The first 2 authors contributed equally to this article.
An integrated MELD model including serum sodium and age improves the prediction of early mortality in patients with cirrhosis†
Article first published online: 30 JUL 2007
Copyright © 2007 American Association for the Study of Liver Diseases
Volume 13, Issue 8, pages 1174–1180, August 2007
How to Cite
Luca, A., Angermayr, B., Bertolini, G., Koenig, F., Vizzini, G., Ploner, M., Peck-Radosavljevic, M., Gridelli, B. and Bosch, J. (2007), An integrated MELD model including serum sodium and age improves the prediction of early mortality in patients with cirrhosis. Liver Transpl, 13: 1174–1180. doi: 10.1002/lt.21197
See Editorial on Page 1080
- Issue published online: 30 JUL 2007
- Article first published online: 30 JUL 2007
- Manuscript Accepted: 14 MAR 2007
- Manuscript Received: 20 SEP 2006
- ISCiii. Grant Numbers: C03/02, PI 06/0623
The Model for End-Stage Liver Disease (MELD) is widely used to predict the short-term mortality in patients with cirrhosis, but potential limitations of this score have been reported. The aim of this study was to improve the score's prognostic accuracy by assessing new objective variables. Data of 310 consecutive patients with cirrhosis who underwent elective transjugular intrahepatic portosystemic shunt placement between July 1995 and March 2005 were analyzed retrospectively. Bivariate and multivariate analyses were performed by proportional hazard Cox regression models. The area under the receiver operating characteristic curve (auROC) and the likelihood ratio test were used to evaluate the performance of the models for predicting early mortality. Findings were validated in a cohort of 451 consecutive patients with cirrhosis on waiting list for liver transplantation. Bivariate analyses showed that the following variables correlated with time to death: age, serum bilirubin, serum creatinine, international normalized ratio of prothrombin time, serum albumin, serum sodium, and MELD. Multivariate analysis revealed that MELD, serum sodium, and age were independently associated with the risk of death. The integrated MELD model (iMELD, incorporating serum sodium and age) was better than original MELD in predicting 12-month mortality: auROC increased by 13.4% and the likelihood ratio statistic from 23.5 to 48.2. The improved accuracy of iMELD was confirmed in the validation sample of 451 patients with cirrhosis on the waiting list for liver transplantation by increasing auROC (+8%) and likelihood ratio statistic (from 41.4 to 82.0). This study shows that in patients with cirrhosis, serum sodium and age are predictors of mortality independent of the MELD score. The incorporation of these variables into the original MELD formula improves the predictive accuracy of time to death. Liver Transpl 13:1174–1180, 2007. © 2007 AASLD.
The Model for End-Stage Liver Disease (MELD) was developed as a prognostic model of short-term mortality in patients with cirrhosis treated with transjugular intrahepatic portosystemic shunt (TIPS).1 The original score was subsequently simplified,2 and it is widely used to predict the short-term mortality in different populations of patients with cirrhosis.3–8 In the United States and in many European centers, the MELD score has replaced the Child-Pugh score for priority allocation in waiting lists for liver transplants.9, 10 The MELD score has the advantage over the traditional Child-Pugh score of being based on objective and readily available variables (serum bilirubin, international normalized ratio of prothrombin time [INR], and serum creatinine) rather than on subjective assessment of degree of clinical abnormalities without universally accepted definitions, and thus widely variable.
However, over the last few years, many investigators have pointed out potential limitations of the MELD score,11–15 which has led to the recommendation of performing additional studies to evaluate the effect of incorporating other objective variables into the model.10, 13, 16 Recent studies have shown that the MELD score's prognostic accuracy may improve by adding additional variables such as serum sodium, ascites, and the hepatic venous pressure gradient.17–20
The aim of the present study was to further improve prognostic modeling by assessing new objective variables in a large cohort of consecutive patients undergoing TIPS in 2 tertiary-care institutions. The model was validated in a large, independent series of patients with cirrhosis entering a waiting list for liver transplantation.
