Pneumonia is the leading cause of sepsis. This study aimed to compare the predictive accuracy for 28-day mortality between Sequential Organ Failure Assessment (SOFA)-2 score and SOFA score in pneumonia-associated sepsis, and to develop and externally validate a composite model incorporating SOFA-2 with key clinical variables. This retrospective cohort study analyzed data from MIMIC-IV (n = 7,150) and externally validated using an Intensive Care Unit (ICU) cohort from Shandong Provincial Hospital (n = 301). Adults meeting Sepsis-3 criteria with pneumonia as the infection source were included. The primary outcome was 28-day all-cause mortality. Multivariable Cox regression, time-dependent AUC analysis, and decision curve analysis were performed. The SOFA-2 composite model was developed and evaluated. In the MIMIC‑IV cohort, the 28‑day all-cause mortality was 28.1%. Multivariable analysis indicated that the SOFA‑2 score was an independent predictor of 28‑day mortality. The Area Under the Curve (AUC) curves of SOFA‑2 and SOFA overlapped substantially, with no significant difference in predictive performance (day 1: 0.78 vs. 0.79; day 28: 0.61 vs. 0.62). A composite SOFA‑2 model significantly improved predictive performance (day‑1 AUC = 0.87). MaxStat analysis identified an optimal risk‑stratification cutoff of 11 for SOFA‑2; patients in the high‑score group (> 11) had significantly lower 28‑day survival than those in the low‑score group (p < 0.001). Subgroup analysis demonstrated that the association between SOFA‑2 and prognosis remained consistent across different characteristic subgroups. External validation further confirmed the independent prognostic value and risk‑stratification ability of SOFA‑2. In patients with pneumonia‑associated sepsis, the SOFA‑2 score exhibited a predictive performance for 28‑day mortality risk similar to that of the conventional SOFA score. Although its standalone use did not show superiority, integrating SOFA‑2 with key clinical variables into a composite model significantly enhanced predictive accuracy.
Wei et al. (2026) studied this question.