We conducted a national longitudinal study to investigate the association between early post-transplant RBC transfusion and transplant outcome. All transplant recipients in France are prospectively included in a mandatory registry (CRISTAL) administered by the French BioMedecine Agency (Agence de BioMédecine). Detailed patient characteristics were collected through this registry. Transfusion information was extracted from the national database of the French transfusion public service (Etablissement Français du Sang) which collects and prepares 100% of blood products issued in France. This database provides access to both pre- and post-transplant transfusions information. The two databases were merged for analyses. The method for building a single database from the cross of the two initial databases is described in the Supplementary Methods. This study was approved by the Committee for Personal Protection Great East II in Besançon (CPP: 13/07, 2013/02/25), by the Advisory Committee on Information Processing for Research in the Field of Health (CCTIRS, 13.138, 2013/03/19), and by French National Computers and Freedom Commission (CNIL, DR-2013-542, 2013/11/25).

Patients were eligible for inclusion if they had received a first kidney transplant in one of the 31 French transplant centers between January 1^t, 2002 and December 31, 2008. All data were extracted from the CRISTAL database. All RBC transfusion episodes that occurred from time of transplantation to 14 days following transplantation were considered. Patients with a graft survival of less than 15 days were excluded from analysis.

All RBC concentrates were produced by the Etablissement Français du Sang and underwent pre-storage deleukocytation (<1x10⁶ leucocytes/RBC unit). A transfusion episode was defined as encompassing consecutive transfusions where the interval between RBC transfusions did not exceed 48 hours (most often 1 to 2 RBC). Patients may have received multiple RBC transfusions from multiple donors.

The primary endpoint was the failure free survival defined as the time between transplantation and transplantation failure. Patients known to be alive without transplantation failure were censored at the date of their last follow-up. A transplantation failure event was defined as graft loss or death in transplantation, whichever occurred first. A graft loss event was defined by the patient’s return to dialysis or retransplantation. Death in transplantation is defined by patient’s death with functional graft. Secondary endpoints were graft and patient survival. Graft survival was defined as the time between transplantation and graft loss. The data of graft survival were censored at the time of death or the last time of follow up for patients alive without graft loss. Patient survival was defined as the time between transplantation and death in transplantation. The data of patient survival were censored at the time graft lost or the last time of follow up for patients without graft loss.

Median (interquartile range), mean values (standard error), and frequency (percentage) were provided for continuous and categorical variables, respectively. Medians, means, and proportions were compared using Student’s t test and chi-square test (or Fisher’s exact test, if appropriate), respectively. Transplantation failure free survival, graft survival, and patient survival were estimated using the Kaplan-Meier method and described using median or rate at specific time points with 95%CI. Follow-up duration was calculated using a reverse Kaplan-Meier estimation (12). Cox proportional hazard models were performed to estimate HR and 95%CI for factors associated with transplant failure, graft failure, and death in transplantation.

The association of 13 baseline parameters with transplant failure was first assessed using univariate Cox analyses, and then parameters with p values of less than 0.05 were entered into a final multivariable Cox regression model, after considering collinearity among variables with a correlation matrix. When used continuously in the Cox model, a potential nonlinear relationship between predictors and transplant failure was first investigated using the fractional polynomials method to determine the best transformation for continuous variables and validated by the restricted cubic splines method with graphical evaluation (13, 14). The assumption of proportionality was checked by plotting log minus- log survival curves and by cumulative martingale process plots. Accuracy of the final model was verified regarding two parameters: discrimination and calibration. The predictive value and the discrimination ability of the final model were assessed with the Harrell’s C-index (15). Random samples of the population were used to derive 95%CI bootstrap percentile for the C-statistic. Calibration was assessed by visual examination of calibration plot. Internal validation of the final model was performed with a bootstrap sample procedure (16). To assess potential bias arising from missing data for parameters involved in the multivariable final model on their significance (P-value) and estimates (b and its standard error), a multiple imputation procedure with a Markov chain Monte Carlo method was performed using SAS MI and MIANALYZE procedure. The final multivariable Cox regression model for transplant failure was applied to graft loss and death in transplantation endpoints.

Several sensitivity analyses were performed. A full multivariable model including all parameters with p values of less than 0.05 in univariate analysis was undertaken to explore the reliability and the robustness of the final multivariable model for transplant failure. A propensity score approach to deal with potential heterogeneity in baseline characteristics between patients with and without transfusion in post transplantation was applied to assess the robustness of the association of post transplantation transfusion with transplant failure evaluated in the primary Cox multivariable analysis.

To address the potential confounding effect of unbalanced factors two methods were used: 1) inverse probability treatment weighting (IPTW), and 2) propensity score matching (17). HR and their 95%CI were estimated with the IPTW cox model. The propensity score matching technique, based on the caliper method, was performed to select two samples with well-balanced characteristics at time of transplantation (17).

