Objectives: There is an 18.9% discard rate among kidney allografts. Here, we aimed to determine predictors of kidney discard and construct an index to identify high-probability discard kidney allografts prior to procurement.

Materials and Methods: A total of 102 246 potential kidney allograft donors from the Organ Procurement and Transplantation Network database were used in this analysis. The cohort was randomized into 2 groups. The training set included 67% of the cohort and was used to derive a predictive index for discard that comprised 21 factors identified by univariate and multivariate logistic regression analysis. The validation set included 33% and was used to internally validate the kidney discard risk index.

Results: In 77.3% of donors, at least 1 kidney was used for transplant, whereas in 22.7% of donors, both kidneys were discarded. The kidney discard risk index was highly predictive of discard with a C statistic of 0.89 (0.88-0.89). The bottom 10th percentile had a discard rate of 0.73%, whereas the top 10th percentile had a discard rate of 83.65%. The 3 most predictive factors for discard were age, creatinine level, and hepatitis C antibody status.

Conclusions: We identified 21 factors predictive of discard prior to donor procurement and used these to develop a kidney discard risk index with a C statistic of 0.89.

Key words : Allocation, Supply, Transplant

Introduction

Any effort to maximize the donor supply of kidney allografts is critically important in the face of the increased disparity between listed candidates and recipients of transplants. The mismatch between the supply and demand is exacerbated by an 18.9% discard rate among kidney allografts.¹ From a different perspective, these discarded kidneys, nearly one-fifth of the potential allografts, represent an opportunity to increase the effective supply for transplant.

To promote efficient organ utilization, it is important to identify allografts that are at increased risk for discard. Many factors have been shown to contribute to kidney discard, including multiple donor characteristics, multiple transplant center characteristics, Organ Procurement Organization sharing, and time of procurement.^2-9 A metric that relies only on donor characteristics applied indepen­dent of location could be useful to evaluate the risk of discard before the time of initial procurement.

A predictive index of combined multiple donor factors to evaluate kidney allograft discard risk could serve 2 purposes. First, the metric could stratify allog­rafts based on discard risk. Second, the metric could be used to preferentially allocate high-probability discard allografts to aggressive transplant centers. We speculate that high-probability discard organs are more likely to be used in aggressive transplant centers, which effectively increases the donor yield.^10-12

Previous investigators have addressed this very issue and have constructed indexes to predict kidney allograft discard by using factors obtained after transfer of the donor to the operating room.^3,9 We aimed to create a model to be used before initial procurement. We considered donor factors to deter­mine predictors of kidney discard and constructed a kidney discard risk index (KDSRI) based on these factors, to provide a scoring system for use in the clinical setting.

Materials and Methods

Study population
We retrospectively analyzed the United Network for Organ Sharing database to recognize all deceased adult donors, age 18 years or older, with at least 1 kidney procured for transplant between January 1, 2002, and September 30, 2016. Our analysis included data from the deceased donor registry obtained through the Organ Procurement and Transplantation Network to identify characteristics predictive of kidney allograft discard. A total of 102 246 potential kidney allograft donors were used in this analysis. In 79 039 of these potential donors (77.3%), at least 1 kidney was used for allograft transplant, whereas in 23 207 cases (22.7%), both kidneys were discarded. Donor demographic and clinical characteristics were summarized. This study was exempt from approval from an ethics board.

Statistical analyses
Data were analyzed with a standard statistical software package, Stata 12.1 (Stata Corp). Continuous variables were reported as mean values ± SD and compared with a t test. Nonparametric variables were compa­red with the Mann-Whitney U test. Contin­gency table analysis was used to compare categorical variables. Results were considered significant at a P value of less than .05. All reported P values were 2-sided. The primary outcome measure was kidney allograft discard. In this analysis, the 2 potential outcomes for deceased donors were either (1) the donor provided at least 1 kidney allograft for transplant or (2) both kidneys were discarded. Deceased donor clinical characteristics and laboratory values were the independent variables, whereas double kidney allograft discard was the dependent variable in logistic regression analysis. Risk factors that were significant in univariate analysis (P < .05) were included in the multivariate logistic regression analysis. We confirmed the stability of the model by stepwise elimination based on a P value greater than .05. The entire cohort was randomized into 2 distinct groups. The training set (68 164) included 67% of the cohort and was used to derive a predictive index for discard. The validation set (34 082) included 33% of the cohort and was used to internally validate the KDSRI.

