- Email: [email protected]
A MELD-Based Model to Determine Risk of Mortality Among Patients With Acute Variceal Bleeding
Gastroenterology 2014;146:412–419
CLINICAL—LIVER A MELD-Based Model to Determine Risk of Mortality Among Patients With Acute Variceal Bleeding Enric Reverter,1,2 Puneeta Tandon,3 Salvador Augustin,4 Fanny Turon,1 Stefania Casu,1 Ravin Bastiampillai,3 Adam Keough,3 Elba Llop,1 Antonio González,4 Susana Seijo,1 Annalisa Berzigotti,1 Mang Ma,3 Joan Genescà,2,4 Jaume Bosch,1,2 Joan Carles García–Pagán,1,2 and Juan G. Abraldes1,2,3 1
CLINICAL LIVER
Hepatic Hemodynamic Laboratory, Liver Unit, Hospital Clínic, Institut d’Investigacions Biomèdiques August Pi i Sunyer, University of Barcelona; 2Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas, Barcelona, Spain; 3Division of Gastroenterology, Department of Medicine, University of Alberta, Edmonton, Canada; 4Liver Unit, Department of Internal Medicine, Hospital Universitari Vall d’Hebron, Vall d’Hebron Institut de Recerca, Universitat Autònoma de Barcelona, Barcelona, Spain This article has an accompanying continuing medical education activity on page e15. Learning Objective: Upon completion of this exam, successful learners will be able to formulate the standard course of action in a patient with cirrhosis and acute variceal bleeding.
See editorial on page 337. BACKGROUND & AIMS: Patients with cirrhosis with acute variceal bleeding (AVB) have high mortality rates (15%–20%). Previously described models are seldom used to determine prognoses of these patients, partially because they have not been validated externally and because they include subjective variables, such as bleeding during endoscopy and Child–Pugh score, which are evaluated inconsistently. We aimed to improve determination of risk for patients with AVB. METHODS: We analyzed data collected from 178 patients with cirrhosis (Child–Pugh scores of A, B, and C: 15%, 57%, and 28%, respectively) and esophageal AVB who received standard therapy from 2007 through 2010. We tested the performance (discrimination and calibration) of previously described models, including the model for end-stage liver disease (MELD), and developed a new MELD calibration to predict the mortality of patients within 6 weeks of presentation with AVB. MELD-based predictions were validated in cohorts of patients from Canada (n ¼ 240) and Spain (n ¼ 221). RESULTS: Among study subjects, the 6-week mortality rate was 16%. MELD was the best model in terms of discrimination; it was recalibrated to predict the 6-week mortality rate with logistic regression (logit, -5.312 þ 0.207 MELD; bootstrapped R2, 0.3295). MELD values of 19 or greater predicted 20% or greater mortality, whereas MELD scores less than 11 predicted less than 5% mortality. The model performed well for patients from Canada at all risk levels. In the Spanish validation set, in which all patients were treated with banding ligation, MELD predictions were accurate up to the 20% risk threshold. CONCLUSIONS: We developed a MELDbased model that accurately predicts mortality among patients with AVB, based on objective variables available at admission. This model could be useful to evaluate the efficacy of new therapies and stratify patients in randomized trials.
Keywords: Cirrhosis; Regression.
TIPS;
Prognostic
Model;
A
cute variceal bleeding (AVB) is a major complication of cirrhosis. Despite the improvement in prognosis in the past 3 decades,1–4 mortality during the bleeding episode remains high and ranges from 24% in unselected cirrhotic variceal bleeders5,6 to about 16% among those receiving the current standard of care (band ligation þ vasoactive drugs þ antibiotics).7 Several factors have been associated with a poor outcome in variceal bleeding. The most consistently reported risk indicators of death are Child–Pugh class or its components, model for end-stage liver disease (MELD) score, renal failure, bacterial infection at admission or shortly after, hypovolemic shock, active bleeding at endoscopy, hepatocellular carcinoma (HCC), and hepatic venous pressure gradient (HVPG) equal or greater than 20 mm Hg.5,7–11 Several prognostic models including these variables have been developed.5,7–9,12 These models could allow patient stratification according to the risk of death and could be used to guide therapeutic decisions, such as treating high-risk patients by a more aggressive approach. However, they are seldom used, in part because of the lack of external validation. Risk prediction in AVB has gained relevance since the demonstration that an early pre-emptive transjugular intrahepatic portosystemic shunt (TIPS) improves the outcome in high-risk patients in 2 randomized trials. In these trials high-risk patients were identified with hemodynamic criteria (HVPG 20 mm Hg shortly after admission)13 or with clinical criteria (Child C up to 13 points or Child B þ active bleeding at endoscopy despite vasoactive
Abbreviations used in this paper: AUROC, area under receiver operating characteristic curve; AVB, acute variceal bleeding; HCC, hepatocellular carcinoma; MELD, model for end-stage liver disease; HVPG, hepatic venous pressure gradient; ROC, receiver operating characteristic; TIPS, transjugular intrahepatic portosystemic shunt.
Logistic © 2014 by the AGA Institute 0016-5085/$36.00 http://dx.doi.org/10.1053/j.gastro.2013.10.018
February 2014
Materials and Methods Study Cohort and Data Collection We prospectively collected a database including all patients admitted to the Hospital Clinic (Barcelona, Spain) for portal hypertension–related bleeding (all sources and all etiologies) from October 2007 to December 2010. This was performed as part of a project to evaluate the impact of current therapies on the prognosis of portal hypertension–related bleeding and was approved by the Institutional Review Board of the Hospital Clinic. All data were gathered in the context of standard practice from the clinical records of the patients, and were anonymized and collected in a protected database. No specific procedures were conducted for the study and informed consent was not required. Patients with cirrhosis and acute bleeding from esophageal varices (all-comers) were considered eligible for the study. The diagnosis of cirrhosis was based on previous clinical history, liver biopsy, and/or unequivocal clinical data and compatible findings on imaging techniques. Esophageal variceal bleeding was confirmed by emergency endoscopy according to Baveno criteria.17 Exclusion criteria were HCC stage D according to the Barcelona Clinic liver cancer staging system.18 This exclusion criterion was based on the high mortality rate of these patients (>60%) and the high rate of therapeutic abstention. Baseline clinical, biochemical, and imaging data of patients were recorded prospectively at admission by a member of the Hepatic Hemodynamic Lab (E.R., E.L., S.S., or J.G.A.). Liver and renal function tests at admission and endoscopic and vasoactive treatments received were available for all patients, as well as rescue therapies and transfusion requirements during the episode. The record of each patient was updated at day 42 or death.
Therapeutic Interventions and Definitions In the period considered for this study, before the implementation of an “early TIPS” policy in our unit,14 patients were treated with the standard of care according to Baveno consensus workshops and American Association for the Study of Liver Diseases guidelines17,19: vasoactive drugs from admission (somatostatin or terlipressin), endoscopic band ligation (sclerotherapy if technically difficult or not feasible) within 12
413
hours of admission, and antibiotic prophylaxis (norfloxacin or ceftriaxone when indicated20). Rescue therapies (ie, Sengstaken–Blakemore balloon tamponade and TIPS) were applied when necessary (re-bleeding or failure to control bleeding). At day 5, combined secondary prophylaxis (b-blockers plus variceal ligation in 2- to 3-week intervals) was started when applicable. Time of admission to hospital was considered as time zero for the follow-up evaluation. The primary end point analyzed in this study was 6-week mortality (bleeding-related mortality).17
Prognostic Models We selected prognostic models offering continuous quantitative estimations of 6-week mortality in AVB. Although MELD was not developed specifically to predict mortality in the bleeding episode,21,22 we also evaluated the performance of this model (as modified by the United Network for Organ Sharing)22 because it has shown excellent predictions in patients with end-stage liver disease,23 and specifically in AVB.12,24,25 Together with MELD, D’Amico9 and Augustin7 models (described in detail in Supplementary Table 1) were selected for validation. In the study by Augustin,7 2 models were presented, a linear one (developed with logistic regression) and a nonlinear one (developed with classification and regression tree analysis). Only the linear model was considered because classification and regression tree analysis models do not provide continuous predictions. Finally, we tested Child–Pugh performance because it repeatedly has been reported as a good prognostic indicator in all clinical situations in cirrhosis, including the AVB episode.5–9,24,25
Statistical Analysis Models performance. To assess the performance of the prognostic models in predicting 6-week mortality, the discrimination, calibration, and overall performance of each model were studied. Discrimination refers to the ability to rank patients according to their risk of developing the outcome, whereas calibration refers to the ability to predict absolute risks (how closely the predicted probabilities agree with the actual outcomes).26 Calibration is of special interest because it allows defining decision thresholds according to the absolute risk of the outcome. The discriminative ability was assessed by receiver operating characteristic (ROC) curve analysis. Calibration was tested by plotting predicted and observed mortalities, and by the Hosmer–Lemeshow goodness-of-fit test after splitting the sample into quintiles.26 In the case of the Augustin7 and D’Amico9 models, absolute predicted mortalities were calculated according to the formulas presented in the original reports. In the case of MELD, the original publication did not report the temporal constant to specifically calculate 6-week mortality.21 After contacting the original authors, they provided a temporal constant of 0.827 for calculating 6-week mortality as described in the original report.21 Child–Pugh calibration could not be assessed because the data are not available to estimate the absolute probability of 6week mortality for each Child–Pugh score. The overall performance of the models was evaluated with R2 and the Brier score.27 A higher R2 and a lower Brier score indicate better performance. Update of MELD calibration. MELD was the best model in terms of discrimination and overall performance, but was significantly miscalibrated. Therefore, we decided to use the Hospital Clinic cohort (initially the validation cohort) to update MELD calibration by fitting a logistic regression model
CLINICAL LIVER
drug treatment).14,15 Both definitions have relevant problems: HVPG measurements are not widely available and may be difficult to perform in bleeding patients, and the definition of active bleeding and some components of the Child classification have subjectivity. Indeed, in a recent trial it was shown that both Child and active bleeding were evaluated inconsistently in different centers.16 Thus, better predictive models, well validated, with objective variables easily and early available from admission clearly are needed. The overall aim of this study was to improve risk prediction in AVB. For this purpose we first tested the performance of recently described prognostic models for AVB in a contemporary series of patients. The MELD model was found to be the best model in terms of discrimination, but it was significantly miscalibrated. Therefore, a second aim of the present study was to update MELD calibration to predict mortality from AVB. The third aim was to validate this new MELD calibration in 2 external series of patients with AVB.
