OP72 Prospective External Validation of Risk Prediction Models for Acute Traumatic Brain Injury in UK Critical Care Units: The Rain Study

Abstract

INTRODUCTION. Acute traumatic brain injury (TBI) is the leading cause of death and disability in adults aged under 40 years. Statistical models have been developed to predict the risk of mortality or unfavourable outcome (death or severe disability) at 6 months following acute TBI but to date these risk prediction models have only been validated using existing data sources. OBJECTIVES. The Risk Adjustment In Neurocritical care (RAIN) Study aimed to prospectively validate risk prediction models for acute TBI among adult patients admitted to UK critical care units. METHODS. Ten risk prediction models were identified: four for mortality at six months (the Hukkelhoven model and IMPACT Core, Extended and Lab models); and six for unfavourable outcome at 6 months (as mortality plus CRASH Basic and CT models). Risk factor data were collected from 67 UK critical care units (including 90 % of regional neuroscience centres) from August 2009 to March 2011. Patients were followed up to 6 months for mortality, by linkage with death registration, and unfavourable outcome (death or severe disability on the extended Glasgow Outcome Scale, administered by postal or telephone questionnaire). The risk prediction models were validated for calibration (c index), discrimination (Hosmer-Lemeshow test and Cox calibration regression) and overall fit (Brier score). Missing data were handled with multiple imputation. RESULTS. Data were collected for 2,975 eligible patients admitted to critical care following acute TBI. 97 % of patients were followed-up for mortality and 81 % for unfavourable outcome at 6 months. Following multiple imputation, mortality and unfavourable outcome at 6 months were 26 % and 57 %, respectively. All risk prediction models had good discrimination for mortality at 6 months (c index 0.75-0.78). The Hukkelhoven and IMPACT Lab models were well calibrated, but the IMPACT Core and Extended models over-predicted mortality. All models for unfavourable outcome at 6 months had worse discrimination (c index 0.69-0.71) and they substantially under-predicted risk of unfavourable outcome. The best performance overall was found for the IMPACT Lab model, which was the most complex model, incorporating laboratory measurements. Models of the next level of complexity (Hukkelhoven, CRASH CT, IMPACT Extended) all performed similarly. CONCLUSIONS. Risk prediction models for acute TBI had acceptable discrimination among a large, representative sample of patients admitted to UK critical care units. Calibration was good with respect to mortality but poor for unfavourable outcome, and these models therefore require recalibration for use in this setting.