ABSTRACT
BACKGROUND: Risk prediction models are used in critical care for risk stratification, summarising and communicating risk, supporting clinical decision-making and benchmarking performance. However, they require validation before they can be used with confidence, ideally using independently collected data from a different source to that used to develop the model. The aim of this study was to validate the Intensive Care National Audit & Research Centre (ICNARC) model using independently collected data from critical care units in Scotland. METHODS: Data were extracted from the Scottish Intensive Care Society Audit Group (SICSAG) database for the years 2007 to 2009. Recoding and mapping of variables was performed, as required, to apply the ICNARC model (2009 recalibration) to the SICSAG data using standard computer algorithms. The performance of the ICNARC model was assessed for discrimination, calibration and overall fit and compared with that of the Acute Physiology And Chronic Health Evaluation (APACHE) II model. RESULTS: There were 29,626 admissions to 24 adult, general critical care units in Scotland between 1 January 2007 and 31 December 2009. After exclusions, 23,269 admissions were included in the analysis. The ICNARC model outperformed APACHE II on measures of discrimination (c index 0.848 versus 0.806), calibration (Hosmer-Lemeshow chi-squared statistic 18.8 versus 214) and overall fit (Brier's score 0.140 versus 0.157; Shapiro's R 0.652 versus 0.621). Model performance was consistent across the three years studied. CONCLUSIONS: The ICNARC model performed well when validated in an external population to that in which it was developed, using independently collected data.