Meld and Variceal Bleeding
414
Reverter et al
Gastroenterology Vol. 146, No. 2
introducing MELD as the only covariable. This procedure (“logistic calibration”) does not modify the discrimination of the model.28 We used penalized maximum likelihood estimation of coefficients to correct for overoptimism because this leads to more accurate predictions in external samples.28,29 The performance of this MELD-based model was validated internally with bootstrapping (200 random bootstrap samples of 178 patients drawn with replacement from the original sample).28
Data-driven model restricted to objective variables. To explore the best data-driven model to predict 6-week mortality in our cohort, we conducted a multivariable logistic regression analysis, restricted to objective variables available at admission.
External validation of MELD recalibration. Because the MELD model was updated, it required CLINICAL LIVER
further external validation. This was performed in 2 series of patients with esophageal variceal bleeding. The first sample included 240 consecutive patients with AVB admitted to 2 providing hospitals from the University of Alberta (Edmonton, Canada) between 2003 and 2008. These patients were managed according to the international guidelines (drugs þ endoscopy þ antibiotics, with TIPS as rescue therapy), and are described in detail in Supplementary Table 2. Adherence to antibiotic prophylaxis was 71%. Second, taking into account that patients in whom band ligation is applied effectively seem to have an excellent prognosis,7 we further explored MELD calibration in 221 patients from the Vall D’Hebron Hospital (Autonomous University of Barcelona, Spain), all treated with drugs þ band ligation þ antibiotics between 2002 and 2012 (Supplementary Table 2). This series was partially reported in a previous study.7 Discrimination and calibration of the MELD-based model in these validation sets were assessed as reported earlier. Data to calculate MELD was available for all patients in both series. Analysis was performed with the IBM SPSS Statistics 19.0 package (SPSS, Inc, Chicago, IL) and R30 (with the aid of the rms and predictABEL packages), by means of auditable and rerunnable script files.
Results Study Cohort, Treatment, and Outcomes A total of 303 patients admitted to the Hospital Clinic for portal hypertension–related bleeding were registered prospectively between October 2007 and December 2010. Among these, 198 patients had cirrhosis and esophageal variceal bleeding. Nineteen patients with advanced HCC (Barcelona Clinic liver cancer stage D) and 1 patient lost to follow-up evaluation were excluded. The final analysis included 178 patients (Figure 1). Baseline characteristics of the patients at admission are reported in Table 1. The overall 6-week mortality was 16%. Causes of death were uncontrolled bleeding in 9 patients (31%), liver failure in 8 patients (28%), and sepsis or multiorgan failure in 12 patients (41%). Supplementary Figure 1 shows the temporal distribution of deaths.
Performance of Prognostic Models for 6-Week Mortality Discrimination was studied by analyzing the ROC curves. MELD showed the highest discrimination (Figure 2),
Figure 1. Flowchart of patients admitted to our hospital for portal hypertension–related bleeding and selection of the study cohort. EV, esophageal variceal.
although the difference between MELD discrimination and that of the other models was not statistically significant. Calibration (ie, the agreement between predicted and observed death rates) is shown in Figure 3. All 3 models were miscalibrated significantly. The D’Amico9 model and MELD systematically overestimated the observed mortality. The Augustin7 model both overestimated and underestimated mortality at different levels of risk. MELD was
Table 1.Baseline Characteristics of Patients at Admission Age, y Male sex, n (%) Etiology of alcohol/HCV/virus þ alcohol/others, % Previous episodes of variceal bleeding, n (%) Albumin level, g/L Bilirubin level, mg/dL, median (IQR) Creatinine level, mg/dL Na level, mEq/L Leukocytes, 109/L INR Ascites present, n (%) Encephalopathy, n (%) MELD score, median (IQR) MELD-Na score, median (IQR) Child–Pugh class of A/B/C, % Child–Pugh score Active bleeding at endoscopy, n (%) Systolic arterial pressure < 100 mm Hg (first 3 h), n (%) Infection at admission, n (%) HCC, n (%) Endoscopic band ligation/sclerotherapy, % Somatostatin/terlipressin, % Need for balloon tamponade, n (%) Rescue TIPS, n (%) 6-Week mortality, n (%)
58.3 (12.4) 130 (73) 39/32/11/18 72 (40) 28.4 (4.8) 1.8 (2.1) 1.1 (0.7) 136.3 (6.2) 8.5 (4.4) 1.6 (0.53) 112 (63) 43 (24) 13.9 (7.5) 17.5 (8.9) 15/57/28 8.4 (2.0) 60 (34) 61 (34) 28 (16) 18 (10) 65/21 80/20 27 (15) 28 (16) 29 (16)
NOTE. Data are expressed as means (SD) unless otherwise stated. HCV, hepatitis C virus; INR, international normalized ratio; IQR, interquartile range.
Meld and Variceal Bleeding
415
Therefore, we decided to update MELD calibration as described in the Materials and Methods section. The equation of the updated model was as follows: logit ¼ –5.312 þ 0.207 $ MELD (nonbootstrapped R2, 0.341; bootstrapped R2, 0.330; Supplementary Table 3). Table 2 provides predicted mortalities for individual MELD values and Supplementary Figure 2 shows the prediction curve of the model. A MELD score of 11 was associated with a 5% risk of mortality, and could be a potentially useful prognostic threshold for defining low-risk patients. A MELD score of 19 was associated with a mortality rate of 20%, and could be a potentially useful threshold for defining a high risk of death. The internal calibration of the MELD-based model was excellent (Figure 4A). Variables reflecting the severity of bleeding, including a systolic arterial pressure less than 100 mm Hg within the first 3 hours from admission and active bleeding at endoscopy, did not significantly add to the predictive value of the MELDbased model (P ¼ .25 and P ¼ .55, respectively).
Figure 2. ROC curves of the models for predicting 6-week mortality in acute variceal bleeding. MELD had the greatest AUROC, indicating the best discrimination. Differences between MELD’s AUROCs and those from other models, however, were not statistically significant (P ¼ .0574 vs Augustin7 model; P ¼ .1194 vs D’Amico9 model; P ¼ .2179 vs Child–Pugh).
the model showing the (Supplementary Table 3).
best
overall
performance
Early Objective Prognostic Indicators for 6-Week Mortality We performed a multivariable analysis with objective variables available on admission to develop the best datadriven model (Supplementary Table 4). Only creatinine, bilirubin, and international normalized ratio (the components of MELD) remained in the final model (area under the ROC curve [AUROC], 0.778; nonbootstrapped R2, 0.314; bootstrapped R2, 0.268). Even this new model was fine-tuned to the data, it was not better than the MELD-based model.
Model Update: MELD Recalibration
External Validation of the MELD-Based Model
MELD was the best model in terms of discrimination and overall performance, and had the additional advantage over the other models of including only objective variables.
Finally, we tested the performance of MELD in 2 additional series of patients with AVB (Supplementary Table 2). The MELD-based model showed excellent discrimination
Figure 3. Calibration plots for (A) MELD, (B) D’Amico,9 and (C) Augustin7 models. To construct these plots, the sample was split into quintiles, and predicted mortality was plotted against observed mortality. Points below the diagonal line (ie, identity line, perfect prediction) indicate overestimation of mortality for that group. Points above the diagonal line indicate underestimation of mortality. P values are those of the Hosmer–Lemeshow goodness-of-fit test (the lower the P value, the worse the agreement between the observed and predicted mortalities). (A) MELD and (B) the D’Amico9 model overestimated mortality. (C) The Augustin7 model both overestimated and underestimated mortality.
CLINICAL LIVER
February 2014
416
Reverter et al
Gastroenterology Vol. 146, No. 2
Table 2.Predicted Mortality for Each MELD Value According to the Updated MELD-Based Model MELD value
CLINICAL LIVER
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Predicted mortality, % 2.5 3.1 3.8 4.6 5.6 6.8 8.2 9.9 11.9 14.3 17.0 20.1 23.6 27.6 31.9 36.6 41.5 46.6 51.7 56.9 61.9 66.6 71.1
(AUROC, 0.873) and calibration (Figure 4B) in the cohort from the University of Alberta (Edmonton, Canada). In the cohort from Vall d’Hebron Hospital (Barcelona, Spain) including patients selected on the basis of receiving band ligation (þ antibiotics þ drugs), the MELD-based model also had good discrimination (AUROC, 0.794). Calibration was acceptable, except in the upper quintile of the patients, in which our model predicted a higher mortality than that actually observed (Figure 4C). In all 3 series, MELD scores potentially defining arbitrary thresholds of low and high risk of death (11 and of 19) consistently were associated with a mortality rate of approximately 5% and 20%, respectively (Figure 4D–F). Therefore, even though the MELD-based model overpredicted mortality for high MELD values in the second validation set, it still would be a useful tool to accurately identify patients with a risk of death of 20% or higher.
