Simple Clinical Tools to Predict Decompensation in Patients With Compensated Cirrhosis: An Unmet Need

The natural history of cirrhosis is characterized by a long, asymptomatic stage (compensated cirrhosis), and by a much shorter stage (decompensated cirrhosis) that is defined by the presence of 1 or more complications of portal hypertension, namely ascites, portal hypertensive bleeding, and/or encephalopathy. The main determinant of survival in cirrhosis is the development of decompensation, with liver-related mortality occurring almost exclusively after the development of decompensation.¹

Although in the past the diagnosis of cirrhosis was mainly established at the decompensated (symptomatic) stage, the recent availability of noninvasive methods to evaluate the degree of liver fibrosis in patients with chronic liver disease has allowed making the diagnosis of cirrhosis at the earlier, asymptomatic stage. Determining the probability of decompensation in these patients is important both clinically and therapeutically.

A subanalysis by Ripoll et al² of a randomized clinical trial that included patients with compensated cirrhosis, mostly from hepatitis C virus (HCV), who did not have varices at inclusion in the study demonstrated that clinically significant portal hypertension (CSPH) defined by a hepatic venous pressure gradient ≥10 mm Hg was the strongest predictor of decompensation. The predictive value of CSPH has been validated in other trials including a recent study in patients with compensated cirrhosis from nonalcoholic steatohepatitis (NASH).³

However, measurement of hepatic venous pressure gradient is a nuanced procedure that is not routinely performed in most centers and its accuracy is highly operator-dependent. Recognizing this fact, the study by Ripoll et al² investigated noninvasive predictors of decompensation in the same cohort (compensated without varices) and found that albumin, Model for End-Stage Liver Disease, aspartate aminotransferase, and platelet count independently predicted decompensation. As an example, a nomogram to calculate the 1-, 2-, and 3-year risk of decompensation based on this model is provided in Figure 1. Other studies have shown that markers of liver sinusoidal endothelial dysfunction, such as the von Willebrand factor,⁴ and markers of portal hypertension or liver stiffness measurements⁵ are also predictive of decompensation.

Figure 1.

Graphical representation in the form of a nomogram of the model described by Ripoll et al.² This model was derived by Cox regression in 213 patients with compensated cirrhosis without varices. The upper panel is used to derive the points contributed for each variable. The total number of points is then used in the lower panel to derive the probability of remaining decompensation-free at 1, 2, and 3 years. ALB, albumin; AST, aspartate aminotransferase; MELD, Model for End-Stage Liver Disease; PLT, platelet count

Refining risk prediction in compensated cirrhosis is key to: (1) enable caregivers to have a meaningful discussion regarding prognosis with a patient, (2) facilitate the design of randomized clinical trials for new therapies aimed at preventing decompensation by improving patient selection and providing accurate statistical powering, and (3) estimate the balance of risks and benefits of a new therapy.

The study by Guha and coworkers⁶ in the current issue of Clinical Gastroenterology and Hepatology introduces a new noninvasive model to predict the risk of decompensation in patients with compensated cirrhosis. It is based on the combination of 2 previously identified scores: ALBI (serum albumin and serum bilirubin) and FIB-4 (age, aspartate aminotransferase, alanine aminotransferase, and platelet count). The ALBI score had been shown to predict survival in a cohort of patients with mostly decompensated hepatitis B cirrhosis,⁷ and the FIB-4 score had been shown to predict survival in a cohort of patients with HCV, only a minority of whom had cirrhosis.⁸ The ALBI-FIB4 model was developed in a prospective cohort of patients with compensated cirrhosis from Nottingham (UK) that consisted predominantly of alcohol- and NASH-related cirrhosis. It was validated in 2 cohorts: one prospective from Dublin that encompassed mostly alcohol and HCV-related cirrhosis; and the other, retrospective, from Egypt that consisted mostly of NASH and HCV etiologies.

