The Decreasing Predictive Power of MELD in an Era of Changing Etiology of Liver Disease

Elizabeth L. Godfrey; Tahir H. Malik; Jennifer C. Lai; Ayse L. Mindikoglu; N. Thao N. Galván; Ronald T. Cotton; Christine A. O'Mahony; John A. Goss; Abbas Rana


American Journal of Transplantation. 2019;19(12):3299-3307. 

In This Article


Study Population

Data for the primary outcome measures included a population of 120 393 individuals waitlisted as recipients. Donor information was available for 56 111 individuals.

Data describing changing indications for transplantation in the modern era included data from 131 644 listed individuals with indications included in the OPTN database.

Declining Predictive Power of the MELD

The c-statistic of the MELD score with 90-day waitlist mortality demonstrated a downward trend since 2002, shown in Figure 1. In 2002, the c-statistic was 0.80, and peaked at 0.802 in 2003 before beginning to decline. By 2007, the c-statistic had dropped to 0.75, a statistically significant decline from every preceding year, and remained statistically lower than the early years of the MELD throughout the analysis period. Numerically, it continued to decrease, reaching 0.70 in 2014 and 2015. This suggests an approximate decrease in correspondence between MELD and mortality of 12.5% in the modern era. The secondary and tertiary definitions of death (waitlist death according to the Social Security death index, and presumptive death after removal from the waitlist) returned higher c-statistics, but also demonstrated a similar numerically downward trend that represented a statistically significant decline from the 2002–2004 period to the 2012–2014 period. An identical analysis was conducted with exceptions cases included rather than excluded, for the purpose of broadly assessing the entire waitlist. It demonstrated a flatter decline and slightly higher overall concordance (Figure 1), beginning at just under 0.80 in 2002, peaking at 0.81 in 2003, and declining to 0.73 in 2014 and 2015. A similar decline from the 2002–2004 period to the 2013–2015 period was found to be statistically significant. The higher concordance between MELD and 90-day mortality with exceptions is likely related to the low MELDs and generally good mortality profile for those individuals who are listed for hepatocellular carcinoma.

Figure 1.

Decline of c-statistic of the MELD and MELD-Na with 90-day mortality. MELD, Model for End-Stage Liver Disease; MELD-NA, sodium Model for End-Stage Liver Disease [Color figure can be viewed at]

Analysis of the MELD-Na demonstrated a less severe decline than the MELD; correspondence with waitlist mortality, however, is still trending downward. The c-statistic can be retrospectively evaluated leading back almost to the institution of MELD; in 2004, it was 0.78 and decreased to 0.71 in 2014 and 0.70 in 2015. Years 2013 to 2015 demonstrated a statistically significant lower c-statistic than 2004 to 2006 by 5% to 10%.

MELD and age at time of listing are increasing; a statistically significant increase in both was observed from 2002 to 2016. Age increased from 50.1 years old at listing in 2002 to 53.9 in 2016 (Figure 2), while listing MELD increased from 18.4 to 22.7 in the same period (Figure 3). This suggests an increasingly aging and sick population on the waitlist. The c-statistic of MELD and mortality was not significantly different between patients over and under 65 (>65 = 0.73, <65 = 0.74; P = .97).

Figure 2.

Increasing age at listing [Color figure can be viewed at]

Figure 3.

Increasing MELD score at listing. MELD, Model for End-Stage Liver Disease [Color figure can be viewed at]

While MELD at listing is increasing, calculated MELD at death is changing less rapidly; it was 34.7 in 2002 and 36.2 in 2016, a statistically significant, but clinically very small increase. In the same interval, most years did not exhibit significantly higher MELDs than that reported in 2002; this indicates minimal change in the maximum severity of disease, as reflected by the MELD, at which patients can be maintained on the list.

The competing risk analysis showed a statistically significant decrease from the 2002–2003 period to the 2013–2015 period in subdistribution hazard ratio, with respect to the relationship between increase in MELD score and 90-day death (Figure 4A). While not every year displayed a significantly lower hazard ratio than the preceding, the overall trend was downward. The MELD-Na demonstrated a higher subdistribution hazard ratio with respect to death within 90 days of listing than the MELD, point by point, but also had a statistically significant, if smaller, decrease between the 2003–2005 period and the 2016–2017 period (Figure 4B). Examining the data in reverse, with transplantation within 90 days as the primary outcome and death the competing risk, a steady rise is observed in the hazard ratio across the modern transplant era (since 2002). That is, the incidence of transplant within 90 days is more strongly predicted by a 1-point rise in MELD or MELD-Na score in 2016 than in 2004. The decreasing hazard ratio of the MELD for death does not mirror the corresponding relationship in transplantation.

Figure 4.

