Abstract
BACKGROUND: Quality benchmarks are increasingly being used to compare the delivery of healthcare, and may affect reimbursement in the future. The University Health Consortium (UHC) expected probability of mortality (EPM) is one such quality benchmark. Although the UHC EPM is used to compare quality across UHC members, it has not been prospectively validated in the critically ill. We aimed to define the performance characteristics of the UHC EPM in the critically ill and compare its ability to predict mortality with the Mortality Prediction Model III (MPM-III). METHODS: The first 100 consecutive adult patients discharged from the hospital (including deaths) each quarter from January 1, 2009 until September 30, 2011 that had an intensive care unit (ICU) stay were included. We assessed model discrimination, calibration, and overall performance, and compared the two models using Bland-Altman plots. RESULTS: Eight hundred ninety-one patients were included. Both the UHC EPM and the MPM-III had excellent performance (Brier score 0.05 and 0.06, respectively). The area under the curve was good for both models (UHC 0.90, MPM-III 0.87, p = 0.28). Goodness of fit was statistically significant for both models (UHC p = 0.002, MPM-III p = 0.0003), but improved with logit transformation (UHC p = 0.41; MPM-III p = 0.07). The Bland-Altman plot showed good agreement at extremes of mortality, but agreement diverged as mortality approached 50 %. CONCLUSIONS: The UHC EPM exhibited excellent overall performance, calibration, and discrimination, and performed similarly to the MPM-III. Correlation between the two models was poor due to divergence when mortality was maximally uncertain.