Prediction of Renal Graft Function 1 Year After Adult Deceased-Donor Kidney Transplantation Using Variables Available Prior to Transplantation

Introduction

Kidney transplantation (KTx) is still regarded as the best therapeutic approach for patients with end-stage renal failure, as it facilitates longer patient survival with a higher quality of life compared to dialysis [1,2]. Additionally, possible complications and substantial expenses associated with dialysis clearly outweigh the risks of KTx along with the subsequent immunosuppressive therapy [3–5]. As the disparity between patients on the waiting list and available donor organs is growing [6–8], kidney allocation remains a challenge for all involved healthcare professionals, requiring the consideration of medical as well as socio-economic, political, and ethical factors [9,10]. While long-term graft function is widely accepted as an overall measurement for successful KTx, providing the recipient with as many dialysis-free years as possible, early renal graft function (eg, after 1 year) is the key factor and an important determinant of long-term graft survival [11,12].

Optimal kidney allocation would avoid futile KTx and enable superior short- and long-term graft function. Kidney function after KTx is most commonly assessed by serum creatinine levels or estimated glomerular filtration rate (eGFR), as determined by various formulas, or, alternatively, by creatinine clearance measurement [13]. Multiple factors have been identified to assess the risk of graft failure or loss, thus facilitating an optimization of peri- and postoperative treatment regimens. These factors include episodes of acute rejection [14–16], acute kidney injury and severe infection after KTx [17,18], and persistent hypertension, diabetes, or proteinuria [15]. A major predictor of functional long-term outcome is a good early functional recovery, mostly described as the absence of delayed graft function (DGF) after KTx [15,16,19–24]. Obviously, at this point, the allocation and surgery have already been performed.

Standard operating procedures and clinical pathways are helpful to standardize intra- and postoperative decision-making. Therefore, a reliable and accurate prognostic model for the prediction of renal graft function after adult KTx using pre-transplant donor and recipient data is highly desirable. Known predictors for adverse outcomes after KTx, ranging from the occurrence of DGF to an increase in mortality, include donor- and recipient-related parameters like age, body weight, and immunization status as well as transplantation-specific factors like the estimated cold ischemia time (CIT) [25–27]. Several available prognostic models must be considered outdated, as they incorporate patient data from the 1990s and even before [26,28], whereas some newer studies are confined to a limited number of donor and recipient parameters while ignoring the third category of transplantation-specific factors [29]. Hence, KTx prognostic model development with pre-transplant data is still the subject of ongoing research [30].

Our current study proposes such a model. We investigated pre-transplant recipient- and donor-related variables regarding their influence on renal graft function after 1 year. Additionally, we evaluated transplantation-specific parameters that were also available in advance of the actual KTx surgery, like the number of HLA mismatches or the CIT. A reliable prognostic model obtained in this manner would aid in the identification of the best recipient/donor-pairings in terms of short-term, and subsequently, long-term graft function, enabling an optimized kidney allocation and reducing the risk of graft failure. Furthermore, recipient/donor-matches with elevated risk for poor graft function could be identified prior to a clinically detectable eGFR decline, hence facilitating a risk-adjusted, close follow-up regimen.

Material and Methods

SETTING AND DATA COLLECTION:

The study setting was provided by a university hospital and tertiary referral center for transplant surgery in Germany. We present a single-center study based on ongoing data collection and retrospective data analysis.

INCLUSION AND EXCLUSION CRITERIA:

We analyzed the data of 2431 patients with KTx performed at Hannover Medical School between January 01, 2000, and December 31, 2012. Exclusion criteria were pediatric KTx (age <18 years), combined organ transplantations and living-related organ donations. Furthermore, cases lacking 1-year follow-up after transplantation (±3 months) were excluded from the study, comprising all cases with early graft loss due to surgical complications or primary non-function. Hence, a study cohort (n=1172) of adult KTx recipients after post-mortal kidney donation was used for prognostic model development. The flow of patients through the study is shown in Figure 1. All enrolled patients received standard triple immunosuppressive therapy during the first year after KTx.