Discussion The present study provides new information to improve risk prediction in patients with cirrhosis and acute variceal bleeding. Specifically, we show that MELD outperforms other models designed to predict prognosis in AVB. Furthermore, we propose a model that assigns a probability of 6-week mortality for each MELD value, and we provide external validation for our MELD calibration in 2 additional series of patients. The recent demonstration that a stratified management of patients with AVB can improve survival in high-risk patients13,14 has emphasized the importance of having
accurate prognostic tools in AVB. However, the criteria used in those trials to define high risk have a low clinical accessibility (HVPG measurement within first 24 h) or include subjective variables (Child–Pugh, active bleeding). In addition, it has been questioned recently whether all patients defined as high risk in the inclusion criteria of the EarlyTIPS trial are indeed high risk, especially those at Child B class þ active bleeding. Some series have shown a low mortality rate in this subgroup of patients7 and others have shown significant heterogeneity in scoring the patients.16 Thus, there is a clear need for better prediction models in AVB, preferably with easily obtained, objective, and reproducible variables available early after patient admission. One of the problems in the use of prognostic models in AVB is the lack of external validation. Thus, our first goal was to evaluate the performance of previously described models. For this purpose we chose 2 models specifically developed for AVB,7,9 which also offered continuous predictions of 6-week mortality. We additionally evaluated MELD because it has shown prognostic value in almost every clinical situation in cirrhosis,23 including variceal bleeding,12,24,25 and Child–Pugh score because it is one of the most widely used bedside prognostic tools in cirrhosis. Our findings show that MELD was the best model in terms of discrimination. However, calibration was poor for all models and good calibration (agreement between expected and observed mortality) is essential to derive decision thresholds to guide therapeutic decisions. This lack of calibration could be explained by several reasons. MELD was developed in a different setting (patients treated with TIPS) in a series of patients from more than a decade ago, before the implementation of several effective treatments for patients with cirrhosis and, therefore, miscalibration could be expected. On the other hand, in the D’Amico9 study many patients did not receive what is currently accepted as the standard of care. Both the D’Amico9 and Augustin7 models included the presence of HCC as a prognostic variable, but without differentiating between HCC stages. Although the presence of HCC up to stage C has little short-term prognostic significance, stage D patients have exceedingly high short-term mortality, and many times these patients are managed with symptomatic measures if presenting with AVB. In the present study, rather than trying to develop a completely new prognostic model, we used a strategy of model updating, which is a better way to develop stable and generalizable models.31 For this purpose, we focused on MELD for several reasons. Because MELD had a good discriminative value it needed less extensive updating. We could recalibrate the model easily to make accurate predictions with logistic calibration, which does not change the structure and the discrimination of the model. The other models would require much more extensive updating, and the sample size of the present study would be inadequate. In addition, MELD is based in simple and reliable measurements, which are important criteria in developing clinically useful prognostic models,31 and it is widely implemented worldwide for the prediction of outcomes in cirrhosis. Of note, variables reflecting the severity of bleeding, which have
Meld and Variceal Bleeding
417
CLINICAL LIVER
February 2014
Figure 4. (A–C) Calibration of the MELD-based model in the 3 samples. The updated MELD-based model showed an excellent calibration in the Hospital Clinic sample (internal calibration of the model), but also in the University of Alberta sample (external calibration). In the Vall D’Hebron sample it showed a good calibration for low values of MELD, but overpredicted mortality in the upper quintile of MELD (P values are those of the Hosmer–Lemeshow goodness-of-fit test). (D–F) The relationship between MELD and mortality in the 3 series of patients. These plots were constructed with nonparametric regression, which shows smoothed actual mortality rates. Tick marks are drawn at actual MELD values. The upper and lower 5% of cases were trimmed. In the 3 series a MELD score of 11 was associated with a mortality rate of approximately 5%, and a MELD score of 19 was associated with a mortality rate of approximately 20%. The Vall D’Hebron series showed a lower mortality rate for MELD scores greater than 20, as compared with the other 2 series.
the additional problem of lacking an unequivocal definition, did not add predictive value to the MELD-based model. Because model updating necessarily requires additional validation, we searched for 2 external validation samples. The first validation set came from a comparable setting: tertiary university hospitals with wide experience in the management of AVB applying the current standard of care,17,19 but in a different country. We showed that the performance of the MELD-based model was excellent in the Alberta series, with an excellent agreement between predicted and observed 6-week mortalities along the whole spectrum of disease severity. Indeed, arbitrary thresholds potentially defining low risk and high risk (MELD scores, 11 and 19, respectively) showed an almost perfect fit to observed 5% and 20% mortality rates (Figure 4E). With the second validation set, from a different University Hospital in Barcelona, we aimed to evaluate the performance of the MELD-based model in a selected sample of patients in whom the currently considered optimal therapy (banding ligation þ vasoactive drugs þ antibiotics) was applicable. These
patients have been shown to have a better overall prognosis than all-comers.6,7 In this sample, MELD-based predictions overestimated mortality only in the most advanced patients. Still, MELD values of 11 and 19 again showed an excellent fit to the observed 5% and 20% mortality rates (Figure 4F), suggesting that the model would be equally useful for risk stratification in this selected category of patients. Our results might set the framework to evaluate the impact of new therapies (including early TIPS) in real-practice outcomes, and to identify the patients who might benefit most from specific treatments. Indeed, the robustness of the model might serve to compare results obtained in different centers, at different periods, and with different treatment protocols. However, it is important to note that these results might not have immediate clinical applicability because the only patienttailored strategy proved to date to improve the outcome in AVB has been early TIPS, and patients were selected on the basis of Child and active bleeding.14,15 In summary, in patients with cirrhosis and AVB, MELD offered an objective and accurate prognostic prediction with
418
Reverter et al
variables available early after admission. MELD could be more efficient than the current criteria for selecting highrisk patients who might benefit from more aggressive treatments. Our proposed MELD-based predictions might be useful in refining the use of early TIPS, in evaluating the impact of new therapeutic strategies on patient prognosis, and in improving risk stratification in future clinical trials.
Supplementary Material Note: To access the supplementary material accompanying this article, visit the online version of Gastroenterology at www.gastrojournal.org, and at http://dx.doi.org/10.1053/j. gastro.2013.10.018. CLINICAL LIVER
References 1. Graham D, Smith J. The course of patients after variceal hemorrhage. Gastroenterology 1981;80:800–806. 2. Chalasani N, Kahi C, Francois F, et al. Improved patient survival after acute variceal bleeding: a multicenter, cohort study. Am J Gastroenterol 2003;98:653–659. 3. Carbonell N, Pauwels A, Serfaty L, et al. Improved survival after variceal bleeding in patients with cirrhosis over the past two decades. Hepatology 2004;40:652–659. 4. Stokkeland K, Brandt L, Ekbom A, et al. Improved prognosis for patients hospitalized with esophageal varices in Sweden 1969-2002. Hepatology 2006; 43:500–505. 5. Augustin S, Muntaner L, Altamirano JT, et al. Predicting early mortality after acute variceal hemorrhage based on classification and regression tree analysis. Clin Gastroenterol Hepatol 2009;7:1347–1354. 6. Amitrano L, Guardascione MA, Manguso F, et al. The effectiveness of current acute variceal bleed treatments in unselected cirrhotic patients: refining short-term prognosis and risk factors. Am J Gastroenterol 2012; 107:1872–1878. 7. Augustin S, Altamirano J, Gonzalez A, et al. Effectiveness of combined pharmacologic and ligation therapy in highrisk patients with acute esophageal variceal bleeding. Am J Gastroenterol 2011;106:1787–1795. 8. Abraldes JG, Villanueva C, Banares R, et al. Hepatic venous pressure gradient and prognosis in patients with acute variceal bleeding treated with pharmacologic and endoscopic therapy. J Hepatol 2008;48:229–236. 9. D’Amico G, de Franchis R. Upper digestive bleeding in cirrhosis. Post-therapeutic outcome and prognostic indicators. Hepatology 2003;38:599–612. 10. Moitinho E, Escorsell A, Bandi JC, et al. Prognostic value of early measurements of portal pressure in acute variceal bleeding. Gastroenterology 1999;117:626–631. 11. Moitinho E, Planas R, Bañares R, et al. Multicenter randomized controlled trial comparing different schedules of somatostatin in the treatment of acute variceal bleeding. J Hepatol 2001;35:712–718. 12. Bambha K, Kim WR, Pedersen R, et al. Predictors of early re-bleeding and mortality after acute variceal haemorrhage in patients with cirrhosis. Gut 2008; 57:814–820.
