External validation of eight different models to predict sepsis mortality in intensive care units

Abstract

Sepsis is a complex, heterogenous syndrome defined as life-threatening organ dysfunction due to severe infection. Existing mortality prediction models may not adequately capture the complexities of sepsis. The objectives of this study were twofold; to clarify to what extent variables belonging to eight different mortality prediction models used in intensive care units (ICU) were collected in routine medical care, and to externally validate these models. A retrospective cohort of 750 patients admitted to three ICU's with a final diagnosis of sepsis at ICU discharge were included. Mortality prediction models were evaluated by calculating the area under receiver operating curve (AUROC) for their ability to predict 30-day mortality. The CSM-4, when used 4 h after ICU admission, predicted ICU episode-of-care mortality best with an AUROC of 0.80. It used only a few variables which are frequently retrieved in routine medical care. ANZROD 24 was the best performing model to be applied 24 h after admission with AUROC of 0.83. Time after admission may decide which prediction model is most useful. Early after ICU admission, the sepsis-specific CSM-4 mortality prediction model performed slightly better than other models. However, at 24 h after admission general models not specific for sepsis, like the ANZROD 24, performed well.