A, Competing risk analysis: increasing relationship between MELD and transplantation over time does not mirror relationship between MELD and death. B, Competing risk analysis: MELD-Na shows similar change but higher overall correspondence with death. MELD, Model for End-Stage Liver Disease; MELD-NA, sodium Model for End-Stage Liver Disease [Color figure can be viewed at]

Another way to assess a measure's ability to predict an outcome is the Brier score. It produces a number between 0 and 1, where 1 indicates perfect failure of prediction (ie, the measure is always wrong) and 0 is perfect success of prediction. The Brier score does, however, square a statistic derived from the error of the prediction, so that a score of 0.25 indicates predicting an event as a 50/50 split (equivalent to chance at predicting a binary outcome). The correspondence of MELD with 90-day mortality over time can be assessed through this score.

Over the years 2002 to 2015, the Brier score for MELD and 90-day mortality was assessed using the probabilities of mortality initially presented by Wiesner et al in their UNOS committee paper that led to the adoption of the MELD for allocation.[17] The Brier score increases almost every year from 2002 to 2015, from 0.0829 to 0.1093. From 2002 to 2005, the Brier score is in the low 0.08 range, from 0.0812 to 0.0829. From 2005 to 2011, the Brier score runs in the high 0.08 s to the low 0.09 s, reaching 0.0928 in 2011. The Brier score then continues to rise to 0.1093 by 2015 (Figure 5). Overall, these indicate that MELD does, in fact, predict 90-day mortality, as they are considerably better than chance, but show a declining correspondence with mortality.

Figure 5.

Increasing (worsening) Brier score of MELD and 90-day mortality. MELD, Model for End-Stage Liver Disease [Color figure can be viewed at]

Declining Incidence of Hepatitis C

The proportion of patients who were listed with an HCV-related diagnosis decreased from 32.5% to 17.1% from 2002 to 2015 (and even more dramatically, to 12.6% in 2016). While not all patients have a documented HCV serostatus at transplant, those who were tested and HCV-positive at transplant declined from 19.9% in 2002 to 13.3% in 2016. Both are statistically significant changes. Where this becomes particularly suggestive with regard to the declining accuracy of MELD is that the c-statistic for the MELD and mortality in HCV-listed patients is higher overall, at 0.766, than in HCV negative patients at 0.733. The change in HCV-positive status is shown in Figure 6.

Figure 6.

Declining rate of HCV-positive status at listing and at transplant. *Points that differ from each other by a statistically significant margin with a P value < .0001. †Points that differ from each other by a statistically significant margin with a P value < .0001. HCV, hepatitis C virus [Color figure can be viewed at]

Of note, while the decline in HCV positivity is in part due to the availability of direct-acting antivirals (DAA), the c-statistic drop for HCV-related listings between the pre-DAA era and post-DAA era is not significantly different from the c-statistic drop for non–HCV-related listings—so as far as can be concluded from the available data, the administration of DAAs themselves is not directly affecting the predictive nature of the MELD or MELD-Na.

Variation in C-statistic by Diagnosis

The MELD displayed substantially different correspondence with 90-day mortality depending on the etiology of liver disease. Table 1 breaks down the patient population by diagnosis at listing. The correspondence between MELD and mortality was highest in primary sclerosing cholangitis, primary biliary cholangitis, and acute liver failure caused by HCV, at 0.80, 0.79, and 0.78, respectively, with cirrhosis secondary to hepatitis B infection and autoimmune hepatitis both following closely at 0.78.

Notably, this correspondence was weaker in fatty liver and alcoholic cirrhosis with or without hepatitis, all falling at approximately 0.73. Collectively, these etiologies accounted for half of listed patients in 2017, up from less than 20% in 2002.

The c-statistic of MELD-Na with 90-day mortality was slightly higher for all etiologies of liver disease than that of the MELD, but order of increasing correspondence is the same as MELD. The cholestatic diseases primary sclerosing cholangitis and primary biliary cirrhosis and HCV cirrhosis are associated with c-statistics of 0.80, 0.79, and 0.79. Fatty liver and alcoholic cirrhosis with or without hepatitis display c-statistics at 0.74, 0.73, and 0.74, respectively.

The lower correspondence of MELD and MELD-Na with 90-day death in fatty liver (NASH) or alcoholic livery disease (ALD) patients does not appear to be due to uniformly worse outcomes; rate of waitlist death is not significantly greater in patients with those etiologies than in others, nor is the risk of waitlist death from cardiac causes higher in NASH as might be expected.

Changing Prevalence of Disease Etiologies

The changing causes of end-stage liver disease can be observed in the large-scale shifts in listing indications in the modern era. Including patients who received exceptions, alcoholic cirrhosis as an indication increased from 13.7% of the listed population, to 24.9%, while HCV cirrhosis (or Type C cirrhosis) has decreased by approximately half, from 30.2% to 14.2%. Fatty liver cirrhosis has increased over a thousand-fold, from 0.14% to 17.5%. These data are in keeping with previously observed trends regarding the relative increase of fatty liver and alcoholic cirrhosis. Other indications are detailed in Table 1.