Risk factors
Deceased donor risk factors for double kidney allograft discard only included factors available in the DonorNet (United Network for Organ Sharing) presentation of the initial organ offer. Laboratory values were terminal values before recovery. Continuous variables were categorized into clinically relevant groupings. Donation after cardiac death was included to create an index that can be universally applied to deceased donors.

Development of a kidney discard risk index
Logistic regression analysis determined the predictors of double kidney allograft discard with the training set. Factors that that were statistically significant in univariate analysis were used in the multivariate analysis to predict double kidney discard. With significant factors from the multivariate analysis, a KDSRI was created to predict double kidney discard. Each donor started with a KDSRI score of 1, and donor characteristics and clinical features were used to determine the final score.

Kidney allograft discard
From the validation data set, we created 10 risk groups with increments of 10 percentile points in the KDSRI to determine average discard rates. Model discrimination was assessed by the area under the receiver operating characteristic (ROC) curve. A C statistic for kidney allograft discard with the kidney donor risk index (KDRI) was also generated. The following equation was used for the KDRI: KDRI = exp([-0.0194 × (donor age - 18) if donor age < 18 y] + [0.0128 × (donor age - 40)] + [0.0107 × (donor age - 50) if donor age > 50 y] + [0.179 if African American donor] + [0.126 if hypertension] + [0.13 if diabetes] + [0.220 × (donor creatinine - 1)] + [-0.209 × (donor creatinine - 1.5) if serum creatinine > 1.5 mg/dL] + [0.0881 if cause of donor death was cerebrovascular accident] + [-0.0464 × (donor height in cm - 170)/10] + [-0.0199 × (weight in kg)/5 if donor weight < 80 kg] + [0.133 if donation after cardiac death] + [0.24 if hepatitis C positive] + [-0.0766 if HLA-B mismatch is 0] + [-0.061 if HLA-B mismatch is 1] + [-0.13 if HLA-DR mismatch is 0] + [0.0765 if HLA-DR mismatch is 2] + [0.005 48 × (cold ischemia time - 20 h)] + [-0.364 if en bloc transplant] + [-0.0148 if double kidney transplant]).

Kidney allograft failure
We applied the KDSRI and KDRI to predict kidney allograft failure in the validation set. The indexes were applied in 141 342 patients who underwent kidney transplant from January 1, 2002, to September 30, 2016. We also performed a retrospective analysis of UNOS deidentified patient-level data. Donor and recipient characteristics were reported at the time of transplant. Follow-up information was collected at 6 months and then yearly after transplant. All recipients were followed up from the time of transplant to last known follow-up (mean follow-up time, 4.5 ± 3.7 years). The primary outcome was time to kidney allograft failure (n = 45 399, 32.1%). Time to graft failure was assessed as the date of kidney transplant to the date of allograft failure. Kaplan-Meier analysis was used for time-to-event analysis. Model discrimination was assessed with the area under the ROC curve.

Development of machine learning models
To test the strength of KDSRI, we built a machine learning model for the same study population (102 246 potential kidney allografts), split randomly into a training set (66.7%) and a validation set (33.3%). The same 21 predictive factors were used to train our models (source code and data available for download, see Gaibo13). We then compared the predictive capability of the KDSRI model and machine learning models.

Results

Study population
Data from a total of 102 246 potential kidney allograft donors were used in this analysis. For 79 039 of these potential donors (77.3%), at least 1 kidney was used for allograft transplant, whereas, for 23 207 (22.7%), both kidneys were discarded. Donor demographics and clinical characteristics are summarized in Table 1. The rate of double kidney discard was consistent over the study period (Figure 1).