The study has several positive features. First, the authors built their model with variables that are readily available in any clinical context and avoid the inclusion of such tools as transient elastography that, although probably more precise, are not widely accessible. Second, they provide an easy-to-use online tool that calculates the absolute risk of decompensation for up to 5 years by entering the values of the 6 variables. Third, they investigate the discriminative ability of the model in 2 external cohorts, from different geographic areas and with different etiologies of cirrhosis. This strength is also a weakness because it would have been important to analyze the performance of the model among the different etiologies. The rate of decompensation may differ among the different etiologies, with recent evidence indicating that NASH cirrhosis may progress more rapidly than cirrhosis from other etiologies.³ Also, such factors as antiviral therapy or alcohol abstinence (which are not discussed in the study) would have had a major impact on the rate of decompensation for HCV and alcohol etiologies, respectively. Another factor that could have had an impact on decompensation rates across cohorts is the proportion of patients with evidence of CSPH (eg, varices) at baseline. Because the criterion by which the diagnosis of cirrhosis was established in the different cohorts is not specified in the study, nor is the proportion of patients with varices, it is uncertain whether patients were comparable regarding their likelihood of having CSPH and, therefore, of decompensating. Regarding the outcomes defining decompensation, HE as a decompensating event was defined as grade 3/4, thereby excluding patients who developed grade 2 HE, which is considered an overt decompensating event⁹ that would have probably presented earlier and in a larger number of patients.

The predictive model itself and its validation have important limitations. Instead of “deconstructing” both models from which it is derived into the 6 individual variables, the authors included the 2 original models separately (ALBI and FIB-4). This is an odd choice, because the weight of the different variables within their new model is most probably substantially different from the weight of the same variables within the 2 predefined models.

Most importantly, the external validation of their new proposed ALBI-FIB4 model is incomplete. Although the authors demonstrate a good discrimination capacity of the model (0.80 in the development cohort, 0.776 in the validation cohorts), slightly higher than the Model for End-Stage Liver Disease and Child-Pugh scores, they do not analyze the calibration of the model (ie, the agreement between the predicted and the observed rates of decompensating events). Demonstration of a consistent calibration among the different cohorts would have been the most important evidence toward the usefulness of the model. The lack of a consistent calibration can be gleaned from data shown in their Table 2, where patients are stratified in “low risk” and “high-risk” and where one can see that the 3- and 5-year decompensation rates are substantially different among the cohorts (eg, whereas the 3-year decompensation rate in the UK low-risk cohort is 9.32%, it is 16.15% in the Egyptian cohort). Although this does not prove miscalibration, it does raise concerns regarding the external validity of the model. A formal external validation in terms of calibration is required before this model can be recommended for use in clinical practice or research. Finally, the dichotomization into 2 different risk strata is unnecessary, leads to oversimplification with loss of important granularity, and could be misleading. According to the online calculator, a risk of decompensation up to 24% at 3 years or 38% at 5 years is considered “low risk” when clinically or in clinical studies these rates could be considered high.

In summary, the availability of a simple noninvasive tool to predict decompensation in cirrhosis is still an unmet need. This tool needs to be carefully validated in terms of calibration, because it would be used in clinical practice to predict the absolute risk of decompensation in individual patients. It is very unlikely that such a tool would work for all etiologies of cirrhosis, and would likely need to be etiology-specific, incorporating specific predictive elements for the specific etiology (eg, diabetes for NASH etiology,¹⁰ abstinence for alcohol-related cirrhosis). These models would require updates as new therapies with an impact on the risk of decompensation become available (as has occurred with effective therapies for HCV) and could incorporate new variables (eg, liver stiffness or new serologic tests) as they become widely available. Additionally, as one could foresee that the tool would be used over time in the same patient, the predictive value of serial updates of the model would also require evaluation. Large multicenter collaborative studies, with large sample sizes for each major etiology, are needed to develop and validate robust models that would be useful in both clinical and research settings.