Predicting mortality in critically ill patients requiring renal replacement therapy for acute kidney injury in a retrospective single-center study of two cohorts

Abstract

Half of the critically ill patients with renal replacement therapy (RRT) dependent acute kidney injury (AKI) die within one year despite RRT. General intensive care prediction models perform inadequately in AKI. Predictive models for mortality would be an invaluable complementary tool to aid clinical decision making. We aimed to develop and validate new prediction models for intensive care unit (ICU) and hospital mortality customized for patients with RRT dependent AKI in a retrospective single-center study. The models were first developed in a cohort of 471 critically ill patients with continuous RRT (CRRT) and then validated in a cohort of 193 critically ill patients with intermittent hemodialysis (IHD) as the primary modality for RRT. Forty-two risk factors for mortality were examined at ICU admission and CRRT initiation, respectively, in the first univariate models followed by multivariable model development. Receiver operating characteristics curve analyses were conducted to estimate the area under the curve (AUC), to measure discriminative capacity of the models for mortality. AUCs of the respective models ranged between 0.76 and 0.83 in the CRRT model development cohort, thereby showing acceptable to excellent predictive power for the mortality events (ICU mortality and hospital mortality). The models showed acceptable external validity in a validation cohort of IHD patients. In the IHD validation cohort the AUCs of the MALEDICT RRT initiation model were 0.74 and 0.77 for ICU and hospital mortality, respectively. The MALEDICT model shows promise for mortality prediction in critically ill patients with RRT dependent AKI. After further validation, the model might serve as an additional clinical tool for estimating individual mortality risk at the time of RRT initiation.