Since February 27, 2002, organ allocation for liver transplantation in the United States has been based on the model for end-stage liver disease (MELD). This prognostic index was originally developed by retrospective multivariate analysis of predictors of 90-day mortality in cirrhotic patients undergoing transjugular intrahepatic portosystemic shunt placement.1 Subsequently a modified version, calculated according to the formula shown in Table 1, was developed for use in liver transplantation.2–5 High MELD scores are strongly associated with early pretransplant wait-list mortality. While MELD may underestimate mortality risk in patients who have persistent ascites and hyponatremia,6–8 it is at least as accurate in predicting short-term cirrhotic survival as more traditional measures of cirrhosis severity, such as the Child-Turcotte-Pugh score.4, 9, 10 Furthermore, it has the advantage of being entirely based on objective laboratory values.
Table 1. The Model for End-stage Liver Disease Score
Ln = natural logarithm (base e)
Cr = creatinine (mg/dL)
INR = international normalized ratio
B = bilirubin (mg/dL)
E = etiology (0 = alcohol, cholestatic; 1 = all other)
MELD modifications adopted by the United Network for Organ Sharing (UNOS) for liver transplant organ allocation in the U.S., implemented 2/27/02.
Minimum values of Cr, B, and INR = 1
Maximum value of Cr = 4
Cr set to 4 if patient received dialysis x > 2 in preceding week
E = 1 for all patients
Score rounded to nearest whole number
In some situations, however, the MELD score may not be a valid prognostic index. One such situation is the patient receiving anticoagulant medications. Warfarin interferes with vitamin K–dependent gamma carboxylation of clotting factors, causing the international normalized ratio (INR) to be elevated.11 By dissociating INR from liver function warfarin therapy largely nullifies the prognostic value of INR. Therapeutic anticoagulation, with the goal of increasing INR to levels as high as 3, may add more than 12 points to the MELD score, conveying a substantial advantage to these patients under the existing system of organ allocation. We have derived an alternative index, normalized to the same scale as MELD but omitting INR, that may permit more appropriate assignment of transplant priority in patients on anticoagulant therapy.
MELD, model for end-stage liver disease; INR, international normalized ratio; DVA, Department of Veterans Affairs; MELD-XI, MELD excluding INR.
PATIENTS AND METHODS
Subjects for analysis (training data set) consisted of 554 cirrhotic veterans referred for consideration of liver transplantation in the Department of Veterans Affairs (DVA) between January 1, 1997, and November 31, 2003. These patients have been characterized more fully in a previous publication.6 Patients known to be receiving oral anticoagulants were excluded. Additional patients were excluded if referred for fulminant hepatic failure, if diagnosis of cirrhosis was suspect or absent, if the patient had undergone a prior transplant, or if insufficient data (concurrent within 30 days) were available to determine MELD scores. All patients included in the analysis were felt by their referring physicians to have been abstinent from alcohol and drug abuse for at least six months (a precondition for transplantation in DVA programs). Initial validation was performed on a holdout dataset of 278 additional patients referred for consideration of liver transplantation through the DVA between December 1, 2003, and October 1, 2005. Survival was determined from the date of the most current liver function test results available at the time of referral. Cases were censored at transplantation. Censoring was not contingent on transplant waiting list status: in addition to patients on the waiting list, veterans who died prior to listing, had not been listed at the time of analysis, or were withdrawn from the transplant list were included in the analysis. Outcome (date of death or transplantation) for all patients was confirmed via both the DVA central office transplant database (updated through March 1, 2006) and the Social Security Death Index (updated in February 2006).
Additional validation was carried out using an independent dataset obtained from the United Network for Organ Sharing, consisting of all patients who were formally listed and awaiting liver transplantation in the United States on May 1, 2001, or who were subsequently added to the transplant list between May 1, 200101 and November 1, 2001, with follow-up through July 22, 2001.9 In this independent validation set, we included only adult cirrhotic patients (age >21 and diagnostic codes 4200–4245) with concurrent (same day) laboratory data obtained between May 1 and November 11, 2001, permitting calculation of true MELD scores; patients with prior liver transplantation, fulminant hepatic failure, noncirrhotic indications for transplantation, or carcinoma as an indication for transplantation were excluded. Patients with cholestatic causes of cirrhosis (primary biliary cirrhosis and primary sclerosing cholangitis) were analyzed separately from noncholestatic cirrhotics. Survival was determined from the date of the most current laboratory available on November 1, 2001, until the date of death. Outcomes in this dataset were censored at date of transplantation or date of withdrawal from the waiting list. In additional analyses reported separately, patients withdrawn from the transplant list because they became too sick to undergo transplantation were treated as deaths on the date of waiting list withdrawal. These studies conformed to the ethical guidelines of the 1975 Helsinki declaration. Data collection and analysis were undertaken with the approval of the McGuire Research Institutional Review Board.