PATIENTS AND METHODS
This study is a retrospective cohort study based on the analysis of 310 consecutive white patients with cirrhosis who underwent elective TIPS placement from July 1999 to March 2005 at 2 tertiary-care European centers (ISMETT, Palermo, Italy [n = 109]; University of Vienna Medical School, Vienna, Austria [n = 201]). The study was approved by the review board of each institution. Informed consent was not specifically required for the study, although written informed consent was obtained for every TIPS procedure. Elective TIPS placement was defined as TIPS being performed in patients with cirrhosis for the prevention of recurrent episodes of variceal bleeding, or for the treatment of refractory ascites or hepatic hydrothorax while the patient was hemodynamically stable. Recurrent variceal bleeding and refractory ascites were defined according to accepted consensus guidelines.16, 21 Hepatic hydrothorax was defined as recurrent pleural effusion that required repeated thoracocentesis to control symptoms despite diuretic therapy. In both institutions, hepatic encephalopathy, hepatocellular carcinoma (HCC), cardiopulmonary comorbidity, active infections, or organic renal disease were contraindications for TIPS.
Only objective variables were entered into the analysis. Clinical variables included age, gender, cause of liver disease, and date of surgery in the case of liver transplantation. Ascites and encephalopathy were not included because their grading is based on subjective evaluation.13, 15 Biochemical variables included serum bilirubin, serum creatinine, INR, and serum albumin. Serum sodium was included in the analysis because its prognostic role has been recently emphasized in patients with cirrhosis listed for liver transplantation.17–19 All laboratory data were determined within the 72 hours preceding the TIPS procedure. In each patient, the MELD score was calculated according to the modified Malinchoc formula6: R = 9.57 loge (creatinine [mg/dL]) + 3.78 loge (bilirubin [mg/dL]) + 11.2 loge (INR) + 6.43. TIPS procedures were performed as previously described by standard techniques.4 All patients were followed up at the outpatient clinic. TIPS patency was investigated at 3 months, 6 months, and every 6 months thereafter by Doppler ultrasound; TIPS was revised whenever occlusion/dysfunction was suspected.
Proportion was used as descriptive statistic for categorical variables, and median and interquartile range for ordinal and continuous variables; 95% test-based confidence intervals (95% CI) were computed for each estimate of interest. Categorical variables were compared with the χ2 test,22 whereas survival probabilities were computed by the Kaplan-Meier method and compared with the log rank test.23
Because the MELD score was developed by means of the Cox proportional hazard model, we adopted the same approach to improve the score's prognostic accuracy. First, we performed bivariate analysis between time to death and possible explanatory variables by the Cox proportional hazard model.23 The main aim of this step was to obtain a value of hazard ratio (HR) for each variable in order to assess the presence of confounders in the subsequent multivariate model. Then, we developed a Cox multivariate model following a mixed backward-forward approach; different models were compared by the likelihood ratio test.23 All available variables were tested, without any preselection criterion.24 Multivariable fractional polynomials25 were used for modeling the functional form of the relationship between the continuous covariates and the outcome. In all time-to-death analyses, patients were censored at the last follow-up date or at the date of liver transplantation. The proportional hazard assumption was checked with the graphical approach of comparing log-log survival curves and by using time-dependent covariates in an extended Cox model.23
The parameters of the final model were used to weight the contribution of each variable in improving the prognostic accuracy of the original MELD. Each parameter was divided by the MELD's parameter and rounded to the nearest integer; the result was used as multiplier of the corresponding variable, and a constant was eventually added to avoid negative numbers. By means of this approach, we obtained the mathematical function to calculate the integrated MELD (iMELD), which integrates the original MELD and the new variables with independent prognostic values.
Two logistic regression models with MELD and iMELD as the only covariates were used in assessing 12-month mortality. Patients censored before 12 months were excluded from these models. The models were compared by the likelihood ratio test.23 Because this test follows the χ2 distribution, the higher the value, the better the model. Moreover, the area under receiver operating characteristic curve (auROC)26 was used to assess the accuracy of the iMELD and to further compare this with the original score in predicting mortality at 3, 6, and 12 months of follow-up. The 95% CIs of the auROC curve was computed with the bootstrap method.27 All analyses were performed with SAS version 9.1.2 (SAS Institute, Cary, NC).
Validation of the Integrated MELD
The performance of iMELD model was evaluated in a data set of 451 consecutive white patients with cirrhosis on the waiting list for liver transplantation from July 1999 through December 2005 at ISMETT, Palermo, Italy.
Both the original MELD and iMELD were calculated in this cohort of patients. Cox proportional hazard and logistic regression models that used each score as the only covariate were applied. The logistic regression model was performed on 12-month mortality; patients censored before 12 months were excluded from this analysis. To compare the 2 scores, the corresponding models were assessed through the value of the likelihood ratio test.23 The auROC26 was used to compare the ability of the 2 scores in discriminating patients who died from those who were alive at 12 months of follow-up.
