0.86), respectively, revealing that the nomogram has good discrimination. After 1,000 bootstrap resamples, the calibration curve (Fig. 4) showed that the anticipated survival probability is nearly identical to the observed survival probability. Furthermore, decision curve analysis curves showed that the established model had a net benefit over about 50%, 50%, and 60% of the risk threshold range in 1 year, 2 years, and 3 years, respectively (Fig. 5). The KM curve (Fig. 6) demonstrated that this model effectively identified patients at various risk levels. (p < 0.0001).

As shown in Figure 3, the AUCs were 0.77 (95% CI = 0.67–0.87), 0.84 (95% CI = 0.76–0.91), and 0.82 (95% CI = 0.74–0.90) for 1 year, 2 years, and 3 years, respectively. The calibration curve showed that the predicted value was close to the actual risk, indicating that the established nomogram can effectively predict the risk of CV event (Fig. 4). In addition, the KM curve (Fig. 6) revealed that our nomogram can be used to discriminate high-risk patients from the low-risk group (log-rank <0.01).

We randomly selected one of the patients enrolled in the study to calculate the probability of CV events. By summing up the scores of various risk factors, we estimated that the chances of facing CV-related events for this specific patient in 1, 2, and 3 years are 1.25%, 1.98%, and 2.91%, respectively.

For sensitivity analysis, we constructed two models to compare the prediction efficiency with the nomogram that we had previously developed. One used all of the variables (model 3), while the other was set by the screening factors of Lasso regression, which involved more variables (model 2). For external validation, the AUCs of model 2 (online suppl. Table 1; for all online suppl. material, see www.karger.com/doi/10.1159/000527856) (1 year: AUC: 0.75, 95% CI: 0.69–0.81; 2 years: AUC: 0.77, 95% CI: 0.72–0.83; 3 years: AUC: 0.79, 95% CI: 0.73–0.85) and model 3 (online suppl. Table 2) (1 year: AUC: 0.73, 95% CI: 0.65–0.80; 2 years: AUC: 0.71, 95% CI: 0.65–0.77; 3 years: AUC: 0.69, 95% CI: 0.63–0.73) were all lower than our established nomogram, proving the satisfactory extrapolation prediction performance of our established nomogram.

An accurate prediction model can help clinicians identify patients with high risk of adverse prognosis. In this study, we constructed a novel nomogram to predict the risk of new CV-related events in patients with NDD-CKD. Previous studies [15, 16] have developed CV risk scoring tools to predict the risk of CV complications for the general population, but these have limited applicability in patients with CKD, given CV risk in patients with CKD is higher than in the general population. These studies also failed to take into consideration biomarkers of CV prognosis in CKD such as eGFR and albuminuria. Other research [7, 8] has focused on developing predictive models for patients receiving dialysis. These individuals, however, have a much higher risk of CV events, and the predictors of these complications (e.g., volume of dialysis filtration and dialysis initial time) differ considerably from those of non-dialysis patients. Our study particularly focused on investigating the risk of CV events for NDD-CKD patients and included individuals from multiple centers in Japan which increase the generalizability of the constructed model in clinical practice. This is also consistent with the concept of precision medicine, which is currently being promoted to facilitate the personalized prediction and treatment of CKD patients [17].

Our results showed that the established nomogram has good predictive performance (AUC: 1 year: 0.77, 2 years: 0.84, 3 years: 0.82), and can successfully discriminate high-risk patients from low-risk groups, proving the model’s potential value for clinical application. Furthermore, five risk factors were included in the established nomogram: eGFR, alb, CVD history, etiology of CKD, and age, but the nomogram demonstrated that the majority of the points in predicting CV complications are determined by age. Increasing age is often accompanied by an increased risk of CV-related events [18]. The 2021 ESC guideline [19] states that elderly people should have their levels of CV risk assessed on a regular basis to prevent poor outcomes. Aside from age, CVD history also accounted for a large percentage of points, which is consistent with results presented in other related studies [20], implying that individuals with a history of CVD are at a higher risk of experiencing new CV complications. This may be explained by some other factors. For instance, patients who have suffered a myocardial infarction are more likely to develop chronic ischemic heart failure [21], which increases the chance of CV death. In addition, due to myocardial ischemia and vagal stimulation, arrhythmias such as sinus bradycardia and atrioventricular block are prone to occur in the early stages of myocardial infarction [22, 23]. Some common risk factors, such as blood glucose, blood lipids, and blood pressure, can also cause new CV events in individuals who have already had a heart attack or stroke [24, 25]. The medical community has already agreed on the fact that the stage of CKD is a crucial factor in predicting CV outcomes. According to the 2016 ESC guideline [3], patients with stage 3 CKD have a high risk of suffering CV complications, and the risks are even higher in stage 4 patients. Other scholars have [26, 27] confirmed that when eGFR <75 mL/min per 1.73m², the risk of CV death increase significantly. Similarly, our results show that the risk of CV events increased alongside the decline in eGFR, which aligns with the current consensus. Furthermore, most of the studies [20, 28‒30] in the literature indicate that hypoalbuminemia is associated with adverse kidney and CV outcomes. Our outcomes also indicate the same phenomenon, which support the importance of routinely testing serum albumin.

