Advancements in therapeutic treatments for liver diseases has led to improved prognosis for decompensated cirrhosis. These improvements have led to the concept of recompensation among candidates on the waitlist for liver transplant. The most updated definition of recompensation is based on Baveno VII consensus, which states recompensation as cure or control of the liver disease, clinical improvement in decompensation symptoms without the need for medications, and restoration of synthetic function of the liver.1 A recent estimate of the prevalence for recompensation is 35% across all etiologies of cirrhosis.2 Breakdown by specific etiologies for liver failure, hepatitis B virus-related cirrhosis had the highest prevalence of recompensation at about 50%, hepatitis C virus cirrhosis at 42%, and alcohol-related cirrhosis at 33%.2 However, the overall rate of delisting from transplant waitlist remains low at 6.2% with hepatitis B virus and alcohol-associated cirrhosis being the most common etiologies to meet criteria for delisting.3,4 There are ongoing efforts to identify a set of criteria to predict which individuals are most likely to be delisted because of clinical improvement. In this retrospective cohort study involving adult candidates on the liver transplant waitlist from 2000 to 2025, Tang et al5 aimed to fill this gap by using machine learning to identify clinical factors associated with successful delisting because of recompensation. Using both random forest modeling and variable importance measures, causes of liver failure (with alcohol-associated liver disease being the most prevalent cause among those who were delisted because of recompensation), candidate age, and serum albumin were identified to be the main predictive factors for delisting because of clinical improvement. Interestingly, Model for End-Stage Liver Disease (MELD) score was not found to have strong association with recompensation but was predictive for death while on waitlist.5 While this finding is interesting, the analysis is limited to only using MELD at the time of waitlisting and MELD changes while on the list may change the overall interpretation. The described model improved prediction at 1 y when compared with Cox modeling. Additionally, their model showed consistency across the era before (pre-2014 era) and after (post-2014) widespread utilization of direct acting antivirals. In fact, their machine learning model was able to predict post-2014 outcomes using pre-2014 data. However, the use of machine learning models such as random survival forests here needs justification for the loss of interpretability that a black box solution comes with. Most standard regression models from statistics come with effect size estimates (eg, slopes in linear regression or hazard ratios in Cox regression) while machine learning models have approximate analogs such as shapely additive explanations parameters or variable importance. These measures help researchers understand how a model result aligns with domain expertise and which variables may be responsible for individual predictions. These measures come with some limitations though. A researcher may know that MELD score at transplant has the highest variable importance in a model, but the actual impact per unit change of that variable relative to others is less clear. Frequently researchers may apply these techniques with little to no improvement in model performance for the sake of novelty even though machine learning methods have been applied to transplant for >30 y.6 In this case, however, the authors show that several variables may be nonlinear. Furthermore, an early improvement in predictive power at 1-y post-listing over a traditional Cox model, which eases equivocation over time, may indicate an issue with proportionality, which machine learning methods may circumvent. An important note is that there are many instances where the benefits of machine learning (ie, nonlinearities, interactions, and nonproportional hazards) can be addressed with more conventional methods. Indeed, in simulated studies and real data Kattan7 showed that Cox models with spline terms can perform just as well as machine learning alternatives. Explainable differences in predictive performance between machine learning models and conventional models may be an indicator of needed changes to model form rather than proving the superiority of machine learning. Additional limitations in this study lead to residual questions for further investigation. In their methodology, candidates with multiple listings were removed from consideration. Furthermore, as mentioned by the authors themselves, lack of granularity in registry-based study may lead to missing important clinical criteria for consideration. Clinical variables such as treatments while on the waitlist, hospital admissions, number of decompensation episodes, and relisting after delisting from the waitlist are potential factors to consider when predicting those who meet criteria for delisting because of clinical improvement. Additionally, while the authors suggest that modeling delisting because of clinical improvement will improve prioritization for liver transplant, it is unclear how such model is substantially different from modeling those who remain alive on the waitlist. Perhaps there is significant overlap and/or differences among these subsets that likely will require more in-depth analysis. Traditionally, the MELD score has been used to prioritize candidates on the liver transplant waitlist. Tang et al5 have shown that the MELD score may be less important when considering the alternative outcome of liver recompensation. By taking a different approach to waitlist outcomes, the authors hint at a new way to prioritize patients. On the other hand, this comes with substantial limitations in both data granularity and analysis methods as stated above. More investigation is needed to determine which set of clinical variables would predict candidates who will achieve recompensation and not require a liver transplant.
Liang et al. (Tue,) studied this question.