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


The major limitation of this study is that it is able to report a statistical change in a tool's predictive power, but is not able to provide a thorough evaluation of the clinical implications of that decreasing predictive power. Since the MELD/MELD-Na is only one piece of complex allocations decisions, it isn't clear that this change has affected how transplant teams make decisions about their patients. However, it is important that clinicians remain aware that the scores that are used to prioritize individuals based on need are not static measures, and that policy makers continually reassess the role of the scoring systems on which we rely.

The temporal decline in predictive ability of both MELD and MELD-Na with 90-day mortality reported here suggests that the score currently relied upon by our current allocation system may have some emerging limitations in predicting the patients most in need of OLT, and therefore that the methods used to assess priority in demand for liver allografts may be less effective than in the past. The current c-statistic between MELD-Na and 90-day mortality is 0.72, which demonstrates that MELD-Na remains predictive of mortality at this time, but determining why the score's predictive capability is declining will be crucial in shaping future policy.

The competing risk analysis corroborates the conclusions suggested by the c-statistic, if to a lesser extent, because it returned a statistically significant decrease since 2002 in sub-distribution hazard ratio between MELD or MELD-Na and death within 90 days, even with death as a competing risk. This indicates that the decreasing magnitude of the relationship between MELD and death is not due merely to increased transplantation or prioritization of transplantation to individuals with high scores. Additionally, the increased magnitude of relationship between MELD/MELD-Na and transplantation from the early 2000s to the present does indicate higher-scored individuals are increasingly prioritized for transplantation. The Brier score also corroborates the decreasing effectiveness at MELD in predicting 90-day mortality, while simultaneously confirming that it maintains some usefulness in this area.

The etiology-associated difference in concordance between MELD and mortality points to the changing epidemiology of liver disease as a major reason for the observed decline. For patients listed with a diagnosis of cholestatic liver diseases such as biliary and sclerosing cholangitis, the MELD and MELD-Na correspond well with mortality, with c-statistics of at least 0.80. Notably, while HCV as a primary indication for transplant shows a middle-of-the-road correspondence between scores and mortality, overall, HCV positive status is associated with a higher c-statistic (0.766 rather than 0.733). The decrease in HCV positive status likely contributes to the decreasing predictive ability of the MELD score.

In contrast to cholestatic diseases and HCV, the predictive ability of MELD/MELD-Na scores with mortality in patients with NASH or alcoholic cirrhosis was at or below 0.74. The scores' shortcomings in predicting NASH outcomes are likely at least partly associated with the disease's indolent presentation and extensive comorbidities.[18] Prior studies have shown that NASH or NAFLD patients are typically listed at similar MELD-Na scores as those with other etiologies of disease, but their MELD scores are slow to progress when still below 15, above which point liver transplantation becomes much more likely.[19,20] NASH sufferers are more likely to die in the first 90 days after waitlisting or be removed from the waitlist for being too sick than either HCV or ALD patients.[19,20] Fatty liver disease is a multisystem disease, not easily graded by liver parameters alone, and far more likely to co-occur with (and contribute to) hypertension, diabetes, chronic kidney disease, and other conditions that increase the risk of waitlist death and decrease the likelihood that a transplant center will offer a waitlisted individual an organ.[21,22]

Conversely, the lower concordance between MELD and ALD is likely related to overall better waitlist outcomes than predicted.[19] Alcoholic liver disease typically presents at a more advanced stage of liver dysfunction, with a more rapid rise in MELD and fewer comorbidities.[19,23] This places patients listed with ALD in a position of greater priority for transplantation, and as their liver function is the driver of their illness with less chronic renal and cardiovascular dysfunction, they are more likely to survive to transplantation.

In the context of disparities between specific liver diseases, fatty and alcoholic liver diseases are on the rise – meteorically in the case of non-alcoholic fatty liver disease (NAFLD), although current projections still suggest ALD will be the most common cause of end stage liver disease in coming years.[24,25] Our data show that ALD already accounted for 1 in 4 listings as of 2017. Unfortunately, these are the diseases for which the MELD and its successor, the MELD-Na, are less consistently predictive of mortality. The change in indications for wait listing is likely part of the reason for the declining accuracy of the MELD score to date. The projected increase in incidence of these particular diseases suggests that the MELD-Na will continue to decline in the future.

This study's findings suggest that a new, or at least improved, MELD may be needed in the near future if it is to continue to be used as a tool to represent mortality risk in liver disease. Unfortunately, due to the limited clinical data available through the OPTN database, it was beyond the scope of this study to determine definitively what is driving this declining correlation with mortality, but continuing research on well-documented limitations of the MELD-Na will likely be key to both answering this question and improving allocation. Evaluating the individual parameters of MELD-Na over time is one potential next step that could help identify and improve elements with decreasing accuracy.

