Development and validation of the critical care outcome prediction equation, version 4.

OBJECTIVE

To revise and validate the accuracy of the critical care outcome prediction equation (COPE) model, version 4.

DESIGN, PARTICIPANTS AND SETTING: Observational cohort analysis of 214 616 adult consecutive intensive care unit admissions recorded from 23 ICUs over 12 years. Data derived from the Victorian Admitted Episode Database (VAED) were used to identify treatment-independent risk factors consistently associated with hospital mortality. A revised version of the COPE-4 model using a random intercept hierarchical logistic regression model was developed in a sample of 35 878 (16.7%) consecutive ICU separations.

MAIN OUTCOME MEASURES

Accuracy was tested by comparing observed and predicted mortality in the remaining 178 741 (83.3%) records and in 23 institutional cohorts. Stability was assessed using the standardised mortality ratio, Hosmer-Lemeshow H10 statistic, calibration plot and Brier score.

RESULTS

The COPE-4 model had satisfactory overall discrimination with an area under receiver operating characteristic curve of 0.82 for both datasets. The development and validation datasets demonstrated good overall calibration with H10 statistics of 13.38 (P = 0.10) and 14.84 (P = 0.06) and calibration plot slopes of 0.99 and 1.034, respectively. Discrimination was satisfactory in all 23 hospitals and one or more calibration criteria were achieved in 19 hospitals (83%).

CONCLUSIONS

COPE-4 model prediction of hospital mortality for ICU admissions has satisfactory performance for use as a risk-adjustment tool in Victoria. Model refinement may further improve its performance.