VARIABLES AND DEFINITIONS:

We only analyzed variables that were available before the actual KTx procedure: Donor- and recipient-related pre-transplant characteristics as well as parameters directly attributable to the transplantation-specific setting (HLA mismatch, CIT). As we intended to build a statistical model that might improve organ allocation, all intra- and postoperative variables, including surgical- or intensive care-related complications and the occurrence of DGF, were deliberately excluded from the analysis.

For definition of KTx urgency and pre-transplant recipient immunization, we applied the criteria of the Eurotransplant Kidney Allocation System [31]: high urgency (HU) status was defined as a lack of access to dialysis, inability to cope with dialysis, or severe bladder problems due to kidney graft failure after simultaneous pancreas/kidney transplantation. Highly immunized (HI) recipients were transplantable patients with high allo-sensitization, as indicated by a panel-reactive antibody (PRA) range of ≥85%. Accordingly, immunized (I) recipients had a PRA range from ≥6% to <85%. Transplantable (T) recipients displayed a PRA range of <6%.

Calculation of the eGFR 1 year after KTx was conducted using the 4-variable Modification of Diet in Renal Disease (MDRD) equation described by Levey et al [32], which was the standard formula applied in our hospital laboratory during the data acquisition period. The resulting GFR categories from KDIGO G1 (normal kidney function, eGFR ≥90 ml/min/1.73 m2) to KDIGO G5 (kidney failure, eGFR <15 ml/min/1.73 m2) were determined according to the method described by Stevens et al [33].

END-POINT AND STUDY HYPOTHESES:

The study end-point was renal graft function 1 year after KTx, as determined by eGFR value and the resulting GFR categories (KDIGO G1-G5).

The primary study hypothesis was that the investigated variables could be used to develop a model for prediction of eGFR values 1 year after KTx with sufficient accuracy (R² >0.800). This model would facilitate a ranking of potential KTx recipients for specific organ offers to identify the recipient/donor-match with the potentially best long-term graft function.

Our secondary hypothesis was that, using the above-mentioned variables, it was possible to predict GFR categories 1 year after KTx with high sensitivity and specificity and an area under the receiver operating characteristic (AUROC) curve >0.700 for each GFR category. This could also prove beneficial for a ranking of potential kidney transplant recipients, with the obvious limitation that due to the ordinal scale of the GFR categories, the resolution of prediction will be reduced. Nevertheless, this model could also provide further insight into the role of specific risk factors contributing to compromised renal graft function after KTx.

STATISTICAL METHODS FOR PROGNOSTIC MODEL DEVELOPMENT:

Potential prognostic factors for a reduced long-term graft function after KTx were identified by a comprehensive literature search as well as consulting with experienced kidney transplant specialists. The parameters were divided into 3 groups: recipient-related, donor-related, and transplantation-specific variables. Variables with more than 5% missing values were eliminated from the subsequent model development (Tables 1–3), which was only applicable to certain donor criteria (Table 2) that were not consistently provided during the organ donation and allocation process.

To test the influence of each variable on eGFR 1 year after KTx and to improve the general understanding of the data, univariable linear regression analyses were performed. Additionally, principal component analyses and correlation matrices were applied to avoid multi-collinearity during regression modelling. In case of a high correlation between 2 variables (2-sided correlation with R >|0.500|), 1 of the 2 correlating variables was excluded from the subsequent multivariable regression analyses. For this exclusion process, a combination of medical expertise and a prioritization of variables with lower p-value during the former univariable regression analysis was used. As they belong to different biological entities prior to transplantation, all recipient and donor variables were investigated in separate principal component analyses, as were the transplantation-specific variables.

Because of the above-mentioned reason, all remaining variables with a lack of inter-variable correlation with R >|0.500| were first used to build separate multivariable linear regression models for recipient and donor (intermediary models). Transplantation-specific variables that belong neither exclusively to the donor nor to the recipient were analyzed separately in a subsequent step. Stepwise backward likelihood elimination of the least significant variables was performed for each of the 2 models, until only significant variables remained within the intermediary model for donor and recipient, respectively.

