Development and validation of routine clinical laboratory data derived marker-based nomograms for the prediction of 5-year graft survival in kidney transplant recipients

Introduction

Benefiting from novel immunosuppressive agents and improved post-transplant management, kidney transplant recipients (KTRs) achieved excellent short-term survival, with a 1-year graft survival rate exceeding 95% [1]. Unfortunately, long-term graft survival has not shown synchronous improvement and remains a critical challenge for both patients and clinicians [2]. Among a variety of factors that contribute to such undesirable outcomes, the lack of a robust prediction system for late renal allograft loss is an important issue that needs to be addressed first. Without reliable evidence of risk stratification, clinicians cannot carry out efficient medical intervention in advance (such as modifying follow-up strategies and adjusting immunosuppression regimens) to prevent the upcoming deterioration of allograft function. In addition, prospective clinical trials on novel preventive or therapeutic agents might be limited due to the absence of early surrogate endpoints that can precisely estimate the risk of late graft failure in KTRs. Therefore, accurately predicting and identifying the probability of allograft loss preemptively is essential to guiding individualized clinical therapeutic decisions in KTRs.

Although a relatively large number of prognostic models have been developed and validated to predict graft failure in KTRs [3, 4], few have been widely implemented in clinical settings due to certain limitations. For example, some promising models were developed and validated on the basis of populations from Occident medical centers [5–7], where patients’ genetic backgrounds, transplant-related management strategies, and health insurance are different from those in Eastern countries. Such a gap may affect their general applicability among transplant centers around the world. Further rigorous validation is necessary or, as an alternative, new prognostic models should be constructed on the basis of local patients to facilitate native clinical application. In addition, many models only include static predictors measured at transplantation, such as recipient and donor demographics and transplant-related parameters [8, 9], or from later on during follow-up, for instance, renal function markers (serum creatinine and urine protein) at a single referring post-transplant point [10–12]. However, limited studies utilized omnibus longitudinal data as predictors [6] despite the fact that this data is the main source of clinical decision-making. Recently, some notable findings demonstrated that routine laboratory index-derived markers that reflect the overall state of a certain period time could serve as valuable predictors for poor outcomes in KTRs. Higher tacrolimus (TAC) trough concentration derived markers, TAC-intrapatient variability (TAC-IPV or TAC-CV) [13, 14], and TAC-time in therapeutic range (TAC-TTR) [15], were extensively proven to be strongly linked with subsequent composite poor outcomes in transplant patients. Furthermore, estimated glomerular filtration rate (eGFR) related novel markers including eGFR-coefficient of variation (eGFR-CV) [16, 17], as well as inflammatory markers calculated from complete blood count (CBC) data, such as neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and monocyte-to lymphocyte ratio (MLR) [18–20], were shown to be associated with adverse events in KTRs. Therefore, we hypothesized that adding these relatively new markers to prognostic models may provide additional information to improve the risk assessment of allograft loss in KTRs.

Taking into account the excellent 1-year graft survival rate and the ubiquitous follow-up strategy where CBC, urine routine test, renal function, and TAC trough concentration were regularly monitored in a majority of KTRs, we adopted the above easily obtained laboratory test data from within the first 6 or 12 months post-transplantation to generate new candidate predictors. The objective of our study was to develop and validate two prognostic nomogram models by combining laboratory data-derived risk factors with demographic and clinical variables for individually predicting the probability of 5-year kidney allograft survival in KTRs. In addition, to increase the practicability of these models, web servers and risk stratification systems were further established.

Results

Discussion

In this study, we successfully developed and validated two novel nomogram-based prognostic models that were capable of individually predicting 5-year allograft loss in KTRs. These nomograms combined early, routinely accessible, yet significant laboratory test-derived indicators (calculated based on 0–6m and 0–12m laboratory results) and demographic variables to generate accurate prognostic predictions of individual patients. Such predictions presented a good ability to discriminate KTRs with low-, moderate-, and high-risk of developing forthcoming graft loss within 1 to 5 years post-transplant.

In the last decade, a growing body of prognostic scoring systems have been reported for the early prediction of long-term renal allograft survival on the basis of various predictors, the majority of which were “static” demographical and clinical variables at certain time points before or after transplantation. Such static Cox or Logistic models incorporating variables measured at time 0 to predict the events occurred throughout the full evaluation period, which might be less effective than using longitudinal data in predicting future risk [21]. More recently, an internationally derived and robustly validated (in Europe and North America centers) prediction system for estimating the risk of graft loss in KTRs was constructed by combining 8 functional, histological and immunological prognostic factors. Strong predictive ability has been shown with C-indexes exceeding 0.8 in different clinical scenarios and subpopulations [7]. However, considering distinct patients’ characteristics, clinical management practices, and the unavailability of histological data for the majority of KTRs in most Chinese transplant centers, it may not be feasible for the implementation in our routine clinical settings. Taking these issues into account, we turned our attention to “dynamic” risk factors which were derived from serial routinely monitored laboratory examination results. We applied multiple calculation forms including average, CV, and proportions of time within or out of therapeutic range to process original laboratory data during early periods post-transplantation. By combining demographic and clinical risk factors with 0-6m and 0-12m processed laboratory predictors, we built up model 1 and model 2 prognostic nomograms, respectively. Both models showed moderate to strong predictive efficacy for early identification of 5-year graft loss, and model 2 was superior to model 1 (C-index for model 1 was 0.78 (0.72–0.85) and for model 2 was 0.85 (0.79–0.90)). We did not conduct direct comparisons between the present study and previously described risk scoring systems due to the lack of some predictor information, such as blood pressure data in DuBay DA’s study [6]. However, when only considering data from a C-index and TD-AUC perspective, our models showed higher values. Taken together, these results indicated that comprehensively integrating laboratory variables was a novel strategy to find risk predictors.