Gastroenterology Vol. 146, No. 2 13. Monescillo A, Martinez-Lagares F, Ruiz-del-Arbol L, et al. Influence of portal hypertension and its early decompression by TIPS placement on the outcome of variceal bleeding. Hepatology 2004;40:793–801. 14. Garcia-Pagan JC, Caca K, Bureau C, et al. Early use of TIPS in patients with cirrhosis and variceal bleeding. N Engl J Med 2010;362:2370–2379. 15. Garcia-Pagan JC, Di Pascoli M, Caca K, et al. Use of early-TIPS for high-risk variceal bleeding: results of a post-RCT surveillance study. J Hepatol 2013;58:45–50. 16. Bosch J, Thabut D, Albillos A, et al. Recombinant factor VIIa for variceal bleeding in patients with advanced cirrhosis: a randomized, controlled trial. Hepatology 2008;47:1604–1614. 17. de Franchis R. Revising consensus in portal hypertension: report of the Baveno V consensus workshop on methodology of diagnosis and therapy in portal hypertension. J Hepatol 2010;53:762–768. 18. Bruix J, Sherman M. Management of hepatocellular carcinoma: an update. Hepatology 2011;53:1020–1022. 19. Garcia-Tsao G, Sanyal AJ, Grace ND, et al. Prevention and management of gastroesophageal varices and variceal hemorrhage in cirrhosis. Hepatology 2007; 46:922–938. 20. Fernandez J, Ruiz dA, Gomez C, et al. Norfloxacin vs ceftriaxone in the prophylaxis of infections in patients with advanced cirrhosis and hemorrhage. Gastroenterology 2006;131:1049–1056. 21. Malinchoc M, Kamath PS, Gordon FD, et al. A model to predict poor survival in patients undergoing transjugular intrahepatic portosystemic shunts. Hepatology 2000; 31:864–871. 22. Available from: www.mayoclinic.org/meld/mayomodel6. html. 23. Kamath PS, Kim WR. The model for end-stage liver disease (MELD). Hepatology 2007;45:797–805. 24. Chalasani N, Kahi C, Francois F, et al. Model for endstage liver disease (MELD) for predicting mortality in patients with acute variceal bleeding. Hepatology 2002; 35:1282–1284. 25. Amitrano L, Guardascione MA, Bennato R, et al. MELD score and hepatocellular carcinoma identify patients at different risk of short-term mortality among cirrhotics bleeding from esophageal varices. J Hepatol 2005; 42:820–825. 26. Altman DG, Vergouwe Y, Royston P, et al. Prognosis and prognostic research: validating a prognostic model. BMJ 2009;338:b605. 27. Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010; 21:128–138. 28. Steyerberg EW. Clinical prediction models. A practical approach to development, validation and updating. New York: Springer ScienceþBusiness Media, LLC, 2009. 29. Moons KG, Donders AR, Steyerberg EW, et al. Penalized maximum likelihood estimation to directly adjust diagnostic and prognostic prediction models for overoptimism: a clinical example. J Clin Epidemiol 2004; 57:1262–1270.
February 2014 30. Available from: www.r-project.org. 31. Moons KG, Altman DG, Vergouwe Y, et al. Prognosis and prognostic research: application and impact of prognostic models in clinical practice. BMJ 2009;338:b606.
Meld and Variceal Bleeding
419
Acknowledgments The authors thank Laura Rocabert, Ángels Baringo, and Rosa Sáez for expert nursing support; and Clara Esteva, Rachel Borowski, and Jackie Love for secretarial assistance. Conflicts of interest The authors disclose no conflicts.
Author names in bold designate shared co-first authorship.
Reprint requests Address requests for reprints to: Juan G. Abraldes, MD, Division of Gastroenterology, 1-51 Zeidler-Ledcor Centre, University of Alberta Campus, Edmonton, Alberta T6G 2X8, Canada. e-mail: [email protected]; fax: (780) 248-1895.
Funding This study was supported in part by grants from the Instituto de Salud Carlos III, Ministerio de Economía y Competitividad (PI11/00883 to J.G.A., PS09/01261 to J.B., and Fondo de Investigación Sanitaria (FIS): PI12/01759 to J.G.); co-financed by Fondo Europeo de Desarrollo Regional (FEDER) funds (EU, “Una manera de hacer Europa”); Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas (CIBERehd) is funded by the Instituto de Salud Carlos III; and a Río Hortega award, Instituto de Salud Carlos III (E.R.).
CLINICAL LIVER
Received May 3, 2013. Accepted October 16, 2013.
419.e1
Reverter et al
Gastroenterology Vol. 146, No. 2
0.10 0.05 0.00
Fraction of deaths
0.15
Temporal distribution of bleeding related deaths
0
10
20
30
40
Days after index bleeding
Probability of death at 6 weeks
Supplementary Figure 1. Temporal distribution of the bleeding-related mortality (42 days). Most deaths occurred within the first 7 days of index bleeding.
0.8
0.6
0.4
0.2
10
20
30
40
MELD
Supplementary Figure 2. Prediction curve of the updated MELD-based model. The grey area represents the amplitude of 95% confidence intervals. The formula of the regression model was as follows: logit ¼ – 5.312 þ 0.207 * MELD. The probability of 6-week mortality was estimated as follows: p ¼ 1 / (1 þ e-logit). The bootstrapped R2 of the model was 0.3295.
February 2014
Meld and Variceal Bleeding
419.e2
Supplementary Table 1.Prognostic Models Evaluated Model
Equation
D’Amico and de Franchis9 (Hepatology 2003) Augustin et al7 (American Journal of Gastroenterology 2011) MELD
Logit ¼ -0.083 þ 0.88 $ encephalopathy (1 ¼ no; 2 ¼ mild/moderate; 3 ¼ severe) þ 0.22 $ bilirubin (mg/dL) þ 1.12 $ HCC (0 ¼ no; 1 ¼ yes) - 0.86 $ albumin (g/dL) Logit ¼ -7.226 þ 0.475 $ Child–Pugh score (5–15 points) þ 1.746 $ HCC (0 ¼ no; 1 ¼ yes) þ 1.179 $ creatinine 1.0 mg/dL (0 ¼ no; 1 ¼ yes) 6.43 þ 11.2 $ ln (INR) þ 9.57 $ ln (creatinine; mg/dL) þ 3.78 $ ln (bilirubin; mg/dL)
NOTE. We studied the performance of MELD score and previously described models for 6-week mortality by D’Amico9 and Augustin.7 MELD score was calculated according to the United Network for Organ Sharing modification22: creatinine and bilirubin were rounded to 1 if below and creatinine was capped to 4 if above or in patients with renal replacement therapy. D’Amico9 and Augustin7 models were calculated with the equations outlined in the table. HCC: hepatocellular carcinoma; INR, international normalized ratio.
Supplementary Table 2.Baseline Characteristics of Patients From Vall d’Hebron Hospital and Edmonton Hospital (Validation Sets)
Age, y Sex, male/female, % Etiology, OH/HCV/virus þ OH/others, (%) Albumin level, g/L Bilirubin level, mg/dL, median (IQR) Creatinine level, mg/dL INR MELD score, median (IQR) Child–Pugh class A/B/C, % Child–Pugh score HCC, n (%) Banding/sclerotherapy, % Balloon tamponade, n (%) Rescue TIPS, n (%) 6-Week mortality, n (%)
Vall d’Hebron (n ¼ 221)
University of Alberta (n ¼ 240)
60.4 (14.4) 67/33 37/36/12/15 27.1 (4.7) 1.6 (1.4) 1.1 (0.6) 1.8 (0.65) 14.9 (7.0) 25/49/26 8.3 (2.2) 23 (10.4) 100/0 25 (11.3) 4 (1.8) 27 (12.2)
56 (12.3) 75/25 48/13/16/23 28 (7.5) 2.0 (2.5) 1.2 (0.7) 1.6 (0.52) 14.3 (8.6) 14/51/35 8.8 (2.1) 19 (7.9) 92.9/3.3 13 (5.4) 16 (6.7) 38 (15.8)
NOTE. Both cohorts were similar to the training set, except for a slightly lower 6-week mortality rate in the Vall d’Hebron cohort. Data are expressed as means (SD) unless otherwise stated. HCV, hepatitis C virus; INR, international normalized ratio; IQR, interquartile range.
419.e3
Reverter et al
Gastroenterology Vol. 146, No. 2
Supplementary Table 3.Performance of the Evaluated Models and of the Updated MELD in the Hospital Clinic Barcelona Cohort
MELD Augustin7 model D’Amico9 model Child–Pugh MELD with updated calibration
R2
Brier score
0.325 0.123 0.161 0.341
0.106 0.133 0.171 0.097
AUROC (95% confidence interval) 0.795 0.723 0.703 0.740 0.795
Hosmer–Lemeshow test (P value)
(0.689–0.901) (0.621–0.824) (0.582–0.824) (0.639–0.841) (0.689–0.901)
.012 .006 .000 .647
NOTE. R2 and Brier score provide a global evaluation of the model (discrimination þ calibration): a higher R2 and a lower Brier score indicate better performance. AUROC evaluates the discrimination of the model; higher values indicate better discrimination. The P value of the Hosmer–Lemeshow test evaluate the calibration of the model; a lower P value indicates worse calibration.
Supplementary Table 4.Univariable and Multivariable Analysis of Objective Variables Associated With 6-Week Mortality Variable Leukocytes, 109/L Na, mEq/L Creatinine level, mg/dL ASAT level, IU/L Albumin level, g/L Bilirubin level, mg/dL, median (IQR) INR Age, y Hepatocellular carcinoma, n (%) Portal vein thrombosis, n (%) Alcoholic etiology, n (%)
Survivors (n ¼ 149) 8.4 136.5 0.98 89 28.6 1.7 1.54 57.8 18 14 58
(4.5) (5.6) (0.6) (141) (4.5) (1.8) (0.4) (12.5) (12.1) (9.4) (38.9)
Dead (n ¼ 29) 9.1 135 1.6 144 26.8 3.1 1.95 60.7 1 2 11
(3.4) (8.6) (0.7) (168) (6.8) (5.4) (0.8) (11.8) (3.4) (6.8) (37.9)
Univariable, P value .419 .389 <.001 .063 .083 <.001 <.001 .248 .168 .668 .920
Multivariable, P value
<.001
.027 .014
NOTE. Data are expressed as means (SD) unless otherwise stated. Creatinine, bilirubin, and INR were log-transformed for the analysis. At multivariable analysis only creatinine, INR, and bilirubin (already included in MELD) were associated with 6-week mortality. INR, international normalized ratio; IQR, interquartile range.