Aside from the aforementioned key factors, CKD etiology, according to our findings, is also associated with the risk of CV complications. Diabetes-related kidney impairment appears to be more likely to lead to CV events, an association that is consistent with what has been suggested by other scholars [31]. This could be due to the fact that diabetes and CKD are both significant risk factors for CV events, and the two together could hasten the progression of CVD, suggesting that patients with diabetes complicated with CKD should pay more attention to their CV health.

It is important to note that some risk factors (e.g., proteinuria, systolic blood pressure, gender, or Hb) for CV events were not selected in the model because we primarily used the simplified model as the starting point of our design. In the sensitivity analysis, when these variables were included in the final model, the prediction performance of the model decreased considerably. We postulate that this might be because the number of CV events is not enough to support the number of included variables. Although our final model did not include previously reported variables associated with CV events, this had no effect on its predictive performance. Lasso regression performs well in reducing the data dimension [32], which means that less positive events may also reflect an accurate prediction performance.

Our study has some strengths. Since the data of enrolled patients came from a multi-center clinical study, our model is considered more reliable and widely applicable. In addition, the data used were mainly based on a prospective cohort study, so the evidence used in our analysis is valid and more reliable compared with data from a retrospective study. Furthermore, the laboratory and physiological indicators in our study are easily available to clinicians, which makes their application more convenient. Moreover, taking into account that our prediction model is mainly based on the clinical data of NDD-CKD patients, our outcomes can be used as guidance for clinicians to better achieve accurate prediction for individuals diagnosed with this condition, facilitating the individual prediction and development of personalized lines of treatment for CKD patients.

Despite all these relevant outcomes, it is equally important to acknowledge the limitations of this study. First, some crucial indicators of CV prognosis, such as brain natriuretic peptide or cardiac troponin, were not investigated due to limited data. Furthermore, patients from Japan were the majority in our study, and our method is yet to be tested among patients in different countries. In addition, we used the definition of the cause of CKD mainly based on laboratory data and the experience of clinicians, which may have caused data heterogeneity. Further validation methods are still important to validate our outcomes and shall be conducted in future studies.

In this study, we established a novel nomogram for predicting CV events in patients with CKD who did not receive dialysis. Our nomogram incorporated five risk factors and demonstrated good prediction performance. This model has great potential for clinical use since it can be conveniently used to facilitate the individualized prediction of CV events in patients with NDD-CKD. Nevertheless, the external applicability of this model still needs to be further validated using data from different settings.

We appreciate the contribution the CKD-ROUTE cohort team and would like to acknowledge their selfless dedication to the original data.

This study is a secondary analysis of data from Dyrad database, which are publicly available. Hence, no institutional review board approval was necessary since written informed consent was already obtained from all participants in the original CKD-ROUTE study.

The authors have no conflicts of interest to declare.

This study did not receive any external funds.

Ning Li and Zhiqiang Li contributed to the concept and design of this study. Ning Li was responsible for statistical analysis and writing of the final report. Haitao Xie, Zhao Wang, Qinglong Gu, and Jun Guo assisted in Statistical Analysis. Zhiqiang Li and Xue Yang reviewed the article and provided critical feedback to shape the content. All authors read and approved the final manuscript.

All the datasets generated and analyzed in this study are available in the Dryad repository (https://datadryad.org/stash/dataset/doi:10.5061%2Fdryad.kq23s).