In our two prognostic nomograms, recipient characteristics including female gender, older age, and lower BMI levels were independent protective factors of long-term graft survival. However, there were some controversies regarding the associations between those individual predictors and renal allograft outcomes. The findings on being female were consistent with some of previous studies [6, 22, 23], but differed from Yohann Foucher [5] and H.Y.Tiong’s studies [24], in which being female was closely associated with poor allograft outcomes in KTRs. Such inconsistency also exists in recipient age and BMI variables among this and other different KTR cohorts [5, 22]. These discordant results suggest that constructing personalized nomograms that host “one size fits all” properties is not an easy task, which may be hampered by the heterogeneity of patient backgrounds and post-transplant management strategies.

In addition to those fixed baseline variables, some potentially amendable post-transplant risk factors were identified in this study. eGFR variability, no matter calculated from 1–6 months eGFR values or 1–12 months eGFR values, was a significant risk factor for graft loss and independent of average eGFR levels. In accordance with Nicole A. Pilch’s [16] and Hoon Young Choi et al’s. [17] studies, our results suggested that converting serial longitudinal eGFR data into eGFR-CV provided more prognostic information than just simply averaging them. Several possible factors may explain its superior ability in early presentations of signs of graft loss. eGFR variability is a rational marker of renal allograft homeostasis. Intrinsic renal factors (such as acute allograft rejection, subclinical rejection, and other damages that directly target the kidney) and extrinsic factors (such as infection, extra-renal organ system injury, and comorbid complications that indirectly affect renal function) may disturb the steady state of the kidney [25, 26]. Even though some deteriorative renal function may revert to its prior “normal” level after appropriate intervention, eGFR variability can still provide comprehensive information on both unmeasured and measurable detrimental events that may have side effects on subsequent graft survival. TAC intra-patient variability (TAC-IPV or TAC-CV) was another important risk factor for graft loss in our study. Accumulating studies have demonstrated that high TAC-CV was associated with poor allograft outcomes such as graft failure and chronic antibody-mediated rejection in various solid organ transplantations [13, 27–29]. Unlike previously published studies, and since we wanted to use TAC-CV to reflect the overall combined impact of medication non-adherence, food and drug interaction, genetic factors, and dose modification due to concurrent diseases or TAC-related toxicity throughout the entire interested post-transplant period [14], TAC-CV was calculated by adopting all TAC trough concentration data from within 0–6 or 0–12 months, rather than eliminating early-phase data that was thought to be unstable because of the frequent dose adjustment to achieve target levels during hospitalization. Final inclusion of TAC-CV in model 2, not in model 1, indicated that replenishing early TAC-CV with relatively late TAC concentration data would be conducive to improve its predictive ability for 5-year graft loss. Interestingly, we found a high average percentage of 0–6 months neutrophil and low 0–12 months lymphocyte mean count emerged as independent risk indicators for subsequent graft loss in different prediction models. Such abnormality of average peripheral leukocyte count may reflect the sustained excessive immunosuppression during certain time periods, which increases the risk of various infections and then threatens the long-term survival [30–32].

Several strengths could be found in the current study. First, our study took full advantage of easily accessible laboratory data during 6 or 12 months post-transplantation to generate novel and cost-effective potential predictors, some of which have been revealed to significantly correlate with poor long-term allograft survival in distinct population-based studies. Nomograms incorporating an early shorter or relatively longer period of laboratory indicators with demographic variables both demonstrated good discrimination and calibration power, which enabled the early identification of individual graft-loss risks without additional cost. Second, we set up the risk thresholds that can classify KTRs into low-, moderate-, and high- risk groups as early as 6 or 12 months after transplantation. This would be helpful to promote personalized health evaluations and carry out precise prevention by adjusting follow-up frequencies, adding protocol biopsies, or modifying immunosuppression regimens. In addition, it would also be favorable in conducting prospective clinical trials by allowing targeted recruitment of moderate- to high-risk patients. Third, considering the clinical practice, nomograms were further converted into web-based apps and predictor calculators were designed to facilitate clinical application.

However, several limitations should be addressed in this study. First, due to the nature of retrospective design, the small number of events, and the single center analysis, the results may be somewhat heterogeneous. Although we have validated these results in a random split-sample cohort, external validation studies from other institutions, or at least from other Chinese institutions, are definitely warranted to confirm these findings. Second, some features would potentially improve the predictive performance of models, such as warm and cold ischemia time and donor specific antibody levels, which were not included in this study due to the incomplete records in our database. Third, these models were only applicable for adult KTRs who maintained their graft function within the first year. However, for those who lost the grafts during the first year, pre- and peri-transplant, as well as early post-transplant parameters, would be feasible to achieve good predictive performance for this patient population.

In summary, we constructed and validated two novel nomograms by incorporating clinical variables and relatively new risk factors that derived from a serial of easily-available laboratory examination data from within the first half- or one-year post-transplant to identify the risk of 5-year graft loss in KTRs early. In addition, the nomogram-based classifiers could provide supportive information on stratifying individual patients into different risk groups, which may assist clinicians and patients in clinical decision-making and adjustment of post-transplant management strategies. Further external validation and prospective application of these models will be performed in our future studies.

Materials and Methods

Supplementary Materials

Supplementary Figures