ACLF-specific prediction models show better discrimination than general models for mortality risk assessmentSpecific prediction models show better accuracy for liver failure patients

BackgroundAcute-on-chronic liver failure (ACLF) is a severe syndrome with rapid progression and high short-to-medium-term mortality. Accurate prognostic risk stratification is essential for guiding clinical decisions and optimizing treatment. While numerous prediction models for ACLF have been developed, their performance and clinical applicability remain unclear, warranting a systematic evaluation to guide evidence-based model selection.MethodsPubMed, Web of Science, Cochrane Library, and Embase were systematically searched from database inception to December 31, 2025. Data extraction and methodological assessment were conducted using the CHARMS. Risk of bias was evaluated using the PROBAST. A meta-analysis of c-statistics was performed using R software (version 4.4.2).ResultsA total of 9,447 studies were identified, with 185 ultimately included, covering 241 external validation cohorts evaluating 75 distinct prognostic models. Approximately 99.51% of analysis units were judged to have a high risk of bias, primarily due to insufficient numbers of outcome events, failure to account for data complexity, and inappropriate assessment of model performance. A total of 24 models met the criteria for meta-analysis at least at one time point, with c-statistics ranging from 0.58 to 0.84. Overall, model discrimination declined with longer prediction horizons, increasing the estimation uncertainty. ACLF-specific models (e.g., CLIF-C ACLF, COSSH ACLF, COSSH ACLF II) showed relatively better discrimination than general models. Among these, the CLIF-C ACLF showed a certain degree of stability across subgroups, though further validation is needed.ConclusionThe overall risk of bias in the included external validation studies was high, with most lacking calibration reports. Therefore, the current evidence primarily supports relative comparisons of model discrimination, but is insufficient to justify their use as precise probability-based tools for direct clinical decision-making. Most prediction models demonstrated moderate to good discrimination, though their performance declined with longer prediction horizons. COSSH ACLF, COSSH ACLF II, and CLIF-C ACLF showed relatively better discrimination than general models in the available evidence, though this advantage needs further confirmation in higher-quality studies. Future research should focus on well-designed, multicenter validation studies, with systematic evaluation of calibration and long-term predictive performance, to further strengthen the evidence base for ACLF prognostic prediction models.