CLINICAL—LIVER A MELD-Based Model to Determine Risk of Mortality Among Patients With Acute Variceal Bleeding Enric Reverter,1,2 Puneeta Tandon,3 Salvador Augustin,4 Fanny Turon,1 Stefania Casu,1 Ravin Bastiampillai,3 Adam Keough,3 Elba Llop,1 Antonio González,4 Susana Seijo,1 Annalisa Berzigotti,1 Mang Ma,3 Joan Genescà,2,4 Jaume Bosch,1,2 Joan Carles García–Pagán,1,2 and Juan G. Abraldes1,2,3 1
CLINICAL LIVER
Hepatic Hemodynamic Laboratory, Liver Unit, Hospital Clínic, Institut d’Investigacions Biomèdiques August Pi i Sunyer, University of Barcelona; 2Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas, Barcelona, Spain; 3Division of Gastroenterology, Department of Medicine, University of Alberta, Edmonton, Canada; 4Liver Unit, Department of Internal Medicine, Hospital Universitari Vall d’Hebron, Vall d’Hebron Institut de Recerca, Universitat Autònoma de Barcelona, Barcelona, Spain This article has an accompanying continuing medical education activity on page e15. Learning Objective: Upon completion of this exam, successful learners will be able to formulate the standard course of action in a patient with cirrhosis and acute variceal bleeding.
See editorial on page 337. BACKGROUND & AIMS: Patients with cirrhosis with acute variceal bleeding (AVB) have high mortality rates (15%–20%). Previously described models are seldom used to determine prognoses of these patients, partially because they have not been validated externally and because they include subjective variables, such as bleeding during endoscopy and Child–Pugh score, which are evaluated inconsistently. We aimed to improve determination of risk for patients with AVB. METHODS: We analyzed data collected from 178 patients with cirrhosis (Child–Pugh scores of A, B, and C: 15%, 57%, and 28%, respectively) and esophageal AVB who received standard therapy from 2007 through 2010. We tested the performance (discrimination and calibration) of previously described models, including the model for end-stage liver disease (MELD), and developed a new MELD calibration to predict the mortality of patients within 6 weeks of presentation with AVB. MELD-based predictions were validated in cohorts of patients from Canada (n ¼ 240) and Spain (n ¼ 221). RESULTS: Among study subjects, the 6-week mortality rate was 16%. MELD was the best model in terms of discrimination; it was recalibrated to predict the 6-week mortality rate with logistic regression (logit, -5.312 þ 0.207 MELD; bootstrapped R2, 0.3295). MELD values of 19 or greater predicted 20% or greater mortality, whereas MELD scores less than 11 predicted less than 5% mortality. The model performed well for patients from Canada at all risk levels. In the Spanish validation set, in which all patients were treated with banding ligation, MELD predictions were accurate up to the 20% risk threshold. CONCLUSIONS: We developed a MELDbased model that accurately predicts mortality among patients with AVB, based on objective variables available at admission. This model could be useful to evaluate the efficacy of new therapies and stratify patients in randomized trials.
Keywords: Cirrhosis; Regression.
TIPS;
Prognostic
Model;
A
cute variceal bleeding (AVB) is a major complication of cirrhosis. Despite the improvement in prognosis in the past 3 decades,1–4 mortality during the bleeding episode remains high and ranges from 24% in unselected cirrhotic variceal bleeders5,6 to about 16% among those receiving the current standard of care (band ligation þ vasoactive drugs þ antibiotics).7 Several factors have been associated with a poor outcome in variceal bleeding. The most consistently reported risk indicators of death are Child–Pugh class or its components, model for end-stage liver disease (MELD) score, renal failure, bacterial infection at admission or shortly after, hypovolemic shock, active bleeding at endoscopy, hepatocellular carcinoma (HCC), and hepatic venous pressure gradient (HVPG) equal or greater than 20 mm Hg.5,7–11 Several prognostic models including these variables have been developed.5,7–9,12 These models could allow patient stratification according to the risk of death and could be used to guide therapeutic decisions, such as treating high-risk patients by a more aggressive approach. However, they are seldom used, in part because of the lack of external validation. Risk prediction in AVB has gained relevance since the demonstration that an early pre-emptive transjugular intrahepatic portosystemic shunt (TIPS) improves the outcome in high-risk patients in 2 randomized trials. In these trials high-risk patients were identified with hemodynamic criteria (HVPG 20 mm Hg shortly after admission)13 or with clinical criteria (Child C up to 13 points or Child B þ active bleeding at endoscopy despite vasoactive
Abbreviations used in this paper: AUROC, area under receiver operating characteristic curve; AVB, acute variceal bleeding; HCC, hepatocellular carcinoma; MELD, model for end-stage liver disease; HVPG, hepatic venous pressure gradient; ROC, receiver operating characteristic; TIPS, transjugular intrahepatic portosystemic shunt.
Logistic © 2014 by the AGA Institute 0016-5085/$36.00 http://dx.doi.org/10.1053/j.gastro.2013.10.018
February 2014
Materials and Methods Study Cohort and Data Collection We prospectively collected a database including all patients admitted to the Hospital Clinic (Barcelona, Spain) for portal hypertension–related bleeding (all sources and all etiologies) from October 2007 to December 2010. This was performed as part of a project to evaluate the impact of current therapies on the prognosis of portal hypertension–related bleeding and was approved by the Institutional Review Board of the Hospital Clinic. All data were gathered in the context of standard practice from the clinical records of the patients, and were anonymized and collected in a protected database. No specific procedures were conducted for the study and informed consent was not required. Patients with cirrhosis and acute bleeding from esophageal varices (all-comers) were considered eligible for the study. The diagnosis of cirrhosis was based on previous clinical history, liver biopsy, and/or unequivocal clinical data and compatible findings on imaging techniques. Esophageal variceal bleeding was confirmed by emergency endoscopy according to Baveno criteria.17 Exclusion criteria were HCC stage D according to the Barcelona Clinic liver cancer staging system.18 This exclusion criterion was based on the high mortality rate of these patients (>60%) and the high rate of therapeutic abstention. Baseline clinical, biochemical, and imaging data of patients were recorded prospectively at admission by a member of the Hepatic Hemodynamic Lab (E.R., E.L., S.S., or J.G.A.). Liver and renal function tests at admission and endoscopic and vasoactive treatments received were available for all patients, as well as rescue therapies and transfusion requirements during the episode. The record of each patient was updated at day 42 or death.
Therapeutic Interventions and Definitions In the period considered for this study, before the implementation of an “early TIPS” policy in our unit,14 patients were treated with the standard of care according to Baveno consensus workshops and American Association for the Study of Liver Diseases guidelines17,19: vasoactive drugs from admission (somatostatin or terlipressin), endoscopic band ligation (sclerotherapy if technically difficult or not feasible) within 12
413
hours of admission, and antibiotic prophylaxis (norfloxacin or ceftriaxone when indicated20). Rescue therapies (ie, Sengstaken–Blakemore balloon tamponade and TIPS) were applied when necessary (re-bleeding or failure to control bleeding). At day 5, combined secondary prophylaxis (b-blockers plus variceal ligation in 2- to 3-week intervals) was started when applicable. Time of admission to hospital was considered as time zero for the follow-up evaluation. The primary end point analyzed in this study was 6-week mortality (bleeding-related mortality).17
Prognostic Models We selected prognostic models offering continuous quantitative estimations of 6-week mortality in AVB. Although MELD was not developed specifically to predict mortality in the bleeding episode,21,22 we also evaluated the performance of this model (as modified by the United Network for Organ Sharing)22 because it has shown excellent predictions in patients with end-stage liver disease,23 and specifically in AVB.12,24,25 Together with MELD, D’Amico9 and Augustin7 models (described in detail in Supplementary Table 1) were selected for validation. In the study by Augustin,7 2 models were presented, a linear one (developed with logistic regression) and a nonlinear one (developed with classification and regression tree analysis). Only the linear model was considered because classification and regression tree analysis models do not provide continuous predictions. Finally, we tested Child–Pugh performance because it repeatedly has been reported as a good prognostic indicator in all clinical situations in cirrhosis, including the AVB episode.5–9,24,25
Statistical Analysis Models performance. To assess the performance of the prognostic models in predicting 6-week mortality, the discrimination, calibration, and overall performance of each model were studied. Discrimination refers to the ability to rank patients according to their risk of developing the outcome, whereas calibration refers to the ability to predict absolute risks (how closely the predicted probabilities agree with the actual outcomes).26 Calibration is of special interest because it allows defining decision thresholds according to the absolute risk of the outcome. The discriminative ability was assessed by receiver operating characteristic (ROC) curve analysis. Calibration was tested by plotting predicted and observed mortalities, and by the Hosmer–Lemeshow goodness-of-fit test after splitting the sample into quintiles.26 In the case of the Augustin7 and D’Amico9 models, absolute predicted mortalities were calculated according to the formulas presented in the original reports. In the case of MELD, the original publication did not report the temporal constant to specifically calculate 6-week mortality.21 After contacting the original authors, they provided a temporal constant of 0.827 for calculating 6-week mortality as described in the original report.21 Child–Pugh calibration could not be assessed because the data are not available to estimate the absolute probability of 6week mortality for each Child–Pugh score. The overall performance of the models was evaluated with R2 and the Brier score.27 A higher R2 and a lower Brier score indicate better performance. Update of MELD calibration. MELD was the best model in terms of discrimination and overall performance, but was significantly miscalibrated. Therefore, we decided to use the Hospital Clinic cohort (initially the validation cohort) to update MELD calibration by fitting a logistic regression model
CLINICAL LIVER
drug treatment).14,15 Both definitions have relevant problems: HVPG measurements are not widely available and may be difficult to perform in bleeding patients, and the definition of active bleeding and some components of the Child classification have subjectivity. Indeed, in a recent trial it was shown that both Child and active bleeding were evaluated inconsistently in different centers.16 Thus, better predictive models, well validated, with objective variables easily and early available from admission clearly are needed. The overall aim of this study was to improve risk prediction in AVB. For this purpose we first tested the performance of recently described prognostic models for AVB in a contemporary series of patients. The MELD model was found to be the best model in terms of discrimination, but it was significantly miscalibrated. Therefore, a second aim of the present study was to update MELD calibration to predict mortality from AVB. The third aim was to validate this new MELD calibration in 2 external series of patients with AVB.
