Internal and Temporal Risk Discrimination in MASLD-Related Cirrhosis Presenting to Emergency Care

Background Risk stratification in MASLD-related cirrhosis presenting to emergency care requires integration of hepatic severity, comorbidity, cardiometabolic burden, and acute presentation. Because AUROC only measures how well a model ranks patients by risk, calibration and temporal validation are needed to assess whether predicted risks remain reliable over time. Methods Among 1,262 MASLD encounters from a matched cohort, we fit six separate multivariable logistic regression models for complete-case MASLD encounters with available prediction features, one for each endpoint: death/ICU, death, ICU admission, MACE, hepatic decompensation, and OHE. For comparability and to avoid endpoint-specific predictor selection, all models used the same prespecified candidate predictors: age, sex, calendar year, MELD-Na, Charlson hepatic score, hepatocellular carcinoma, clinically significant portal hypertension, cardiovascular disease, diabetes, chronic kidney disease, triage, and mSOFA. Coefficients and predicted risks were estimated separately for each endpoint. Internal validation used patient-level cross-validation. Performance was summarized with AUROC and patient-level bootstrap 95% CI, Brier score, calibration intercept, and calibration slope. Temporal validation trained models in 2016–2021 period and tested them in 2022–2024 period. Standard scores, including MELD, MELD-Na, CLIF-AD, Charlson, qSOFA, and mSOFA, were retained as comparators. Results In MASLD encounters, the endpoint-specific logistic models showed the highest cross-validated discrimination for hepatic decompensation (AUROC, 0.864; 95% CI, 0.839 to 0.884; Brier 0.141; calibration slope 0.938) and MACE (AUROC, 0.845; 95% CI, 0.816 to 0.871; Brier 0.125; slope 0.934). OHE prediction was also useful (AUROC, 0.809; 95% CI, 0.777 to 0.839; Brier 0.112; slope 0.819). Death/ICU showed more modest discrimination (AUROC, 0.702; 95% CI, 0.654 to 0.749; Brier 0.105; slope 0.845). Temporal validation preserved useful ranking but showed weaker calibration: AUROC 0.834 for hepatic decompensation, 0.789 for MACE, 0.783 for OHE, and 0.685 for death/ICU, with slopes between 0.59 and 0.77 for several endpoints. MACE prediction depended substantially on cardiovascular history: AUROC was 0.845 with cardiovascular disease included and 0.731 without it. Single liver-score benchmarks approached the multivariable death/ICU model but did not capture cardiometabolic outcomes as well. Conclusions Six endpoint-specific multivariable logistic models showed useful internal discrimination for MACE, OHE, and hepatic decompensation in MASLD-related cirrhosis. However, temporal calibration was imperfect. These models should be interpreted as risk-stratification tools requiring recalibration and external validation before clinical implementation.