HDCN Article Review/Hyperlink

Douma CE, Redekop WK, Van der Meulen JHP, Van Olden RW, Haeck J, Struijk DG, Krediet RT

Predicting mortality in intensive care patients with acute renal failure treated with dialysis

J Am Soc Nephrol (Jan) 8:111-117 1997

Patients in intensive care units with acute renal failure (ARF) continue to have a high mortality. Prediction models have been attempted to predict mortality in such patients, perhaps with an ultimate goal of limiting initial dialysis therapy when prognosis is hopeless. In practice, prediction models are of more usefulness in clinical studies of ARF when trying to determine if use of modality or membrane, for example, affects mortality, and when correction needs to be made for underlying risk.

The main model used in an ICU setting is the APACHE score, the Acute Physiology and Chronic Health Evaluation system. It consists of two parts: an Acute Physiology Score (APS) based on 12 physiologic measures, and on the presence of chronic health problems. This is used with the reason for ICU to compute an estimated risk of death in a given patient. The APACHE III is the most recent version of this evaluation system. Other variants include a Simplified Acute Physiology Score, which includes selected physiology measures from the APACHE system without the chronic health problem part. There are a number of other variants. There also exist some ARF-specific predictive models based on work in ARF patients.

In this study, a number of these scores were retrospectively applied to 238 patients who received a first dialysis treatment in the intensive care unit. The overall in-hospital mortality for the group was 78%. Receiver operating curves were determined to assess the sensitivity and specificity of each prediction system to predict death.

In general, the models derived from ICU patients as a whole tended to underpredict mortality in the patients with ARF. A model developed by Liano (Nephron 63:21-31, 1993), which was derived in ARF patients, seemed to predict observed mortality rather well. Of the general models, the APACHE III evaluation was the best. However, none of the models had high sensitivity and specificity. Both the APACHE III and Liano model were good at identifying very high risk patients. When patients in the top quintiles with each method were examined, observed mortality rates were 97% and 98%, respectively.

Comment: The fact that all models tested showed poor to moderate discriminatory ability should inject some caution into interpretation of studies of ARF mortality where APACHE scores were used to ensure comparability of groups. In particular, the underestimation of mortality in ARF patients with the APACHE score is noteworthy. Should one consider withholding dialysis in patients within the highest quintiles of APACHE and Liano scores? This will always remain an individual, clinical decision. In any case, cutoff scores for the highest quintiles are not given, and the questionnaires would need to be recalibrated at a given institution.

The techniques of meta-analysis are well-developed. Perhaps one could combine all of these predictive equations into a neural net type of algorithm which might have improved sensitivity and specificity. (John T. Daugirdas, M.D., University of Illinois at Chicago)