Development and validation of a novel mortality prediction score (ICU-MARS) based on routine clinical and laboratory parameters in respiratory intensive care unit patients

This retrospective study aimed to evaluate the association between routinely measured clinical and laboratory parameters and ICU mortality in patients admitted to a Respiratory Intensive Care Unit (RICU), and to develop a novel prognostic score based on these variables. This retrospective, single-center cohort study included 2,417 unique adult patients admitted to the RICU between January 2021 and December 2025. The cohort was divided chronologically into a development cohort ( n = 1,932) and a temporal validation cohort ( n = 485). The primary outcome was ICU mortality, defined as death during the index RICU stay. Multivariable logistic regression was used for model development. Internal validation was performed using 1,000 bootstrap resamples, and model performance was assessed using discrimination, calibration, and the Brier score. Overall ICU mortality was 29.3% (709/2,417). The reduced ICU-MARS model included age, urea, base excess, serum albumin, malignancy, and mechanical ventilation at ICU admission. In the temporal validation cohort, 475 patients were included in the complete-case analysis, and the observed ICU mortality rate was 25.3%. ICU-MARS demonstrated good discrimination, with an AUC of 0.836 (95% CI, 0.797–0.874). The calibration intercept was − 0.237 (95% CI, − 0.480 to 0.007), the calibration slope was 1.008 (95% CI, 0.798–1.218), and the Brier score was 0.140. ICU-MARS showed significantly better discrimination than APACHE II, whereas the difference in discrimination between ICU-MARS and SOFA was not statistically significant. Routinely available clinical and laboratory parameters at ICU admission have significant predictive value for ICU mortality in RICU patients. The ICU-MARS score may provide a practical approach for early mortality risk stratification, demonstrating significantly higher discrimination than APACHE II, while no statistically significant difference in discrimination was observed compared with SOFA. External validation in independent and multicenter cohorts is warranted before broader clinical implementation.