Thereafter, the pre-transplant donor- and recipient-related variables with a significant, independent influence on eGFR 1 year after KTx underwent a second multivariable linear regression analysis. As this second analytical step was used to model the setting of the actual transplantation, now all transplantation-specific variables were also included in the evaluation. The second multivariable modelling step was again conducted with backward likelihood elimination of the least significant variables. Furthermore, a purposeful selection of co-variables was performed until only significant and medically plausible variables remained in the final linear regression model. Multi-collinearity between continuous variables within the final model was excluded by a separate principal component analysis.

After the first study hypothesis had to be discarded, the independent significant variables identified during linear regression modelling were utilized for a subsequent multivariable ordinal-logistic regression analysis to investigate their impact on the GFR category 1 year after KTx. A receiver operating characteristic (ROC) curve analysis with determination of the AUROC was conducted to test the prognostic capability of the final regression model to predict each GFR category separately with high sensitivity and specificity. An unimpaired kidney function 1 year after KTx (KDIGO G1) was used as a reference.

All statistical analyses were performed using JMP Pro 13 software (SAS Institute, Cary, NC, USA). A p-value <0.050 was defined as statistically significant.

ETHICAL CONSIDERATIONS AND CONSENT:

This retrospective study was approved by the Ethics Committee of Hannover Medical School. Patients provided informed consent that their data could be used for scientific purposes at the time of hospital admission for transplantation, which is the general policy of our institution. All patient data were anonymized and de-identified.

Results

DESCRIPTIVE STATISTICS:

The distribution of the investigated groups of variables within the study cohort is shown in Tables 1–3. Mean (median), range, and standard deviation are given for continuous variables to represent the data structure so that non-normally distributed variables can be identified immediately. The mean (median) eGFR 1 year after KTx was 43.1 (40.8) mL/min/1.73 m2 with a standard deviation of 17.3 mL/min/1.73 m2 (95%-CI: 42.1–44.1 mL/min/1.73 m2) and an interquartile range of 22.4 mL/min/1.73 m2. Mean and median eGFR 1 year after KTx equaled GFR category G3b, corresponding with moderately to severely decreased kidney function.

RISK FACTOR ANALYSIS WITH UNIVARIABLE LINEAR REGRESSION AND PRINCIPAL COMPONENT ANALYSES:

The results of the univariable linear regression analyses are summarized in Tables 4–6. Recipient-related risk factors with a significant influence on eGFR 1 year after KTx were age, weight, height, BMI, a waiting list status of HU, and a chronic glomerulonephritis or unknown etiology of kidney failure as an indication for KTx. Furthermore, the following donor-related variables had a significant influence on eGFR 1 year after KTx: age, height, BMI, female sex, and a cause of death due to cerebral edema, trauma of the central nervous system, cerebro-vascular bleeding, subarachnoid hemorrhage, respirational causes, or traumatic injury. Regarding transplantation-specific variables, the number of HLA mismatches for HLA-A, HLA-B, and HLA-DR displayed a significant influence on eGFR 1 year after KTx.

The principal component analysis for continuous, recipient-related variables revealed a correlation exceeding the threshold of R >|0.500| between recipient weight, height, and BMI, between waiting time for KTx and time since first dialysis in years, and between current and highest PRA values in percent, respectively. Therefore, we decided to exclude recipient weight, recipient height, time since first dialysis, and highest PRA value from the subsequent multivariable regression analyses. The donor-related variables showed a correlation above R >|0.500| for donor weight, height, and BMI. Accordingly, we excluded donor weight and donor height from the subsequent modelling process. The transplantation-specific variables displayed no relevant factor correlations.