Statistical analyses were performed using the Statistical Package for the Social Sciences (SPSS) release 11.5 for Windows (SPSS, Inc., Chicago, IL). Initially to derive an alternative prognostic index based on bilirubin and creatinine, we performed binary logistic regression analysis of these 2 variables as predictors of 90-day pretransplant survival in the training group, using the forward stepwise approach. We employed the natural logarithms of bilirubin and creatinine rather than the native values to be consistent with the MELD score. The resulting predictive formula was normalized to the same scale as the MELD score by linear regression. We termed the resulting prognostic index MELD-XI (for MELD excluding INR). To assess the predictive accuracy of this new prognostic index, we compared MELD-XI to the original MELD score as predictors of pretransplant mortality at time points ranging from 30 to 180 days by determining areas under receiver operating characteristic curves (concordance or c-statistics).
Characteristics of Training and Validation Populations
Features of the patient populations in the training, holdout validation, and independent validation datasets are summarized in Table 2. The training and holdout groups, consisting of patients from the DVA health system, were predominantly males with noncholestatic liver disease. Mean MELD scores were 16.2 and 15.9, respectively. Etiology of liver disease was hepatitis C and/or past ethanol use in more than 85%. The independent validation set derived from the national U.S. transplant waiting list in the pre-MELD era was healthier (mean MELD, 13.7) and more diverse (40% were women). It also included a substantial minority (13%) of patients with chronic cholestatic conditions such as primary biliary cirrhosis and primary sclerosing cholangitis, sufficient to permit subgroup analysis. In the veteran population, pretransplant 90-day mortality was 13.6% in the training cohort and 12.1% in the hold out validation cohort. In the United Network for Organ Sharing wait-list independent validation dataset, 90-day pretransplant mortality was 3.9%, with an additional 0.6% of patients withdrawn from the waiting list within 90 days because they became too sick to transplant. In both groups, mortality was strongly associated with elevations of each of the MELD components. Individual c-statistics ± SE for bilirubin, creatinine, and INR as predictors of 90-day pretransplant mortality in the DVA patients were 0.777 ± 0.028, 0.782 ± 0.027, and 0.756 ± 0.027, respectively. In the United Network for Organ Sharing dataset the c-statistics for these 3 parameters were 0.811 ± 0.016, 0.711 ± 0.020, and 0.788 ± 0.016 respectively.
Table 2. Patient Characteristics
Holdout Validation Set
Independent Validation Set
NOTE: The training and holdout validation patients were cirrhotic veterans referred for consideration of liver transplantation in the Department of Veterans Affairs in the periods 1/1/97–11/30/03 (training) or 12/1/03–9/30/05 (holdout validation). The independent validation set from the OPTN database includes adult cirrhotic patients listed for transplantation in the United States between 5/1/01 and 11/1/01. See text for details.
We performed logistic regression analysis of 90-day pretransplant mortality in the training dataset using logarithmic transformations of bilirubin and creatinine and excluding INR. Following the customary practice of the United Network for Organ Sharing, values of bilirubin and creatinine less than 1 were set to 1 to avoid negative values, values of creatinine greater than 4 were capped at 4, and creatinine was set to 4 in patients receiving dialysis (Table 1). This analysis confirmed that bilirubin and creatinine were independent predictors of 90-day pretransplant survival (each P < 0.001). Optimal discrimination was provided by the formula
D = 1.437 Ln(bilirubin) + 3.304 Ln(creatinine) − 4.409.
Odds ratios for 90-day mortality associated with 1 unit increases in Ln(bilirubin) and Ln(creatinine) with 95% confidence intervals were 4.2 (range, 2.8–6.4) and 27.0 (range, 11.8–62.5), respectively.
By linear regression, we found that the best linear fit between D (as independent variable) and the MELD score (dependent) in the training set was provided by a line with slope 3.558 and y-intercept 25.13. Correlation between the two was good, with r = 0.930. Based on this analysis, we concluded that the best estimate of MELD score in the absence of INR (MELD-XI) is obtained by the formula MELD-XI = 3.558 D + 25.13, or, substituting for D, MELD-XI = 5.112 ln(bilirubin) + 11.76 ln(creatinine) + 9.44.
