Background: A recent prognostic model, predicting 60-day case fatality in critically ill patients requiring renal replacement therapy (RRT), has been developed (Acute Renal Failure Trial Network [ATN] study). Because many prognostic models are suggested in literature, but just a few have found its way into clinical practice, we aimed to externally validate this prediction model in an independent cohort.
Methods: A total of 1,053 critically ill patients requiring RRT from the MIMIC-III database were analyzed. The models' discrimination was evaluated using c-statistics. Calibration was evaluated by Hosmer-Lemeshow (H-L) test and GiViTi calibration belt.
Results: In a case-mix population, including patients with normal or altered serum creatinine (sCr) at intensive care unit admission, discrimination was moderate, with a c-statistic of 0.71 in the nonintegerized risk model. In patients with altered baseline sCr, better discrimination was achieved with the integer risk model (0.76, 95% confidence interval, 0.71-0.81). As for the calibration, although the H-L test was good only in patients with normal/slightly altered sCr at admission, the calibration belt disclosed no significant deviations from the bisector line for any of the models in patients, regardless of admission sCr. Of note, a refitted model had a c-statistics of 0.85, similar to the derivation cohort.
Conclusions: The ATN prognostic model can be useful in a broad cohort of critically ill patients. Although it showed only moderate discrimination capacity when patients with elevated admission sCr were included, using a refitted model improved it, illustrating the need for continuous external validation and updating of prognostic models over time before their implementation in clinical practice.