Performance of the model for end-stage liver disease score for mortality prediction and the potential role of etiology

Highlights

• Discrimination of MELD is widely reported as fair to good, although its calibration is still unclear.

• In 2 cirrhosis cohorts we found barely acceptable c-statistics, which were significantly worse in patients with non-viral etiology.

• Calibration was largely unsatisfactory with the Mayo and UNOS MELD versions.

• Validated recalibrations of MELD-Mayo and UNOS versions are presented which allow reliable predictions for clinical practice.

• Age, albumin and ascites as the indication for TIPS are candidate variables for an update to the MELD-TIPS score.

Background & Aims

Although the discriminative ability of the model for end-stage liver disease (MELD) score is generally considered acceptable, its calibration is still unclear. In a validation study, we assessed the discriminative performance and calibration of 3 versions of the model: original MELD-TIPS, used to predict survival after transjugular intrahepatic portosystemic shunt (TIPS); classic MELD-Mayo; and MELD-UNOS, used by the United Network for Organ Sharing (UNOS). We also explored recalibrating and updating the model.

Methods

In total, 776 patients who underwent elective TIPS (TIPS cohort) and 445 unselected patients (non-TIPS cohort) were included. Three, 6 and 12-month mortality predictions were calculated by the 3 MELD versions: discrimination was assessed by c-statistics and calibration by comparing deciles of predicted and observed risks. Cox and Fine and Grey models were used for recalibration and prognostic analyses.

Results

In the TIPS/non-TIPS cohorts, the etiology of liver disease was viral in 402/188, alcoholic in 185/130, and non-alcoholic steatohepatitis in 65/33; mean follow-up±SD was 25±9/19±21 months; and the number of deaths at 3-6-12 months was 57-102-142/31-47-99, respectively. C-statistics ranged from 0.66 to 0.72 in TIPS and 0.66 to 0.76 in non-TIPS cohorts across prediction times and scores. A post hoc analysis revealed worse c-statistics in non-viral cirrhosis with more pronounced and significant worsening in the non-TIPS cohort. Calibration was acceptable with MELD-TIPS but largely unsatisfactory with MELD-Mayo and -UNOS whose performance improved much after recalibration. A prognostic analysis showed that age, albumin, and TIPS indication might be used to update the MELD.

Conclusions

In this validation study, the performance of the MELD score was largely unsatisfactory, particularly in non-viral cirrhosis. MELD recalibration and candidate variables for an update to the MELD score are proposed.

Lay summary

While the discriminative performance of the model for end-stage liver disease (MELD) score is credited to be fair to good, its calibration, the correspondence of observed to predicted mortality, is still unsettled. We found that application of 3 different versions of the MELD in 2 independent cirrhosis cohorts yielded largely imprecise mortality predictions particularly in non-viral cirrhosis. Thus, we propose a recalibration and suggest candidate variables for an update to the model.

Introduction

The model for end-stage liver disease (MELD) score is used worldwide to predict the risk of mortality in patients with cirrhosis and to prioritize patients for orthotopic liver transplant (OLT). The original MELD was developed using Cox regression to predict survival after elective transjugular intrahepatic portosystemic shunt (TIPS) in patients with cirrhosis.¹ It included disease etiology, bilirubin, creatinine and international normalized ratio (INR), as predictors. We will refer to this score as the MELD-TIPS. Subsequently, MELD-TIPS was adapted by removing the predictor “etiology” and multiplying the predictors’ coefficients by 10.² This is the classic MELD and is commonly adopted to predict mortality in a broader range of patients with advanced liver disease. We will refer to this as the MELD-Mayo score.

The MELD-Mayo score was later modified by the United Network for Organ Sharing (UNOS) in 2002 to restrict the range of possible predictions,³ and in 2016 to account for hyponatremia.⁴ This modified score, which we will refer to as the MELD-UNOS, is commonly used for organ allocation priority for OLT.

Therefore, 3 different versions of the MELD have entered clinical practice and online calculators are available for each of them,[5], [6], [7] while it is not always clear which one should be used.

Several studies have investigated the performance of the MELD (mostly MELD-Mayo) models, reporting promising discrimination with concordance statistic ranging from 0.66 to 0.83.^8,9 However, some studies have identified unsatisfactory performance in several patient subgroups, which prompted exceptions to the MELD and model revisions.^10,11 Moreover, studies of the correspondence between observed and expected mortality (calibration) at defined observation times are lacking.⁹

Therefore, while the MELD helps physicians in ranking patients according to risk, it is hardly applicable when mortality probability is a key for clinical decisions or simply to inform the patient on his expected survival.

In the present study we assessed the discrimination and calibration performance of mortality predictions by the 3 aforementioned MELD scores in 2 independent cohorts of patients with cirrhosis. We also explored recalibrating and updating the model.