Meld and Variceal Bleeding
414
Reverter et al
Gastroenterology Vol. 146, No. 2
introducing MELD as the only covariable. This procedure (“logistic calibration”) does not modify the discrimination of the model.28 We used penalized maximum likelihood estimation of coefficients to correct for overoptimism because this leads to more accurate predictions in external samples.28,29 The performance of this MELD-based model was validated internally with bootstrapping (200 random bootstrap samples of 178 patients drawn with replacement from the original sample).28
Data-driven model restricted to objective variables. To explore the best data-driven model to predict 6-week mortality in our cohort, we conducted a multivariable logistic regression analysis, restricted to objective variables available at admission.
External validation of MELD recalibration. Because the MELD model was updated, it required CLINICAL LIVER
further external validation. This was performed in 2 series of patients with esophageal variceal bleeding. The first sample included 240 consecutive patients with AVB admitted to 2 providing hospitals from the University of Alberta (Edmonton, Canada) between 2003 and 2008. These patients were managed according to the international guidelines (drugs þ endoscopy þ antibiotics, with TIPS as rescue therapy), and are described in detail in Supplementary Table 2. Adherence to antibiotic prophylaxis was 71%. Second, taking into account that patients in whom band ligation is applied effectively seem to have an excellent prognosis,7 we further explored MELD calibration in 221 patients from the Vall D’Hebron Hospital (Autonomous University of Barcelona, Spain), all treated with drugs þ band ligation þ antibiotics between 2002 and 2012 (Supplementary Table 2). This series was partially reported in a previous study.7 Discrimination and calibration of the MELD-based model in these validation sets were assessed as reported earlier. Data to calculate MELD was available for all patients in both series. Analysis was performed with the IBM SPSS Statistics 19.0 package (SPSS, Inc, Chicago, IL) and R30 (with the aid of the rms and predictABEL packages), by means of auditable and rerunnable script files.
Results Study Cohort, Treatment, and Outcomes A total of 303 patients admitted to the Hospital Clinic for portal hypertension–related bleeding were registered prospectively between October 2007 and December 2010. Among these, 198 patients had cirrhosis and esophageal variceal bleeding. Nineteen patients with advanced HCC (Barcelona Clinic liver cancer stage D) and 1 patient lost to follow-up evaluation were excluded. The final analysis included 178 patients (Figure 1). Baseline characteristics of the patients at admission are reported in Table 1. The overall 6-week mortality was 16%. Causes of death were uncontrolled bleeding in 9 patients (31%), liver failure in 8 patients (28%), and sepsis or multiorgan failure in 12 patients (41%). Supplementary Figure 1 shows the temporal distribution of deaths.
Performance of Prognostic Models for 6-Week Mortality Discrimination was studied by analyzing the ROC curves. MELD showed the highest discrimination (Figure 2),
Figure 1. Flowchart of patients admitted to our hospital for portal hypertension–related bleeding and selection of the study cohort. EV, esophageal variceal.
although the difference between MELD discrimination and that of the other models was not statistically significant. Calibration (ie, the agreement between predicted and observed death rates) is shown in Figure 3. All 3 models were miscalibrated significantly. The D’Amico9 model and MELD systematically overestimated the observed mortality. The Augustin7 model both overestimated and underestimated mortality at different levels of risk. MELD was
Table 1.Baseline Characteristics of Patients at Admission Age, y Male sex, n (%) Etiology of alcohol/HCV/virus þ alcohol/others, % Previous episodes of variceal bleeding, n (%) Albumin level, g/L Bilirubin level, mg/dL, median (IQR) Creatinine level, mg/dL Na level, mEq/L Leukocytes, 109/L INR Ascites present, n (%) Encephalopathy, n (%) MELD score, median (IQR) MELD-Na score, median (IQR) Child–Pugh class of A/B/C, % Child–Pugh score Active bleeding at endoscopy, n (%) Systolic arterial pressure < 100 mm Hg (first 3 h), n (%) Infection at admission, n (%) HCC, n (%) Endoscopic band ligation/sclerotherapy, % Somatostatin/terlipressin, % Need for balloon tamponade, n (%) Rescue TIPS, n (%) 6-Week mortality, n (%)
58.3 (12.4) 130 (73) 39/32/11/18 72 (40) 28.4 (4.8) 1.8 (2.1) 1.1 (0.7) 136.3 (6.2) 8.5 (4.4) 1.6 (0.53) 112 (63) 43 (24) 13.9 (7.5) 17.5 (8.9) 15/57/28 8.4 (2.0) 60 (34) 61 (34) 28 (16) 18 (10) 65/21 80/20 27 (15) 28 (16) 29 (16)
NOTE. Data are expressed as means (SD) unless otherwise stated. HCV, hepatitis C virus; INR, international normalized ratio; IQR, interquartile range.
Meld and Variceal Bleeding
415
Therefore, we decided to update MELD calibration as described in the Materials and Methods section. The equation of the updated model was as follows: logit ¼ –5.312 þ 0.207 $ MELD (nonbootstrapped R2, 0.341; bootstrapped R2, 0.330; Supplementary Table 3). Table 2 provides predicted mortalities for individual MELD values and Supplementary Figure 2 shows the prediction curve of the model. A MELD score of 11 was associated with a 5% risk of mortality, and could be a potentially useful prognostic threshold for defining low-risk patients. A MELD score of 19 was associated with a mortality rate of 20%, and could be a potentially useful threshold for defining a high risk of death. The internal calibration of the MELD-based model was excellent (Figure 4A). Variables reflecting the severity of bleeding, including a systolic arterial pressure less than 100 mm Hg within the first 3 hours from admission and active bleeding at endoscopy, did not significantly add to the predictive value of the MELDbased model (P ¼ .25 and P ¼ .55, respectively).
Figure 2. ROC curves of the models for predicting 6-week mortality in acute variceal bleeding. MELD had the greatest AUROC, indicating the best discrimination. Differences between MELD’s AUROCs and those from other models, however, were not statistically significant (P ¼ .0574 vs Augustin7 model; P ¼ .1194 vs D’Amico9 model; P ¼ .2179 vs Child–Pugh).
the model showing the (Supplementary Table 3).
best
overall
performance
Early Objective Prognostic Indicators for 6-Week Mortality We performed a multivariable analysis with objective variables available on admission to develop the best datadriven model (Supplementary Table 4). Only creatinine, bilirubin, and international normalized ratio (the components of MELD) remained in the final model (area under the ROC curve [AUROC], 0.778; nonbootstrapped R2, 0.314; bootstrapped R2, 0.268). Even this new model was fine-tuned to the data, it was not better than the MELD-based model.
Model Update: MELD Recalibration
External Validation of the MELD-Based Model
MELD was the best model in terms of discrimination and overall performance, and had the additional advantage over the other models of including only objective variables.
Finally, we tested the performance of MELD in 2 additional series of patients with AVB (Supplementary Table 2). The MELD-based model showed excellent discrimination
Figure 3. Calibration plots for (A) MELD, (B) D’Amico,9 and (C) Augustin7 models. To construct these plots, the sample was split into quintiles, and predicted mortality was plotted against observed mortality. Points below the diagonal line (ie, identity line, perfect prediction) indicate overestimation of mortality for that group. Points above the diagonal line indicate underestimation of mortality. P values are those of the Hosmer–Lemeshow goodness-of-fit test (the lower the P value, the worse the agreement between the observed and predicted mortalities). (A) MELD and (B) the D’Amico9 model overestimated mortality. (C) The Augustin7 model both overestimated and underestimated mortality.
CLINICAL LIVER
February 2014
416
Reverter et al
Gastroenterology Vol. 146, No. 2
Table 2.Predicted Mortality for Each MELD Value According to the Updated MELD-Based Model MELD value
CLINICAL LIVER
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Predicted mortality, % 2.5 3.1 3.8 4.6 5.6 6.8 8.2 9.9 11.9 14.3 17.0 20.1 23.6 27.6 31.9 36.6 41.5 46.6 51.7 56.9 61.9 66.6 71.1
(AUROC, 0.873) and calibration (Figure 4B) in the cohort from the University of Alberta (Edmonton, Canada). In the cohort from Vall d’Hebron Hospital (Barcelona, Spain) including patients selected on the basis of receiving band ligation (þ antibiotics þ drugs), the MELD-based model also had good discrimination (AUROC, 0.794). Calibration was acceptable, except in the upper quintile of the patients, in which our model predicted a higher mortality than that actually observed (Figure 4C). In all 3 series, MELD scores potentially defining arbitrary thresholds of low and high risk of death (11 and of 19) consistently were associated with a mortality rate of approximately 5% and 20%, respectively (Figure 4D–F). Therefore, even though the MELD-based model overpredicted mortality for high MELD values in the second validation set, it still would be a useful tool to accurately identify patients with a risk of death of 20% or higher.