All analyses were performed using SAS version 9.4 (SAS Institute, Cary NC) and R software version 2.15.2 (R Development Core Team, Vienna, Austria; http://www.r-project.org). P values of less than 0.05 were considered statistically significant, and all tests were two-sided. Details on the interpretation of important statistical concepts are given in the Supplementary Methods.

The CRISTAL database identified 12,945 transplant patients having received a first kidney transplant between January 1, 2002 and December 31, 2008, among which 12,559 (97%) patients had a functional kidney allograft 15 days post-transplant (Figure 1). The linkage with the EFS database identified 3,483 (28%) patients having received an early post-transplant RBC transfusion. The majority of patients experienced one transfusion episode (n=2512, 72%), most often consisting of two RBC units (n=1739, 69%, Figure 1).

Transfused and non-transfused patients differed by several parameters (Table 1A). Nevertheless, patient follow-up was similar in both groups (7.8 95% confidence intervals, CI [7.6-8.0] vs 7.6 95%CI [7.4-7.7]; p=0.1523). Before transplantation, 1,609 patients had undergone transfusion (any history of transfusion), regardless of transfusion occurrence early after transplantation. Patients who received RBC after transplantation were more likely to have received transfusions before transplant (25.4% vs 8%; p<0.0001). Table 1B summarizes differences between patients having received RBC before transplantation and those who did not (Table 1B).

The frequency of patients receiving at least one RBC unit after transplantation remained unchanged throughout the study period (Supplementary Data and Table S1).

In univariate analysis, age, male gender, dialysis before transplantation, primary renal disease, increasing number of HLA mismatch, CMV exposure, body mass index (BMI) higher than 24.9 kg/m², donor type, and cold ischemia time > 12h were associated with transplant failure (Table 2). Both pre- and early post-transplant transfusions were also associated with transplant failure.

A correlation matrix was used to detect significant correlations between investigated parameters and to select the variable the most clinically relevant and/or with less missing data (Supplementary Data and Figure S1). Cox analysis included ten independent risk factors: recipient age, gender, BMI, pre-transplant dialysis, primary renal disease, donor-recipient HLA mismatch(es), anti-HLA immunization, CMV exposure, donor type, cold ischemia time, and pre- and early post-transplantation transfusion (Supplementary Data and Table S2). Considering the very high rate of missing data regarding CMV exposure (n=4407/12949) and the lack of association with the outcome in the full multivariable model, this parameter was not included in the final multivariable model analysis (Table 2).

Final multivariable analysis revealed that pre- and post-transplant transfusions were associated with increased transplant failure (Hazard Ratio, HR 1.635, 95%CI [1.492;1.791] p<0.0001 and HR 1.650, 95%CI [1.538;1.771] p<0.0001, respectively) (Table 2 and Figure 2).

A multiple imputation analysis based on 500 imputed data sets provided similar results for all the variables (Supplementary Data and Table S3). The multivariable model exhibited good discrimination ability (Concordance Index, C-index= 0.619, 95%CI [0.610-0.628]). The calibration plots showed an optimal agreement between model prediction and actual observation for predicting transplant failure probability at one, three, five, and ten years (Supplementary Data and Figure S2). Moreover, transplantation failure free survival curves continue to diverge several years after transfusion (Figure 2). In the internal validation, uncertainties around hazard ratio measured with a bootstrapping procedure reflected the robustness of the final model (Table 2). The multivariable model that considers separately past history of transfusion and post-transplant transfusion showed that these two parameters were both independently associated with higher transplant failure (Supplementary Data and Table S4).

A propensity score analysis was performed (Supplementary Data and Table S5A). An inverse probability of treatment weighting (IPTW) analysis, post-transplant transfusion, remained associated with a higher risk of transplant failure (HR = 1.363; 95%CI [1.264-1.471]; p<0.0001) (Supplementary Data and Table S5B). After propensity score matching of patients with similar characteristics (Supplementary Data and Table S5C and Supplementary Data and Figure S3), transplant failure rate was found to be higher in patients having received post-transplant RBC as compared to those not transfused after transplant (Supplementary Data and Figure S4).

In multivariable analysis, both pre- and post-transplant transfusions were independently associated with an increased risk of graft loss and patient death in transplantation (Table 3A, B). Dialysis before transplantation, primary renal disease, and donor type were also associated with graft loss (Table 3A). Regarding death in transplantation, age, male gender, dialysis before transplantation and primary renal disease are significantly associated in a multivariable model (Table 3B). These results are also evidenced by Kaplan-Meier graft and patient survival curves according to the post-transplantation transfusion (Figures 3A, B).