Data entry rate
All variables investigated in univariate analysis demonstrated at least 99% entry completion for our study population except history of myocardial infarc­tion (93%) and donation after cardiac death (11%).

Univariate and multivariate analysis
All factors analyzed by univariate analysis are listed in Table 2 with corresponding entry completion rate, odds ratio (OR), CI, and P value. Factors found to be statistically significant in univariate analysis were used in multivariate analysis and are listed in Table 3 with corresponding donor percent that met factor criteria, OR, CI, P value, and relative weight to predict double kidney discard. The 3 most significant risk factors for double kidney allograft discard were age (18-24 years [OR, 0.49; CI, 0.44-0.56], ≥ 70 years [OR, 20.66; CI, 17.85-23.91]), creatinine level (1.0-1.9 mg/dL [OR, 2.16; CI, 2.04-2.29], ≥ 4.0 mg/dL [OR, 79.24; CI, 70.79-88.70]), and hepatitis C antibodies status (OR, 9.77; CI, 8.88-10.74).

Relative weight
Relative weight for each variable was determined for clinicians to gauge the effect of each factor but was not used in the KDSRI (Table 3). The total point value for each factor was determined by addition or subtraction of 1 risk point for every 10% increase or 10% decrease in the odds ratio as determined by the multivariate analysis.

Kidney allograft discard
Table 4 contains the KDSRI equation for prediction of double kidney allograft discard. The KDSRI score was calculated with reference to the deceased donor with the following characteristics: age 40 to 49 years, BMI 25 to 34, creatinine < 1.0 mg/dL, white ethnicity, region 4, male, donation after brain death, blood group O, and height 25th to 74th percentile.

The area under the ROC curve (C statistic) for the KDSRI score was 0.887 (range, 0.883-0.891) in the validation group (n = 34 082) and 0.885 (range, 0.883-0.888) in the training group (n = 68 164) (Figure 2). When age alone was used as a continuous variable, it predicted discard risk with a C statistic of 0.7434 in the validation group and 0.7429 in the training group.

With the KDSRI, we created risk groups based on 10-point percentile increments (Table 4). A KDSRI score of less than 0.425 corresponded to the bottom 10%, and a score greater than 58.664 corresponded to the top 10%. Discard percentage for each group was significantly greater than the bottom 10th percentile (Figure 3). The bottom 10% had a discard rate of 0.73%, whereas the top 10% had a discard rate of 83.65%.

The KDRI predicts double kidney allograft discard risk with a C statistic of 0.8246 (range, 0.8194-0.8298). Recipient and procedure specific factors in the KDRI could not be used to predict discard risk. These factors include HLA mismatch, cold ischemia time, en bloc transplant, and double kidney transplant.

Kidney allograft failure
Among the 10 KDSRI percentile groups, graft survival at 1 year was significantly lower in the 40th to 100th percentile groups compared with the 10th percentile. Graft survival at 10 years was signifi­cantly lower in the 20th to 100th percentile groups compared with the 10th percentile (Table 4).

Even though the KDSRI was designed to predict double kidney allograft discard, it also correlated with kidney allograft survival time after transplant (Table 5). The KDSRI achieved a C statistic of 0.557 (range, 0.553-0.561) and the KDRI achieved a similar C statistic of 0.583 (range, 0.579-0.587).

Support of kidney discard risk index with machine learning models
We built several machine learning models and picked the 2 best models to compare with KDSRI. The 2 models were built with Support Vector Machine and Logistic Regression, and these predicted kidney allograft discard with C statistics of 0.889 and 0.897, respectively (Figure 4). These values are comparable with the KDSRI C statistic of 0.887 (range, 0.883-0.891) and strengthen the support for KDSRI. We show that, when trained with the same index, our best machine learning models achieve a predictive power that is statistically equivalent to that of KDSRI.