Characteristics of the Patients and Clinical Outcome
The clinical characteristics of the 310 patients included in the study are listed in Table 1. During the first 12 months of follow-up, 36 patients underwent liver transplantation, and 67 patients died. In the Palermo subgroup, 80% of mortality was related to chronic liver disease, including liver failure, HCC, spontaneous bacterial peritonitis, gastrointestinal bleeding, and hepatorenal syndrome; 10% was related to infections and 10% to extrahepatic diseases. In the Vienna subgroup, mortality was due to chronic liver disease 70%, pneumonia 4%, extrahepatic disease 10.5%, other cancer (HCC excluded) 9%, and other causes 6.5%. No statistically significant differences in survival were found between the 2 institutions (data not shown). Overall survival probability computed by the Kaplan-Meier method was 88% at 3 months, 84% at 6 months, and 77% at 12 months.
|Age (yr), mean (SD)||55.3 (9.8)|
|Male gender (%)||70|
|Alcoholic/nonalcohol cirrhosis (%)||50/50|
|Serum bilirubin (mg/dL), mean (SD)||2.6 (3.1)|
|Serum creatinine (mg/dL), mean (SD)||1.2 (0.9)|
|INR, mean (SD)||1.4 (0.4)|
|Serum albumin (g/L), mean (SD)||32.5 (6.8)|
|Serum sodium (mEq/L), mean (SD)||135.5 (5.9)|
|MELD score, mean (SD)||12.8 (5.7)|
The bivariate analysis by the Cox proportional hazard regression model showed 7 variables associated with mortality: age, serum bilirubin, serum creatinine, INR, serum albumin, serum sodium, and MELD (Table 2). Gender and cause of liver disease were not associated with survival (Table 2).
|HR||95% CI||P value|
|Gender (female vs. male)||1.05||0.72–1.51||0.807|
|Cause of cirrhosis (alcohol vs. nonalcohol)||1.20||0.86–1.68||0.295|
|Serum bilirubin (mg/dL)||1.28||1.04–1.58||0.021|
|Serum creatinine (mg/dL)||1.87||1.37–2.57||<0.0001|
|Serum albumin (mg/dL)||0.96||0.94–0.98||0.0004|
|Serum sodium (mEq/L)||0.95||0.92–0.97||0.0001|
The median MELD score was 12, and 92.5% of patients had a MELD score of <21. In bivariate analysis, increasing MELD score as continuous variable was associated with an increased risk of mortality (HR, 1.07; 95% CI, 1.04-1.09; P < 0.0001) (Table 2).
The median serum sodium was 136 mEq/L. In bivariate analysis, decreasing serum sodium as continuous variable was associated with an increased risk of mortality (HR, 0.95; 95% CI, 0.92-0.97; P = 0.0001) (Table 2). A serum sodium level of ≤130 mEq/L (hyponatremia) was noted in 56 patients (18%). Patients with hyponatremia differed significantly from those without hyponatremia with regard to lower albumin (29.5 ± 6.2 vs. 33.1 ± 6.8 g/dL, P = 0.0003), and higher MELD (14.5 ± 6.1 vs. 12.2 ± 5.4, P = 0.005). Twelve-month mortality was 17% in patients with normal sodium and 43% in patients with hyponatremia (P < 0.0001).
The median age was 55.3 years. Increasing age as a continuous variable was associated with increased mortality (HR, 1.02; 95% CI, 1.00-1.04; P = 0.027) (Table 2).
Multivariate analysis disclosed a model explaining time to death. This model contained MELD, serum sodium, and age (Table 3). Risk of mortality increased as MELD and age increased, and serum sodium decreased.
|HR||95% CI||P value|
|Serum sodium (mEq/L)||0.96||0.93–0.99||0.0029|
Integrated MELD Score
Cox proportional hazard regression model, weighting the coefficient of the MELD to a value of 1, assigned a coefficient of 0.3 per each year of patient's age and −0.7 to the value of serum sodium. We then added a constant of 100 to avoid negative numbers. Thus, for instance, a 50-year-old patient with serum sodium of 128 and an original MELD score of 14 has an iMELD score of 14 + (50 × 0.3) − (0.7 × 128) + 100 = 39.4.