Discussion The present study provides new information to improve risk prediction in patients with cirrhosis and acute variceal bleeding. Specifically, we show that MELD outperforms other models designed to predict prognosis in AVB. Furthermore, we propose a model that assigns a probability of 6-week mortality for each MELD value, and we provide external validation for our MELD calibration in 2 additional series of patients. The recent demonstration that a stratified management of patients with AVB can improve survival in high-risk patients13,14 has emphasized the importance of having
accurate prognostic tools in AVB. However, the criteria used in those trials to define high risk have a low clinical accessibility (HVPG measurement within first 24 h) or include subjective variables (Child–Pugh, active bleeding). In addition, it has been questioned recently whether all patients defined as high risk in the inclusion criteria of the EarlyTIPS trial are indeed high risk, especially those at Child B class þ active bleeding. Some series have shown a low mortality rate in this subgroup of patients7 and others have shown significant heterogeneity in scoring the patients.16 Thus, there is a clear need for better prediction models in AVB, preferably with easily obtained, objective, and reproducible variables available early after patient admission. One of the problems in the use of prognostic models in AVB is the lack of external validation. Thus, our first goal was to evaluate the performance of previously described models. For this purpose we chose 2 models specifically developed for AVB,7,9 which also offered continuous predictions of 6-week mortality. We additionally evaluated MELD because it has shown prognostic value in almost every clinical situation in cirrhosis,23 including variceal bleeding,12,24,25 and Child–Pugh score because it is one of the most widely used bedside prognostic tools in cirrhosis. Our findings show that MELD was the best model in terms of discrimination. However, calibration was poor for all models and good calibration (agreement between expected and observed mortality) is essential to derive decision thresholds to guide therapeutic decisions. This lack of calibration could be explained by several reasons. MELD was developed in a different setting (patients treated with TIPS) in a series of patients from more than a decade ago, before the implementation of several effective treatments for patients with cirrhosis and, therefore, miscalibration could be expected. On the other hand, in the D’Amico9 study many patients did not receive what is currently accepted as the standard of care. Both the D’Amico9 and Augustin7 models included the presence of HCC as a prognostic variable, but without differentiating between HCC stages. Although the presence of HCC up to stage C has little short-term prognostic significance, stage D patients have exceedingly high short-term mortality, and many times these patients are managed with symptomatic measures if presenting with AVB. In the present study, rather than trying to develop a completely new prognostic model, we used a strategy of model updating, which is a better way to develop stable and generalizable models.31 For this purpose, we focused on MELD for several reasons. Because MELD had a good discriminative value it needed less extensive updating. We could recalibrate the model easily to make accurate predictions with logistic calibration, which does not change the structure and the discrimination of the model. The other models would require much more extensive updating, and the sample size of the present study would be inadequate. In addition, MELD is based in simple and reliable measurements, which are important criteria in developing clinically useful prognostic models,31 and it is widely implemented worldwide for the prediction of outcomes in cirrhosis. Of note, variables reflecting the severity of bleeding, which have
Meld and Variceal Bleeding
417
CLINICAL LIVER
February 2014
Figure 4. (A–C) Calibration of the MELD-based model in the 3 samples. The updated MELD-based model showed an excellent calibration in the Hospital Clinic sample (internal calibration of the model), but also in the University of Alberta sample (external calibration). In the Vall D’Hebron sample it showed a good calibration for low values of MELD, but overpredicted mortality in the upper quintile of MELD (P values are those of the Hosmer–Lemeshow goodness-of-fit test). (D–F) The relationship between MELD and mortality in the 3 series of patients. These plots were constructed with nonparametric regression, which shows smoothed actual mortality rates. Tick marks are drawn at actual MELD values. The upper and lower 5% of cases were trimmed. In the 3 series a MELD score of 11 was associated with a mortality rate of approximately 5%, and a MELD score of 19 was associated with a mortality rate of approximately 20%. The Vall D’Hebron series showed a lower mortality rate for MELD scores greater than 20, as compared with the other 2 series.
the additional problem of lacking an unequivocal definition, did not add predictive value to the MELD-based model. Because model updating necessarily requires additional validation, we searched for 2 external validation samples. The first validation set came from a comparable setting: tertiary university hospitals with wide experience in the management of AVB applying the current standard of care,17,19 but in a different country. We showed that the performance of the MELD-based model was excellent in the Alberta series, with an excellent agreement between predicted and observed 6-week mortalities along the whole spectrum of disease severity. Indeed, arbitrary thresholds potentially defining low risk and high risk (MELD scores, 11 and 19, respectively) showed an almost perfect fit to observed 5% and 20% mortality rates (Figure 4E). With the second validation set, from a different University Hospital in Barcelona, we aimed to evaluate the performance of the MELD-based model in a selected sample of patients in whom the currently considered optimal therapy (banding ligation þ vasoactive drugs þ antibiotics) was applicable. These
patients have been shown to have a better overall prognosis than all-comers.6,7 In this sample, MELD-based predictions overestimated mortality only in the most advanced patients. Still, MELD values of 11 and 19 again showed an excellent fit to the observed 5% and 20% mortality rates (Figure 4F), suggesting that the model would be equally useful for risk stratification in this selected category of patients. Our results might set the framework to evaluate the impact of new therapies (including early TIPS) in real-practice outcomes, and to identify the patients who might benefit most from specific treatments. Indeed, the robustness of the model might serve to compare results obtained in different centers, at different periods, and with different treatment protocols. However, it is important to note that these results might not have immediate clinical applicability because the only patienttailored strategy proved to date to improve the outcome in AVB has been early TIPS, and patients were selected on the basis of Child and active bleeding.14,15 In summary, in patients with cirrhosis and AVB, MELD offered an objective and accurate prognostic prediction with
418
Reverter et al
variables available early after admission. MELD could be more efficient than the current criteria for selecting highrisk patients who might benefit from more aggressive treatments. Our proposed MELD-based predictions might be useful in refining the use of early TIPS, in evaluating the impact of new therapeutic strategies on patient prognosis, and in improving risk stratification in future clinical trials.
Supplementary Material Note: To access the supplementary material accompanying this article, visit the online version of Gastroenterology at www.gastrojournal.org, and at http://dx.doi.org/10.1053/j. gastro.2013.10.018. CLINICAL LIVER
References 1. Graham D, Smith J. The course of patients after variceal hemorrhage. Gastroenterology 1981;80:800–806. 2. Chalasani N, Kahi C, Francois F, et al. Improved patient survival after acute variceal bleeding: a multicenter, cohort study. Am J Gastroenterol 2003;98:653–659. 3. Carbonell N, Pauwels A, Serfaty L, et al. Improved survival after variceal bleeding in patients with cirrhosis over the past two decades. Hepatology 2004;40:652–659. 4. Stokkeland K, Brandt L, Ekbom A, et al. Improved prognosis for patients hospitalized with esophageal varices in Sweden 1969-2002. Hepatology 2006; 43:500–505. 5. Augustin S, Muntaner L, Altamirano JT, et al. Predicting early mortality after acute variceal hemorrhage based on classification and regression tree analysis. Clin Gastroenterol Hepatol 2009;7:1347–1354. 6. Amitrano L, Guardascione MA, Manguso F, et al. The effectiveness of current acute variceal bleed treatments in unselected cirrhotic patients: refining short-term prognosis and risk factors. Am J Gastroenterol 2012; 107:1872–1878. 7. Augustin S, Altamirano J, Gonzalez A, et al. Effectiveness of combined pharmacologic and ligation therapy in highrisk patients with acute esophageal variceal bleeding. Am J Gastroenterol 2011;106:1787–1795. 8. Abraldes JG, Villanueva C, Banares R, et al. Hepatic venous pressure gradient and prognosis in patients with acute variceal bleeding treated with pharmacologic and endoscopic therapy. J Hepatol 2008;48:229–236. 9. D’Amico G, de Franchis R. Upper digestive bleeding in cirrhosis. Post-therapeutic outcome and prognostic indicators. Hepatology 2003;38:599–612. 10. Moitinho E, Escorsell A, Bandi JC, et al. Prognostic value of early measurements of portal pressure in acute variceal bleeding. Gastroenterology 1999;117:626–631. 11. Moitinho E, Planas R, Bañares R, et al. Multicenter randomized controlled trial comparing different schedules of somatostatin in the treatment of acute variceal bleeding. J Hepatol 2001;35:712–718. 12. Bambha K, Kim WR, Pedersen R, et al. Predictors of early re-bleeding and mortality after acute variceal haemorrhage in patients with cirrhosis. Gut 2008; 57:814–820.