MULTIVARIABLE LINEAR REGRESSION MODELLING FOR EGFR 1 YEAR AFTER KTX:

The final, multivariable linear regression model revealed that only the following 5 variables had an independent, significant influence on eGFR 1 year after KTx: donor age, recipient BMI, a re-transplantation beyond the second KTx procedure (ie, third, fourth, or fifth re-transplantation of the same recipient), cold ischemia time, and donor creatinine (Table 7). R2 of the multivariable linear regression model was below 0.800, indicating a highly questionable usefulness because of a lack of correlation of the predictive model (R2=0.212). Thus, the first hypothesis was discarded after completion of the linear regression modelling process.

MULTIVARIABLE ORDINAL-LOGISTIC REGRESSION MODELLING FOR GFR CATEGORY 1 YEAR AFTER KTX:

Multivariable ordinal-logistic regression analysis revealed that all of the 5 previously identified variables with an independent, significant influence on eGFR 1 year after KTx also showed an independent, significant influence on the GFR category 1 year after KTx. The p-values of the multivariable ordinal-logistic regression model are summarized in Table 8.

FORMULA OF THE FINAL ORDINAL-LOGISTIC REGRESSION MODEL FOR PREDICTION OF GFR CATEGORIES 1 YEAR AFTER KTX:

The linear logit of the final ordinal-logistic regression model for the GFR category 1 year after KTx was defined as follows:

The predicted probability for GFR category G1 was calculated as follows:

The predicted probability for GFR category G2 was calculated as follows:

The predicted probability for GFR category G3a was calculated as follows:

The predicted probability for GFR category G3b was calculated as follows:

The predicted probability for GFR category G4 was calculated as follows:

The predicted probability for GFR category G5 could not be calculated reliably.

POTENTIAL USEFULNESS OF THE FINAL MULTIVARIABLE ORDINAL-LOGISTIC REGRESSION MODEL AS A PREDICTION MODEL:

The final multivariable ordinal-logistic regression model is suitable to predict the GFR categories G2, G3a, G3b, and G4 1 year after KTx with high sensitivity and specificity, as shown by ROC curve analysis. The AUROCs for all of the mentioned predictions were greater than 0.700, fulfilling the conventional criteria for a prediction model of potential clinical usefulness (Figure 2). The AUROC for GFR category G5 was 0.6109, indicating that the highest GFR category cannot be predicted with sufficient sensitivity and specificity. GFR category G1 was used as a reference.

Discussion

This study proposes a new model for the prediction of graded renal graft function 1 year after deceased-donor adult KTx. Unlike preliminary studies that solely investigated the influence of donor variables on creatinine clearance to identify marginal donor organs [34], we present the first prognostic model comprising donor-related, recipient-related, and transplantation-specific parameters in a statistical approach that was designed to resemble the unique clinical setting of organ transplantation by use of a 2-step multivariable regression model. As it is intended to improve recipient/donor-matching and optimize organ allocation concerning short- and thereby also long-term function, only variables available in advance of the actual transplantation surgery were included. The multivariable regression model shows that donor age, donor serum creatinine levels, recipient BMI, re-transplantations beyond the second KTx procedure, and cold ischemia time have a significant influence on KDIGO-graded renal graft function after 1 year. All identified significant variables are clinically plausible and generally correspond with previous findings, as presented below.