Both the original MELD and integrated MELD results were statistically significantly associated with 12-month mortality by the logistic regression model. The auROC for iMELD was 13.4% greater than that for the original MELD (Fig. 1). Similarly, the iMELD model had a better likelihood ratio statistic than original MELD (48.2 vs. 23.5). Finally, the higher accuracy of the iMELD over the original score was confirmed also in predicting 3- and 6-month mortality (Table 4).
|Mortality||Model||auROC (95% CI)*|
|Training cohort||Validation cohort|
|3 months||Original MELD||0.70 (0.61–0.79)||0.79 (0.70–0.87)|
|iMELD||0.76 (0.68–0.83)||0.83 (0.74–0.90)|
|6 months||Original MELD||0.71 (0.62–0.79)||0.77 (0.69–0.83)|
|iMELD||0.79 (0.72–0.86)||0.82 (0.76–0.87)|
|12 months||Original MELD||0.69 (0.62–0.77)||0.75 (0.68–0.81)|
|iMELD||0.78 (0.72–0.84)||0.81 (0.76–0.86)|
In the training cohort, a comparison between auROC values of iMELD and original MELD showed that in patients with original MELD scores of <15 (n = 213), the auROC values were 0.58 (0.50-0.67; 95% CI) for MELD and 0.75 (0.65-0.84; 95% CI) for iMELD, whereas in patients with original MELD scores of ≥15 (n = 97), the auROC scores were 0.61 (0.47-0.63; 95% CI) for MELD and 0.67 (0.55-0.80; 95% CI) for iMELD.
Validation of the iMELD
The validation sample included 451 patients with cirrhosis on waiting list for liver transplantation. The mean (SD) age was 55.0 (7.8) years, and 73% of these patients were male. In this cohort of patients, mean (SD) value of serum bilirubin was 3.53 (5.08) mg/dL, serum creatinine 0.98 (0.68) mg/dL, INR 1.4 (0.44), and MELD score 11.8 (6.2); serum sodium was 137.5 (4.7) mEq/L. A serum sodium ≤130 mEq/L was noted in 29 patients (6%). During the first 12 months of follow-up, 88 patients died without receiving liver transplants, and 168 received liver transplants. Overall survival probability by Kaplan-Meier curve was 93%, 85%, and 76% at 3, 6, and 12 months, respectively.
In this independent sample, both the original MELD and iMELD scores were statistically significantly associated with time to death (Cox proportional hazard model) and 12-month mortality (logistic regression model). However, in both models, iMELD had a better predictive value, as showed by a much higher likelihood ratio statistic (82.0 vs. 41.4 in the Cox model; 81.3 vs. 50.5 in the logistic model). The auROC for iMELD was 8% greater than that for the original MELD (Fig. 2). In the validation cohort, the higher accuracy of the iMELD over the original score was confirmed in predicting 3- and 6-month mortality (Table 4).
The present study confirmed the prognostic value of the MELD score in predicting short-term mortality in patients with cirrhosis undergoing elective TIPS placement and in candidates for liver transplantation. In addition, this study clearly demonstrated that serum sodium and age have additional prognostic significance, independent of the MELD score in predicting mortality.
We found that the risk of mortality increased as serum sodium decreased. Serum sodium predicted better survival as continuous variable than as dichotomous variable, which allowed a more precise modeling. This has the additional advantage of reducing the effect of small variations of sodium level on the final score. The role of hyponatremia as independent predictor of mortality had already been emphasized by others17–19 and probably reflects the underlying hemodynamic derangement associated with advanced liver disease28 and the risk of hepatorenal syndrome.29
This is the first study demonstrating that age independently influences survival in addition to the MELD score. Age was associated to the risk of mortality as a continuous variable, with older patients having worse survival. The association of aging with mortality in cirrhosis is neither new nor surprising because several studies have shown this correlation in the past.30–32 Most recently, a systematic review of 118 prognostic studies in patients with cirrhosis showed that age is the most important independent prognostic factor of survival that is not part of the Child-Pugh score.33 It has been suggested that aging may reflect a longer duration of cirrhosis and a more severe liver disease. Because of the above, it is surprising that age was not assessed in some recent studies.
The new variables were incorporated into the original MELD by using an integrated formula (iMELD) including Cox proportional hazard regression coefficients of serum sodium and age and weighting the coefficient of the MELD to a value of 1. It should be remarked that, similarly to the components of the original MELD, new variables introduced in the MELD (serum sodium and age) are objective, reliable, reproducible, and readily available parameters.