Discussion

We conducted an extensive multivariate analysis of predictors of double kidney allograft discard risk with information available from DonorNet (United Network for Organ Sharing) at the initial organ offer before the donor was transferred to the operating room for initial procurement. The 3 most significant risk factors for double kidney allograft discard were as follows: age (18-24 years [OR, 0.49; CI, 0.44-0.56] and ≥ 70 years [OR, 20.66; CI, 17.85-23.91]), creatinine level (1.0-1.9 mg/dL [OR, 2.16; CI, 2.04-2.29] and ≥ 4.0 mg/dL [OR, 79.24; CI, 70.79-88.70]), and hepatitis C antibodies status (OR, 9.77; CI, 8.88-10.74). With all the significant risk factors from multivariate analysis, we constructed a weighted index, the KDSRI, to predict double kidney discard risk. The KDSRI had a very good ability to predict double kidney allograft discard risk with a C statistic of 0.89.

The study of kidney allograft discard is significant because nearly 20% of kidney allografts are discarded. The unmet need in kidney transplant is growing. Reduction of discard rate could serve as an additional source for sorely needed kidney allografts. An index to predict discard is important because the suggested strategies to reduce discard involve preferential allocation of high-probability discard allografts to aggressive transplant centers.¹⁴ An index is necessary to designate a kidney donor as high probability for discard prior to allocation. Previous investigators have produced indexes that rely on factors determined after transfer of the donor to the operating room. Marrero and colleagues³ constructed an index with a C statistic of 0.84. The KDSRI has a superior predictive ability, with a C statistic of 0.89, and can be determined before initial procurement.

There have been several proposals to reduce discard rates by alteration of allocation schemes to preferentially allocate marginal allografts to aggressive transplant centers.15-18 To facilitate these proposals, an accurate index is needed to predict discard risk. In this analysis, we identified 21 factors predictive of kidney discard and used these to develop KDSRI with a C statistic of 0.89. Furthermore, these predictors were highly effective to train supervised learning models with statistically equivalent C statistics of 0.89. Among these models, KDSRI can be best modeled by a straightforward, linear equation without loss of accuracy. This index can be used to identify kidneys with high probability for discard prior to organ procurement and prior to transfer of the donor to the operating room.

Limitations
The Organ Procurement and Transplantation Network deceased donor database only includes candidates who were transferred to the operating room with the intent to procure organs for transplant. Eligible donors who were outright refused for all solid-organ transplants were not included in this analysis. A more complete analysis would include all eligible donors, but these data are not available in the deceased donor Organ Procurement and Transplantation Network database. Another significant limitation is that this analysis does not account for center-specific behaviors or recipient factors. If a risk-averse center procures the allograft and refuses to proceed with the allograft, one may assume that the likelihood of allograft reallocation will be less than would be that case at a more aggressive transplant center. Despite these limitations, the KDSRI did achieve a C statistic of 0.89 when applied to over 100 000 deceased donors.

Since the passage of the National Organ Transplant Act of 1984 in the United States, data entry has been mandatory for all US transplant centers. Nevertheless, all patient registries suffer from variability in data entry. The findings from this study were based on large cohorts of patients and are unlikely to be significantly affected by small amounts of missing data.

Conclusions

We identified 21 factors predictive of kidney discard and used these factors to develop a KDSRI that was very predictive of double kidney allograft discard, with a C statistic of 0.89. This index can be used to identify kidneys of high probability of discard before transfer of the donor to the operating room for initial procurement.

References:

1. Organ Procurement and Transplantation Network: National Data. Richmond, Virginia; 2019. https://optn.transplant.hrsa.gov
   
   
2. Mohan S, Chiles MC, Patzer RE, et al. Factors leading to the discard of deceased donor kidneys in the United States. Kidney Int. 2018;94(1):187-198. doi:10.1016/j.kint.2018.02.016
   CrossRef - PubMed
   
3. Marrero WJ, Naik AS, Friedewald JJ, et al. Predictors of deceased donor kidney discard in the United States. Transplantation. 2017;101(7):1690-1697. doi:10.1097/TP.0000000000001238
   CrossRef - PubMed
   