Regarding donor-related variables, a higher donor age is commonly known as a major risk factor for poor long-term graft function [14–16,35]. Si Nga et al found that donor age was the only parameter that correlated with outcome, as older donors displayed a higher risk for occurrence of DGF and worse kidney function 6 months after KTx [36]. Dempster et al also showed that higher donor age, but not recipient age, leads to higher rates of graft loss 1 year after KTx [37]. The fact that the median donor age in this study was relatively high (52.0 years) has likely increased the influence of a higher donor age on post-transplant kidney function. On the contrary, the influence of donor serum creatinine is more controversial – while elevated terminal creatinine levels have repeatedly been identified as a risk factor for DGF after KTx [35,38,39], a creatinine-based estimation of donor kidney function is not advantageous for the prediction of long-term renal graft function compared to a prediction based on donor age alone [40]. Shaheen et al even reported that neither donor age nor a terminal rise in donor serum creatinine significantly affects graft function, graft survival, or patient mortality [41]. This finding might, however, be attributable to a low number of older donors (n=43 donors of age >50 years) within the presented study cohort and a low cutoff (>150 μmol/l) for the donor group with the highest serum creatinine at time of organ recovery [41]. Considering the observation that transplantation of expanded criteria donation (ECD) or donation after cardiac death (DCD) kidneys has a potentially negative impact on KTx outcome [24,35], it seems even more plausible that donor age and donor serum creatinine levels are relevant variables for the prediction of long-term graft function. In contrast, a sex mismatch between donor and recipient showed no impact on graft survival, which is consistent with previous findings [42].

Also, in line with published literature, recipient BMI was identified as a significant, independent risk factor for a higher GFR category 1 year after KTx [14,15,35,43]. Liese et al showed an association between an increased recipient BMI and different adverse outcomes, especially the occurrence of DGF and a lower eGFR after 1 year [44]. As a high recipient BMI is associated with an increased risk for new onset of diabetes mellitus after KTx [45] and the recipient diabetic state is an independent risk factor for reduced long-term patient survival [46], the cardiovascular and other diabetic changes imposed by obesity and metabolic syndrome might also explain the effect a high recipient BMI has on long-term graft function after KTx [47]. Particularly, this manifests in diabetic patients undergoing simultaneous pancreas/kidney transplantation – within this group, recipient age has also been identified to have a significant effect on long-term renal graft function, most probably because of the detrimental effects a prolonged diabetic state imposes on the cardiovascular system [48].

Additionally, we identified re-transplantations beyond the second KTx as an independent risk factor for higher GFR categories. This aligns with previous studies demonstrating lower graft survival rates in re-transplantations of higher grade (ie, beyond the second KTx procedure) [49]. Kim et al found a higher risk for DGF and other complications in third KTx recipients, resulting in lower allograft survival [50]. Graft survival rates might be especially low for re-transplantations in patients with high PRA ranges (>80%) and in patients who lost their previous grafts early after the last transplantation [51]. However, due to the low number of re-transplantations beyond the second KTx within our study cohort (only 2.7% of all cases), this specific sub-cohort will need a more focused, detailed analysis in future studies.

Of all transplantation-specific parameters, only CIT had a significant impact on long-term graft function. This is congruent with several other studies that identified CIT as an independent and significant risk factor for early graft failure and DGF [35,52,53], inferior graft survival and function [41,53], and even patient survival [52]. Frei et al calculated that every hour of CIT increases the risk of graft loss by 3% [52], while a longer CIT of more than 14 hours elevates the risk for an impaired graft survival for up to 5 years after KTx [54]. Corresponding with the literature, our findings show the need to keep the CIT as short as possible. This might be of particular importance during the transplantation of ECD and DCD organs [24].

The number of HLA mismatches did not show a significant impact on the GFR category 1 year after KTx in our multivariable regression model, whereas univariable regression analysis demonstrated a significant influence for all 3 types of HLA mismatches (Table 6). The latter observation is in line with most studies that identified HLA mismatches as significant and independent risk factors for graft failure after KTx [55,56]. Our observations might be partially attributable to an early rejection of HLA-mismatched kidneys, leading to a drop-out of patients from the study cohort before the follow-up after 1 year was reached. Nevertheless, as the HLA-mediated decline in renal graft function is finally attributable to rejection processes, it seems more likely that this adverse effect can be avoided or at least controlled by modern immunosuppressive strategies. The applied immunosuppressive medication itself has been identified to have an additional effect on the functional KTx outcome [57,58].