Gastroenterology Vol. 146, No. 2 13. Monescillo A, Martinez-Lagares F, Ruiz-del-Arbol L, et al. Influence of portal hypertension and its early decompression by TIPS placement on the outcome of variceal bleeding. Hepatology 2004;40:793–801. 14. Garcia-Pagan JC, Caca K, Bureau C, et al. Early use of TIPS in patients with cirrhosis and variceal bleeding. N Engl J Med 2010;362:2370–2379. 15. Garcia-Pagan JC, Di Pascoli M, Caca K, et al. Use of early-TIPS for high-risk variceal bleeding: results of a post-RCT surveillance study. J Hepatol 2013;58:45–50. 16. Bosch J, Thabut D, Albillos A, et al. Recombinant factor VIIa for variceal bleeding in patients with advanced cirrhosis: a randomized, controlled trial. Hepatology 2008;47:1604–1614. 17. de Franchis R. Revising consensus in portal hypertension: report of the Baveno V consensus workshop on methodology of diagnosis and therapy in portal hypertension. J Hepatol 2010;53:762–768. 18. Bruix J, Sherman M. Management of hepatocellular carcinoma: an update. Hepatology 2011;53:1020–1022. 19. Garcia-Tsao G, Sanyal AJ, Grace ND, et al. Prevention and management of gastroesophageal varices and variceal hemorrhage in cirrhosis. Hepatology 2007; 46:922–938. 20. Fernandez J, Ruiz dA, Gomez C, et al. Norfloxacin vs ceftriaxone in the prophylaxis of infections in patients with advanced cirrhosis and hemorrhage. Gastroenterology 2006;131:1049–1056. 21. Malinchoc M, Kamath PS, Gordon FD, et al. A model to predict poor survival in patients undergoing transjugular intrahepatic portosystemic shunts. Hepatology 2000; 31:864–871. 22. Available from: www.mayoclinic.org/meld/mayomodel6. html. 23. Kamath PS, Kim WR. The model for end-stage liver disease (MELD). Hepatology 2007;45:797–805. 24. Chalasani N, Kahi C, Francois F, et al. Model for endstage liver disease (MELD) for predicting mortality in patients with acute variceal bleeding. Hepatology 2002; 35:1282–1284. 25. Amitrano L, Guardascione MA, Bennato R, et al. MELD score and hepatocellular carcinoma identify patients at different risk of short-term mortality among cirrhotics bleeding from esophageal varices. J Hepatol 2005; 42:820–825. 26. Altman DG, Vergouwe Y, Royston P, et al. Prognosis and prognostic research: validating a prognostic model. BMJ 2009;338:b605. 27. Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010; 21:128–138. 28. Steyerberg EW. Clinical prediction models. A practical approach to development, validation and updating. New York: Springer ScienceþBusiness Media, LLC, 2009. 29. Moons KG, Donders AR, Steyerberg EW, et al. Penalized maximum likelihood estimation to directly adjust diagnostic and prognostic prediction models for overoptimism: a clinical example. J Clin Epidemiol 2004; 57:1262–1270.
February 2014 30. Available from: www.r-project.org. 31. Moons KG, Altman DG, Vergouwe Y, et al. Prognosis and prognostic research: application and impact of prognostic models in clinical practice. BMJ 2009;338:b606.
Meld and Variceal Bleeding
419
Acknowledgments The authors thank Laura Rocabert, Ángels Baringo, and Rosa Sáez for expert nursing support; and Clara Esteva, Rachel Borowski, and Jackie Love for secretarial assistance. Conflicts of interest The authors disclose no conflicts.
Author names in bold designate shared co-first authorship.
Reprint requests Address requests for reprints to: Juan G. Abraldes, MD, Division of Gastroenterology, 1-51 Zeidler-Ledcor Centre, University of Alberta Campus, Edmonton, Alberta T6G 2X8, Canada. e-mail: [email protected]; fax: (780) 248-1895.
Funding This study was supported in part by grants from the Instituto de Salud Carlos III, Ministerio de Economía y Competitividad (PI11/00883 to J.G.A., PS09/01261 to J.B., and Fondo de Investigación Sanitaria (FIS): PI12/01759 to J.G.); co-financed by Fondo Europeo de Desarrollo Regional (FEDER) funds (EU, “Una manera de hacer Europa”); Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas (CIBERehd) is funded by the Instituto de Salud Carlos III; and a Río Hortega award, Instituto de Salud Carlos III (E.R.).
CLINICAL LIVER
Received May 3, 2013. Accepted October 16, 2013.
419.e1
Reverter et al
Gastroenterology Vol. 146, No. 2
0.10 0.05 0.00
Fraction of deaths
0.15
Temporal distribution of bleeding related deaths
0
10
20
30
40
Days after index bleeding
Probability of death at 6 weeks
Supplementary Figure 1. Temporal distribution of the bleeding-related mortality (42 days). Most deaths occurred within the first 7 days of index bleeding.
0.8
0.6
0.4
0.2
10
20
30
40
MELD
Supplementary Figure 2. Prediction curve of the updated MELD-based model. The grey area represents the amplitude of 95% confidence intervals. The formula of the regression model was as follows: logit ¼ – 5.312 þ 0.207 * MELD. The probability of 6-week mortality was estimated as follows: p ¼ 1 / (1 þ e-logit). The bootstrapped R2 of the model was 0.3295.
February 2014
Meld and Variceal Bleeding
419.e2
Supplementary Table 1.Prognostic Models Evaluated Model
Equation
D’Amico and de Franchis9 (Hepatology 2003) Augustin et al7 (American Journal of Gastroenterology 2011) MELD
Logit ¼ -0.083 þ 0.88 $ encephalopathy (1 ¼ no; 2 ¼ mild/moderate; 3 ¼ severe) þ 0.22 $ bilirubin (mg/dL) þ 1.12 $ HCC (0 ¼ no; 1 ¼ yes) - 0.86 $ albumin (g/dL) Logit ¼ -7.226 þ 0.475 $ Child–Pugh score (5–15 points) þ 1.746 $ HCC (0 ¼ no; 1 ¼ yes) þ 1.179 $ creatinine 1.0 mg/dL (0 ¼ no; 1 ¼ yes) 6.43 þ 11.2 $ ln (INR) þ 9.57 $ ln (creatinine; mg/dL) þ 3.78 $ ln (bilirubin; mg/dL)
NOTE. We studied the performance of MELD score and previously described models for 6-week mortality by D’Amico9 and Augustin.7 MELD score was calculated according to the United Network for Organ Sharing modification22: creatinine and bilirubin were rounded to 1 if below and creatinine was capped to 4 if above or in patients with renal replacement therapy. D’Amico9 and Augustin7 models were calculated with the equations outlined in the table. HCC: hepatocellular carcinoma; INR, international normalized ratio.
Supplementary Table 2.Baseline Characteristics of Patients From Vall d’Hebron Hospital and Edmonton Hospital (Validation Sets)
Age, y Sex, male/female, % Etiology, OH/HCV/virus þ OH/others, (%) Albumin level, g/L Bilirubin level, mg/dL, median (IQR) Creatinine level, mg/dL INR MELD score, median (IQR) Child–Pugh class A/B/C, % Child–Pugh score HCC, n (%) Banding/sclerotherapy, % Balloon tamponade, n (%) Rescue TIPS, n (%) 6-Week mortality, n (%)
Vall d’Hebron (n ¼ 221)
University of Alberta (n ¼ 240)
60.4 (14.4) 67/33 37/36/12/15 27.1 (4.7) 1.6 (1.4) 1.1 (0.6) 1.8 (0.65) 14.9 (7.0) 25/49/26 8.3 (2.2) 23 (10.4) 100/0 25 (11.3) 4 (1.8) 27 (12.2)
56 (12.3) 75/25 48/13/16/23 28 (7.5) 2.0 (2.5) 1.2 (0.7) 1.6 (0.52) 14.3 (8.6) 14/51/35 8.8 (2.1) 19 (7.9) 92.9/3.3 13 (5.4) 16 (6.7) 38 (15.8)
NOTE. Both cohorts were similar to the training set, except for a slightly lower 6-week mortality rate in the Vall d’Hebron cohort. Data are expressed as means (SD) unless otherwise stated. HCV, hepatitis C virus; INR, international normalized ratio; IQR, interquartile range.
419.e3
Reverter et al
Gastroenterology Vol. 146, No. 2
Supplementary Table 3.Performance of the Evaluated Models and of the Updated MELD in the Hospital Clinic Barcelona Cohort
MELD Augustin7 model D’Amico9 model Child–Pugh MELD with updated calibration
R2
Brier score
0.325 0.123 0.161 0.341
0.106 0.133 0.171 0.097
AUROC (95% confidence interval) 0.795 0.723 0.703 0.740 0.795
Hosmer–Lemeshow test (P value)
(0.689–0.901) (0.621–0.824) (0.582–0.824) (0.639–0.841) (0.689–0.901)
.012 .006 .000 .647
NOTE. R2 and Brier score provide a global evaluation of the model (discrimination þ calibration): a higher R2 and a lower Brier score indicate better performance. AUROC evaluates the discrimination of the model; higher values indicate better discrimination. The P value of the Hosmer–Lemeshow test evaluate the calibration of the model; a lower P value indicates worse calibration.
Supplementary Table 4.Univariable and Multivariable Analysis of Objective Variables Associated With 6-Week Mortality Variable Leukocytes, 109/L Na, mEq/L Creatinine level, mg/dL ASAT level, IU/L Albumin level, g/L Bilirubin level, mg/dL, median (IQR) INR Age, y Hepatocellular carcinoma, n (%) Portal vein thrombosis, n (%) Alcoholic etiology, n (%)
Survivors (n ¼ 149) 8.4 136.5 0.98 89 28.6 1.7 1.54 57.8 18 14 58
(4.5) (5.6) (0.6) (141) (4.5) (1.8) (0.4) (12.5) (12.1) (9.4) (38.9)
Dead (n ¼ 29) 9.1 135 1.6 144 26.8 3.1 1.95 60.7 1 2 11
(3.4) (8.6) (0.7) (168) (6.8) (5.4) (0.8) (11.8) (3.4) (6.8) (37.9)
Univariable, P value .419 .389 <.001 .063 .083 <.001 <.001 .248 .168 .668 .920
Multivariable, P value
<.001
.027 .014
NOTE. Data are expressed as means (SD) unless otherwise stated. Creatinine, bilirubin, and INR were log-transformed for the analysis. At multivariable analysis only creatinine, INR, and bilirubin (already included in MELD) were associated with 6-week mortality. INR, international normalized ratio; IQR, interquartile range.