Validation of MELD-XI
MELD and MELD-XI values correlated well in all 3 groups, as shown in Figure 1. In the training group a regression line relating MELD on the horizontal axis to MELD-XI on the vertical axis yielded correlation coefficient r = 0.930; by definition, slope = 1 and y-intercept = 0. In the holdout validation, correlation coefficient was 0.954, slope 0.870 ± 0.016 and y-intercept 2.42 ± 0.29 In the independent validation group, correlation coefficient was 0.902, slope 0.814 ± 0.005, and y-intercept 3.546 ± 0.068.
In the training group, MELD-XI and MELD were comparable as discriminators of pretransplant survival at all time points from 30 to 180 days (Table 3 and Fig. 2). C-statistics of the 2 indices differed by less than 0.01 at all time points from 30 to 180 days. In the holdout validation group, MELD was slightly more accurate than MELD-XI for predicting 30-day survival (c-statistic ± SE 0.939 ± 0.023 vs. 0.906 ± 0.033, respectively), but differences narrowed at 60 days (0.860 ± 0.049 vs. 0.841 ± 0.050) and 90 days (0.842 ± 0.049 vs. 0.829 ± 0.048) and the 2 scores were equally accurate as predictors of 180-day mortality (0.795 ± 0.044 vs. 0.797 ± 0.043).
Table 3. Area Under Receiver Operating Characteristic Curve for MELD vs. MELD-XI as Predictors of Pretransplant Mortality at Various Time Points: Training and Holdout Validation Datasets
Pretransplant mortality time point
Training Set (n = 554)
Holdout Validation Set (n = 278)
NOTE: Data presented as c-statistics ± SE.
0.924 ± 0.030
0.927 ± 0.028
0.939 ± .023
0.906 ± .033
0.883 ± 0.027
0.886 ± 0.025
0.860 ± .049
0.841 ± .050
0.860 ± 0.029
0.865 ± 0.027
0.842 ± .049
0.829 ± 0.048
0.817 ± 0.025
0.811 ± 0.024
0.795 ± 0.044
0.797 ± 0.043
In the independent validation group, MELD-XI was comparable to MELD as a predictor of pretransplant 90-day survival (Table 4 and Fig. 3). In these analyses, cases were censored at time of transplantation or withdrawal from the waiting list. C-statistics for MELD and MELD-XI were comparable in patients with cholestatic liver diseases (0.905 ± 0.030 vs. 0.894 ± 0.031) as well as noncholestatic causes of cirrhosis (0.857 ± 0.016 vs. 0.843 ± 0.016). Similarly, when patients withdrawn from the list because they became too ill to undergo transplantation (withdrawn sick) were treated as deaths at the time of waiting list withdrawal, c-statistics for MELD vs. MELD-XI as predictors of 90-day survival in cholestatic patients were 0.911 ± 0.027 and 0.896 ± 0.029, respectively, while c-statistics in noncholestatic patients were 0.848 ± 0.015 and 0.837 ± 0.015, respectively.
Table 4. Area Under Receiver Operating Characteristic Curve for MELD vs. MELD-XI as Predictors of Pretransplant Mortality at Various Time Points: Independent Validation Dataset of 7,203 Adult Cirrhotic Patients Awaiting Transplantation in the United States Through the United Network for Organ Sharing, May 1 to October 31, 2001
Independent Validation Set
NOTE: Data presented as c-statistics ± SE. Data were censored at waiting list withdrawal or transplantation. Mortality is determined in 2 ways: time until death (censored at “withdrawn sick”) or time until either death or “withdrawn sick” as endpoints.
90 days (death)
90 days (death or withdrawn sick)
In the DVA population, data were recorded prior to transplant referral or listing, and the fraction of patients undergoing transplantation within 90 days was small (2.0%) relative to 90-day pretransplant deaths (13.3%). Table 5A shows the distribution of MELD-XI and MELD scores in the combined DVA patient groups (training and holdout validation, n = 816 after censoring for early transplants), and their association with early pretransplant mortality. In each range of scores, mortality of MELD and MELD-XI were comparable. The 90-day pretransplant mortality in the subset of 68 patients with MELD scores greater than 25 was 75%; in the 57 patients with MELD-XI >25, 90-day pretransplant mortality was 79%. In the independent validation population (Table 5B), consisting of patients already listed for transplantation, overall mortality in each range of MELD scores was lower, reflecting in part the much larger proportion of patients transplanted within 90 days (9.0% transplanted vs. 3.9% who died or were withdrawn sick). Once again, however, equal MELD and MELD-XI scores were associated with comparable prognosis.