The iMELD performed better than original MELD in predicting 3-, 6-, and 12-month mortality, as shown by a greater auROC over of original MELD. The auROC, which is a measure of discrimination (i.e., the ability of the model to distinguish patients who live from patients who die), increased by 8.6% at 3 months (from 0.70 to 0.76), by 11.3% at 6 months (from 0.71 to 0.79), and by 13.4% at 12 months (from 0.69 to 0.78). These improvements are of greater magnitude than that reported by Wiesner et al.6 in the study comparing MELD and Child-Pugh scores in predicting mortality on waiting list, in which MELD increased the accuracy over Child-Pugh by 7%. In addition, iMELD markedly increased the likelihood ratio statistic compared with original MELD (48.2 vs. 23.5). Therefore, our results support that adding serum sodium and age to the original MELD improves its accuracy in predicting time to death.
It should be noted that the iMELD confirmed to be more accurate than original MELD in predicting 3-, 6-, and 12-month mortality in an independent cohort of patients with cirrhosis listed for liver transplantation (validation sample).
It is of interest that although in the training sample patients who underwent TIPS placement had no HCC at the time of the procedure, to make the findings of this study more generalizable and to avoid a bias toward more advanced liver diseases, the validation sample included patients with HCC and used their calculated biological MELD score for the analysis. Interestingly, the iMELD remained a better predictor of survival compared with the original MELD, regardless of the analysis being performed, including patients with or without HCC.
The auROC for original MELD in our derivation cohort was lower than that observed by some investigators2, 3, 6, 18, 19 but similar to that reported by others.4, 5, 7, 17, 20 Heuman et al.17 showed that a MELD of <21 had a reduced discrimination power in detecting early mortality with a c statistic of 0.689 compared with 0.767 in the subgroup of patients with a MELD score >21 and suggested that the addition of sodium to the original MELD resulted in better mortality prediction at lower scores than at the upper ranges. In our population, 93% of patients had a MELD of <21, which could explain the relatively low auROC of original score. Of note, a similar prevalence of low MELD patients was reported in 12,996 adults listed in the United States for liver transplantation.34 Our data showed a higher accuracy of iMELD than original MELD; however, the predictive improvement of iMELD was greater in patients with original MELD scores of <15 than in patients with higher original score (23% vs. 9%). Finally, it should be noted that according to the auROC, the predictive ability of original MELD was lower in patients undergoing TIPS placement compared with patients on waiting list for liver transplantation (0.69 vs. 0.75, respectively) and that the predictive improvement by using iMELD was higher in the former group (+13.4%, vs. +8.0%, respectively). This may be because, at least in part, the training series of TIPS patients included many patients treated because of refractory ascites, which accounts for the higher prevalence of hyponatremia in the training vs. validation series (serum sodium ≤130: 18% vs. 6%), as well as for the lower level of serum sodium (mean 135.5 vs. 137.5 mEq/L). We suggest that differences in serum sodium or in hyponatremia may explain differences in the performance of the original MELD observed in other studies.2–8, 15, 17–20
Our study clearly showed that iMELD compared with the original score more accurately selects patients with higher risk of mortality after TIPS. This observation suggests that TIPS should be cautiously performed in patients with advanced liver disease and older age, and if possible, only as a bridge to liver transplantation. The iMELD also more accurately defined the risk of mortality in patients with cirrhosis on waiting list for liver transplantation, and therefore, it could be useful to define the priority of liver allocation. However, the inclusion of age into the iMELD might raise some ethical issues in this set of patients because conflicting data exist on the risk of postoperative death in elderly liver transplant recipients.35–40
In conclusion, this study shows that in patients with cirrhosis, serum sodium and age are predictors of early mortality, independent of the MELD score. Integration of these readily available objective variables into the original MELD formula greatly improves the predictive accuracy of time to death. The usefulness of iMELD in defining the priority of organ allocation in patients on the waiting list for liver transplants should be confirmed in larger series of patients.
- 5Comparison of MELD, Child-Pugh, and Emory model for the prediction of survival in patients undergoing transjugular intrahepatic portosystemic shunting. Am J Gastroenterol 2003; 98: 1167–1174., , , , , , et al.Direct Link:
- 22Practical Statistics for Medical Research. London, England: Chapman & Hall; 1994..
- 38Recent findings concerning liver transplantation in the United States. In: CeckaJM, TerasakiPI, eds. Clinical Transplants 1996. Los Angeles: UCLA Tissue Typing Laboratory; 1996.: 15., , .