Many have concluded that creatinine is a particularly "weak link" in the MELD, with limitations in both detecting and tracking acute kidney injury and kidney disease secondary to cirrhosis, particularly when used without accounting for patients' baseline values; this is particularly pronounced in female patients and may be responsible for previously observed reduced rates of liver transplantation in women.[26–29] Aside from the value of adjusting the MELD-Na to make it more gender neutral, using a more robust metric of kidney function would have utility in the NASH population, as kidney disease is even more prevalent in that cohort. Replacing creatinine with a more gender-neutral and/or more robust measure of glomerular filtration rate (GFR) has been proposed by numerous researchers, with several suggestions offered.[30–32] Members of our research group continue to explore the replacement of creatinine in the current MELD with cystatin C or an estimated GFR using creatinine and cystatin C.[33,34] No large studies have yet been performed to better evaluate what the impact of integrating one of these metrics into the MELD would be, which remains a gap in current knowledge.

A number of studies have evaluated additional markers and metabolites unrelated to the existing model that could be prognostic of severity or diagnostic of etiology of liver disease. Metabolomics studies, while only conducted in small cohorts to date, have showed that plasma metabolites including tyrosine may have high discriminatory value in predicting short-term mortality in both alcoholic hepatitis and decompensated cirrhosis.[35,36] C-reactive protein, frailty indices, and even cardiac function markers have also shown promise in this area.[37] Given the rising prevalence of NASH, some of these markers of more systemic dysfunction might be valuable in gauging the overall health of a listed individual.

Beyond individual parameter studies, emerging methodologies that may lead the way in improving allocation systems include machine-learning-supported model building and testing. Numerous studies have endeavored to replace the MELD with other models and have fallen short; however, advances in technology and computing power are allowing for more profound investigation of factors affecting patient survival. One promising study used artificial intelligence self-learning algorithms on a large sample data set of transplant patients to identify the best predictors of mortality, and developed the Optimized Prediction of Mortality model (OPOM), which uses components of MELD, age, time on the waitlist, and rate of illness progression as modeled by change in metabolites over time.[27] The OPOM study found that the MELD and MELD-Na are less able to predict death within 90 days in those patients with greater acuity of illness and found that their OPOM model was more effective in those patients, suggesting that a similar model might be better suited for today's increasingly sick, aging waitlist population.

While the chief issue brought up by these findings is the future of the MELD, the changing population over time also raises important questions. The increase in the MELD at listing from 18.4 to 22.7 from 2002 to 2016 indicates that the overall severity of liver disease at listing has increased (excluding the spring 2016 transition to the sodium MELD). This finding quantifies an observation that has been previously made by experts in the field that patients are growing sicker at listing.[38] Not only is the population being listed for transplant sicker (at least in terms of bilirubin, creatinine, and international normalized ratio), it is also increasing in mean age, having increased from 50.1 years old at listing in 2002 to 53.9 years in 2016. With an older, sicker population being listed for transplant, the question of when and for whom liver transplant would be an inappropriate therapy is increasingly vital. Given the as-yet limited prospects for alternative hepatic replacement therapy, determining how to best utilize the limited resource of cadaveric grafts in increasingly at-risk patients will be a critical policy question. It may be that the increasing prognostic uncertainty of the MELD and MELD-Na is primarily due to a combination of the increasing comorbidities and frailty of individuals on the waitlist over time, and the advances in life support technology and practice that allow critically ill individuals to remain on the waitlist longer. Given this climate, we must establish methods to ensure our allocation system remains responsive to the continued need to optimize the allocation of liver allografts to those waitlisted patients who are both most severely ill and most able to benefit from transplantation.

In conclusion, while MELD remains predictive of mortality, its predictive accuracy is declining. It corresponds best with viral and cholestatic indications, rather than the fastest-growing indications: NAFLD and ALD. Patients being listed for liver transplant are presenting an increasing challenge to manage; they are sicker with higher MELD scores and are typically older, as well as having an increasing proportion of NAFLD and ALD, the latter of which accounted for one in four listings in 2017 and has no effective medical therapies available.[7] Finally, viral cirrhosis has been decreasing as an indication for transplantation, and the relative proportion of listings prefigures an even faster decline in the coming years. These conclusions, while suggestive, are preliminary, and extensive further study needs to be conducted to determine if the MELD is the best model to continue to use to prioritize the sickest patients, or if allocation should follow other guiding models given the changing population with liver disease, and how the MELD and allocation as a whole can be revolutionized using improved computational resources and advances to maximize their effectiveness.