In prognostic modelling, variables may become statistically significant after creating ordinal categories that were previously insignificant when using numerical variables. This is likely, since numerical variables may have a non-linear statistical influence in ordinal regression analysis. Such non-linear statistical influences that become apparent during analysis are nonetheless potentially clinically relevant. On the other hand, the fact that multivariable linear regression was not able to predict eGFR 1 year after KTx may be attributable to non-normally distributed data. The residuals were not tested for normality and no data transformations were applied to meet the assumptions for linear regression. We decided to apply multivariable ordinal-logistic regression for prognostic modelling, aiming to predict GFR categories with high sensitivity and specificity. We believe that this approach also leads to a model with higher clinical usefulness, because the sensitivity and specificity of prediction for each GFR category can be calculated. The proposed prognostic model does not take non-conventional variable interactions into account, such as multiplicative instead of additive risk increases. However, the goal of this study was to find a parsimonious prognostic model that is still able to predict GFR categories with high sensitivity and specificity, because we believe that such an approach would potentially enhance clinical usefulness in model application.

The strengths of this study include a comparatively large patient cohort as well as a wide range of pre-transplant data. Despite the promising overall results, the presented study is also subject to certain limitations. Due to the single-center, retrospective study design, including only 1 ethnic group, selection bias may have occurred. Furthermore, an external validation of the prognostic model in other transplant centers is needed before a broad application of the proposed model becomes reasonable. For clinical application it will also be necessary to estimate CITs in advance; this could be implemented during the early allocation procedure. Based on our clinical and surgical experience, operating time is only a minor proportion of CIT, whereas the largest, and therefore most relevant, proportion of CIT is attributable to the time required for organization and logistics, including donor organ allocation and transportation. These times can be estimated with good accuracy as long as the location and beginning of the donor operation are known and the logistics of transplantation have been tentatively organized in advance. Finally, a prediction of kidney graft failure after 1 year as represented by GFR category G5 (eGFR <15 ml/min/1.73 m2) was not possible with sufficient sensitivity and specificity. We attribute this to the low number of patients with GFR category G5 (n=19) within our study cohort, most probably because patients who returned to dialysis due to graft failure were excluded from our analysis because of a lack of post-transplant follow-up 1 year after KTx (Figure 1).

In conclusion, we developed a clear and concise model for the prediction of graded renal graft function 1 year after KTx that incorporates recipient- and donor-related as well as transplantation-specific variables. All included parameters are easily available during routine clinical practice and are known (or, in the case of CIT, calculable) prior to transplantation. The resulting multivariable regression model can predict GFR categories after 1 year with high sensitivity and specificity (AUROC >0.700), with the above-mentioned exception of GFR category G5. Our prediction model could facilitate an allocation of specific donor kidneys to specific recipients to achieve a lower GFR category 1 year after KTx, thereby also facilitating superior long-term graft function. This could also help to avoid futile KTx, not only for overall patient risk reduction, but also for minimization of recipient mortality due to the return to dialysis after graft failure [59]. Another benefit of the presented model is the possible identification of recipient/donor-matches at elevated risk for impairment of graft function before an eGFR decline is clinically detectable. Re-transplantations beyond the second KTx in combination with long CIT, an older donor organ, high donor creatinine, or an obese recipient are at increased risk for poor long-term graft function, as indicated by a higher GFR category, and should therefore be avoided. If a more suitable recipient is not available, the prediction of a higher GFR category should be the reason for heightened clinical vigilance during KTx follow-up. Finally, the results of our multivariable regression analyses provide additional insights into the variables that are relevant as independent significant risk factors influencing eGFR and GFR category after deceased-donor adult KTx.

Conclusions

It is possible to predict GFR categories 1 year after KTx with high sensitivity and specificity using only the pre-transplant variables donor age, donor creatinine, recipient BMI, re-transplantations beyond the second KTx, and CIT. The resulting prognostic model is of high clinical usefulness. Besides the obvious benefit of an optimized recipient/donor-matching with regard to renal graft function, it facilitates a risk-adjusted follow-up regimen for patients at an elevated risk of developing a higher GFR category. A combination of multiple adverse factors can be avoided in advance.