Table 5. Pretransplant 90-day mortality as a function of MELD or MELD-XI. Data for Dept. of Veterans Affairs patients are from combined training and holdout validation sets (n = 823) censored at time of transplantation (n = 14). In analyzing data from U.S. national transplant list, patients “withdrawn sick” (n = 45) are counted as deaths at time of waiting list withdrawal; other cases are censored at transplantation (n = 645) or waiting list withdrawal (n = 43)
MELD 90d mortality
MELD-XI 90d mortality
Dept. of Veterans Affairs (training and holdout), deaths
U.S. national transplant wait list 5/1–10/31/01 (independent validation), death or withdraw sick within 90 days
These data demonstrate that most of the predictive value of MELD resides in 2 of its components: bilirubin and creatinine. A modified MELD score, MELD-XI, calculated based solely on these 2 parameters, is almost as accurate as MELD for predicting pretransplant mortality. Because MELD-XI scores are normalized to the same range and distribution as MELD scores, MELD-XI can reasonably be substituted for MELD in those situations where INR is unavailable or unreliable as a test of liver function, and MELD-XI would be expected to provide a better indication of prognosis in these situations.
There are several circumstances under which INR may be misleading as a liver function test. The most obvious of these is warfarin anticoagulation. Warfarin use is not uncommon in patients with liver disease, related to indications as diverse as hepatic venous thrombosis,12 pulmonary hypertension,13 portal venous thrombosis,14 atrial fibrillation, valvular heart disease, deep venous thrombosis, pulmonary embolism, and hypercoagulable states.15 Dissociation of INR from liver function also occurs when conditions such as cholestasis, pancreatic insufficiency, or small intestinal disease cause vitamin K malabsorption, unless patients receive adequate parenteral vitamin K supplementation.16 In disseminated intravascular coagulation, prolongation of the INR reflects clotting factor consumption and anticoagulant effects of fibrin split products rather than synthetic liver function.17, 18 Conversely administration of fresh frozen plasma or factor VIIa to prevent or control hemorrhage transiently lowers the INR and may cause MELD to underestimate true disease severity.19 A prognostic index that excludes the INR avoids these concerns.
The allocation of transplant priority in patients receiving warfarin has been an area of contention in the United States in the 4 years since implementation of MELD-based organ allocation. Because of the lack of a suitable alternative, the MELD score in these patients has remained the basis for priority. While deliberate manipulation of the INR through warfarin administration probably is uncommon, physicians are aware that warfarin treatment increases their patients' likelihood of being transplanted, and this fact undoubtedly influences decisions regarding anticoagulation. A policy of substituting MELD-XI for MELD in any patient who has received anticoagulants in the previous week would eliminate this undesirable incentive and permit more rational risk assessment and organ allocation. Because MELD-XI is normalized to the same scale as MELD and carries equivalent prognostic information, the substitution could be made automatically and would not require a MELD exception. Programs for calculation of MELD could easily be modified to include a question regarding current use of anticoagulants, and to automatically substitute MELD-XI for MELD in this situation.
There is a moderate degree of systematic variability between laboratories in measurement of INR, and this variability has been identified as a potential area of bias in organ allocation.20, 21 One immediate application of MELD-XI may be in monitoring and correcting for such discrepancies. By measuring the gap between MELD and MELD-XI in large numbers of samples from many patients, oversight agencies can readily compare institutions and identify those that deviate significantly from the norm. Additional testing could then be undertaken to determine whether the methods employed to determine INR are the source of the difference. Where systematic biases are confirmed, corrective action, such as introduction of an institutional correction factor for INR values or mandated use of a reference laboratory, may be appropriate.
Our data, though collected prospectively, were analyzed retrospectively, and the populations studied may not be universally representative. Prospective confirmation in other populations is desirable. In certain conditions, such as Budd-Chiari syndrome, it can be argued that the MELD-XI score may underestimate risk and the higher MELD priority resulting from anticoagulant treatment may be justified; current data are insufficient to confirm or refute this hypothesis. Finally, inclusion of other predictors such as hyponatremia or persistent ascites may improve the predictive accuracy of MELD and may need to be considered in calculation of MELD-XI as well.6–8 Despite these limitations, we propose that MELD-XI fills a void by offering a rational approach to assessment of prognosis in cirrhotic patients who require treatment with anticoagulants.
The authors thank Brenda Salvas, Transplant Program Manager, for the Department of Veterans Affairs, for her assistance in tracking outcomes of patients referred for consideration of liver transplantation, and Denise Tripp and Patricia Paquette of the United Network for Organ Sharing in Richmond, VA, for providing data from the Organ Procurement and Transplant Network permitting independent validation.