Enhancing short-term predictions of mortality risk for liver transplant candidates through random survival forests.

MELD-based scores do not fully account for systemic and metabolic risk in liver transplant candidates, with notable gaps for those with metabolic dysfunction-associated steatotic liver disease (MASLD). In this retrospective, outcome-anchored, single-center proof-of-concept study, we analyzed adult patients listed for liver transplantation at our institute from 2004 to 2025. We analyzed clinical and laboratory data collected 3-12 months prior to death, transplantation, or censoring, focusing on advanced disease stages. Exclusion criteria included hepatocellular carcinoma and incomplete MELD 3.0 data. We developed a competing-risk random survival forest (RSF) model using nine metabolic and systemic variables, and compared its performance to MELD-Na and MELD 3.0 via time-dependent discrimination and calibration metrics. Among 434 candidates (median age 56.8 years, 60.4% male), 43.3% died on the waitlist, 45.6% underwent transplantation, and 11.1% were alive at study end. The RSF model showed numerically higher discrimination for 180-day waitlist mortality (AUC = 0.81) versus MELD 3.0 and MELD-Na (AUC = 0.75 each), but this exploratory improvement did not demonstrate statistically significant superiority in pairwise AUC comparison. Albumin, diabetes mellitus, and hemoglobin were the strongest predictors. Exploratory reclassification analyses demonstrated a positive Net Reclassification Index (NRI = 0.30), driven primarily by improved identification of patients who died on the waiting list, with minimal misclassification among survivors. Exploratory subgroup analyses showed numerically higher discrimination among candidates with MASLD. These findings support further evaluation of the RSF model as a research framework for outcome-anchored 180-day waitlist mortality assessment. Prospective, multicenter, and external validation will be required before any consideration of clinical use.