4. Bae S, Massie AB, Luo X, Anjum S, Desai NM, Segev DL. Changes in discard rate after the introduction of the Kidney Donor Profile Index (KDPI). Am J Transplant. 2016;16(7):2202-2207. doi:10.1111/ajt.13769
   CrossRef - PubMed
   
5. Cohen JB, Shults J, Goldberg DS, Abt PL, Sawinski DL, Reese PP. Kidney allograft offers: Predictors of turndown and the impact of late organ acceptance on allograft survival. Am J Transplant. 2018;18(2):391-401. doi:10.1111/ajt.14449
   CrossRef - PubMed
   
6. Mohan S, Foley K, Chiles MC, et al. The weekend effect alters the procurement and discard rates of deceased donor kidneys in the United States. Kidney Int. 2016;90(1):157-163. doi:10.1016/j.kint.2016.03.007
   CrossRef - PubMed
   
7. Kim SY, Choi C, Han A, Min SI, Min SK, Jung IM, Kim SJ, Ha J. Reasons of kidney discard after allocation and procurement. Transplantation. 2015;99(10):S93-S93.
   
   
8. Butala NM, King MD, Reitsma W, et al. Association between organ procurement organization social network centrality and kidney discard and transplant outcomes. Transplantation. 2015;99(12):2617-2624. doi:10.1097/TP.0000000000000773
   CrossRef - PubMed
   
9. Massie AB, Desai NM, Montgomery RA, Singer AL, Segev DL. Improving distribution efficiency of hard-to-place deceased donor kidneys: Predicting probability of discard or delay. Am J Transplant. 2010;10(7):1613-1620. doi:10.1111/j.1600-6143.2010.03163.x
   CrossRef - PubMed
   
10. Maluf DG, Carrico RJ, Rosendale JD, Perez RV, Feng S. Optimizing recovery, utilization and transplantation outcomes for kidneys from small, ≤20 kg, pediatric donors. Am J Transplant. 2013;13(10):2703-2712. doi:10.1111/ajt.12410
    CrossRef - PubMed
    
11. Garonzik-Wang JM, James NT, Weatherspoon KC, et al. The aggressive phenotype: center-level patterns in the utilization of suboptimal kidneys. Am J Transplant. 2012;12(2):400-408. doi:10.1111/j.1600-6143.2011.03789.x
    CrossRef - PubMed
    
12. Narvaez JRF, Nie J, Noyes K, Leeman M, Kayler LK. Hard-to-place kidney offers: system-level predictors of discard. Transplantation. 2018;102:S290. doi:10.1111/ajt.14712
    CrossRef - PubMed
    
13. Gaibo Y. Supervised learning models for KDSRI and provided data. Version 1.1. June 9, 2020. https://zenodo.org/badge/latestdoi/270105982
    
    
14. Locke JE, Sellers MT. Get on with it! A novel allocation strategy to reduce kidney discards. Am J Transplant. 2019;19(11):2971-2972. doi:10.1111/ajt.15471
    CrossRef - PubMed
    
15. Vinkers MT, Smits JM, Tieken IC, de Boer J, Ysebaert D, Rahmel AO. Kidney donation and transplantation in Eurotransplant 2006-2007: minimizing discard rates by using a rescue allocation policy. Prog Transplant. 2009;19(4):365-370
    CrossRef - PubMed
    
16. O’Connor KJ, Cmunt K. Early Experience with new kidney allocation system: a perspective from the Organ Procurement Agency. Clin J Am Soc Nephrol. 2017;12(12):2057-2059. doi:10.2215/CJN.06360617
    CrossRef - PubMed
    
17. Garcia DL, Herrera ET, Lemonche CI, et al. MP79-07: Usefulness of kidney preimplantation biopsies from deceased donors for kidney transplant allocation. J Urol. 2015;193(4):e1013-e1013. doi:10.1016/j.juro.2015.02.2864
    CrossRef
    
18. Tanriover B, Mohan S, Cohen DJ, et al. Kidneys at higher risk of discard: expanding the role of dual kidney transplantation. Am J Transplant. 2014;14(2):404-415. doi:10.1111/ajt.12553
    CrossRef - PubMed