Serial versus updated MELD for conditional 28-day mortality in critically ill patients with cirrhosis and sepsis: a landmark cohort study

Whether an admission-window Model for End-stage Liver Disease (MELD) score adds predictive information once an updated score is known remains uncertain in critically ill patients with cirrhosis and sepsis. We studied adults in MIMIC-IV v3.1 who were alive and hospitalized 72 h after ICU admission with observed classic MELD in D0 [0,24 h) and D2 [48,72 h). C1 contained shared baseline covariates and D2 MELD; C2 added D0 MELD. Latest-time laboratory ties were summarized by their median. Patient-grouped five-fold cross-validation was repeated 100 times with five training-fold completed datasets. Confidence intervals used 1,000 conditional patient-cluster bootstrap samples, with 200 full-pipeline samples for the primary comparison. A discharge-censored sensitivity used inverse-probability-of-censoring weighting (IPCW). The primary population comprised 1,136 ICU episodes from 1,007 patients, with 348 recorded linked deaths through ICU day 28. C-indices were 0.7673 for C1 and 0.7674 for C2; Brier scores were 0.1613 and 0.1614. The paired C-index difference was 0.0001 (95% CI − 0.0016 to 0.0019), and Brier reduction was − 0.0001 (− 0.0010 to 0.0007). Full-pipeline intervals were − 0.0045 to 0.0020 and − 0.0022 to 0.0011, respectively. D0 and D2 MELD were highly correlated (Spearman 0.9101); 15.23% of episodes changed by at least five points. Flexible models did not show better predictive performance. The discharge-censored comparison gave a C-index difference of − 0.0008 (− 0.0064 to 0.0044) and Brier reduction of 0.0005 (− 0.0031 to 0.0038). Adding D0 classic MELD to baseline covariates and D2 MELD produced small estimated changes in internally validated prediction in this selected population with largely stable MELD. These findings do not establish equivalence or justify reducing measurements. Secondary comparisons require interpretation within their measurement-defined populations.