Kidney transplantation (KT) is optimal treatment for end-stage renal disease (ESRD), whether for the patient’s survival, his quality of life or the cost of treatment (1,2,3).

Over the last two decades, the number of KTs has increased worldwide, in the number of transplantations from both living donors (4) and cadaveric donors (5,6). This phenomenon is due to an increase of patients with kidney failure, largely linked to longer life expectancy, and to an increase in prevalence of type 2 diabetes (7). As a result, there has been a situation of kidney transplant shortage for several years which led to a rise in the number of patients on KT waiting list (6).

Reducing the need for repeated transplantation by avoiding graft loss (GL) is a major objective. Despite progress in post-transplant management, particularly in terms of immunosuppressive treatments (8,9,10), long-term graft survival has only slightly improved over recent years (10,11), with a median estimated at 14 years in France (7).

Many predictive factors of long-term graft survival have been identified in recent years such as ischemia-reperfusion syndrome (12,13), delayed graft function (DGF) (14,15). Currently, post-transplant recovery assessment of the graft is based on consensual but not very reliable parameters. Plasma creatinine is a marker of interest for the clinician, but data on its kinetics in the first days following KT are scarce. Nadir creatinine and time to reach a creatinine level below 2.83 mg/dL during the transplant patient’s hospital stay are parameters commonly used by clinicians since they are considered to be independent prognostic factors for graft survival.

Daily creatinine values during the transplant patient’s hospital stay are available and consideration of their dynamic over time could yield additional predictive information.

Many statistical models have been developed to study the evolution of quantitative markers over time. One of them is the latent class mixed model (LCMM), which is an extension of the standard linear mixed model (16). The model computes the heterogeneity of individual trajectories and identifies subgroups of patients with similar trajectory profiles, independently of their observed characteristics. This method has shown its interest in identifying subgroups of trajectories of renal function in a cohort of patients with chronic kidney disease (CKD) (17) and, more recently, in diabetic patients (18,19). While the LCMM has been applied for repeated measures of creatinine or glomerular filtration rate (GFR) (20), it has never been used to describe subgroups of trajectories of renal recovery in the days following transplantation in renal recipients. LCMM could reveal unknown heterogeneity in post-KT trajectories of renal function which could help identify kidney recipients at high risk of GL.

Therefore, our aims were 1) to evaluate the ability of the LCMM to identify subgroups of trajectories of creatinine immediately following renal transplantation, 2) to identify the clinical factors associated with the different trajectories and 3) to test the association between subgroups of trajectories and graft outcomes.

In the ASTRE cohort, 496 patients underwent a rank 1 KT at the Poitiers University Hospital. After applying the selection criteria, 435 recipients (258 males, 177 females aged 56 years) with kidney transplants from a DBD donor were included in the analysis (Figure 1).

Recipients, donors and transplants background characteristics are displayed in Table 1.

Between 24 h post-KT and hospital discharge, 6,211 serum creatinine measurements were recorded in the 435 patients corresponding to a median of 12 (11,12,13,14,15) measurements per patient per stay in agreement with median hospital stay duration. As shown in Table 2, delayed graft function was observed in 80 patients (18%) and a median number of 2 (2,3,4) hemodialysis sessions was needed during the hospital stay.

During median follow-up of 73 (48–107) months, 68 patients (16%) lost their graft. At year one 7 GL (2%) and 39 T cell acute rejections (9%) were recorded.

In the present study, using a retrospective cohort of 435 kidney transplant recipients, we performed a LCMM and highlighted substantial heterogeneity in renal function recovery immediately following KT from DBD. Interestingly, we identified four distinct trajectories showing non-linear decrease in daily creatinine levels during the initial hospitalization for transplantation.

These trajectories showed heterogeneity between patients regarding both creatinine level at day 1 and speed of the decrease, which reflects graft recovery. Patients with “optimal” or “good recovery” represented 37% and 10% of the study population, respectively, and their creatinine values dropped quickly. Patients with “intermediate recovery” had higher creatinine values at day one and a slower decrease over time, and they were the most numerous, representing 47% of the population. Finally, 6% of patients presented a trajectory markedly different from the others, with high initial values and a delayed and small decrease (“poor recovery”). The slight increase observed in the trajectory of the “optimal recovery” group could be due to the renal toxicity of the calcineurin inhibitors.

Although baseline characteristics were poor predictors of these trajectories, some clinically relevant variables such as cold ischemia time were associated with the trajectories. Interestingly, kidney donor hypertension, but not ECD, was also associated with the trajectories: in the “optimal recovery” trajectory, the lowest cold ischemia time and rate of hypertensive donors were found. The use of machines perfusion did not differ significantly between the different latent classes.

The four trajectories were also associated with relevant post-KT characteristics. As expected, patients classified in the “poor recovery” class were more likely to have DGF. Furthermore, they had the highest nadir creatinine and the longest time in days needed to reach the creatinine threshold of 2.83 mg/dL. “Good” and “optimal recovery” trajectories did not differ significantly in terms of time needed to reach the 2.83 mg/dL creatinine threshold.

Regarding graft outcome, the four trajectories were found to be an independent prognostic factor. In the literature, there are a few risk scores predicting graft loss, which include mainly donor and baseline transplant characteristics as prognostic factors (27,28). Cold ischemia time is identified as a risk factor for reduced graft and patient survival and DGF (29). The hypothermic perfusion machine also reduces the DGF incidence from 38% to 24% (30). Moreover, it is demonstrated that the survival of grafts at 1 year is 92.3% for grafts placed on a hypothermic machine perfusion, versus 80.2% for grafts with cold storage (31). During the first few months post-KT, 3-, 6- and 12-month eGFR have been reported as prognostic values (32) such as creatinine and eGFR trajectories beyond one year of follow-up (20,33). Nevertheless, none of these risk scores takes into consideration the evolution of daily creatinine levels in early post-KT. To our knowledge, our study is the first to characterize renal recovery trajectories immediately following KT, using the latent class analysis tool. Consideration of these daily determinations of creatinine during the hospital stay could be more relevant to predict GL than isolated values such as nadir of creatinine level.

The use of LCMM for the identification of trajectories of renal function evolution has raised a lot of interest in recent years (17,34). This novel unsupervised approach overlooks prior assumptions on the evolution of creatinine as a means of classifying patients. Our model was fitted on several covariates known to influence renal recovery, and this statistical fitting results in independency between these covariates and the identified trajectories. More specifically, the impact of hemodialysis sessions was taken into account as a time-dependent variable since it induces a drop in creatinine levels that gradually affects the creatinine levels, up to 24–48 h after the session. In a sensitivity analysis, we applied the model on the subset of data free of post-hemodialysis creatinine values and obtained four latent classes comparable to the previous ones.

Regarding the clinical applicability of our models, the type of recovery for a given recipient could be estimated outside the framework of this study. Indeed, the available hospital stay creatinine values and the covariates present in the model could be calculated in a dedicated application which would indicate the class membership probabilities. Hence, transplant recipients at higher risk of graft loss could be identified upon their hospital discharge, enabling personalized follow-up frequency and better management.

This study has several limitations, particularly its monocentric design. Selection bias regarding recipients and donors might have conditioned class-membership. External validation is needed to confirm the general applicability of a LCMM. Moreover, an analysis conducted on Maastricht-III and living donors could be of particular interest.

In conclusion, substantial inter-individual heterogeneity in creatinine trajectories immediately following KT was highlighted. Insight into a recovery class might lead to more precise estimation of risk of graft loss for a given patient and be conducive to optimized post-KT management.