Novel models for the prediction of mortality after traumatic brain injury requiring intensive care

Introduction

Major head injury is a common reason for admission to the ICU. Knowledge of factors that predict mortality provides clues to the pathophysiology of head injury, how clinicians' interventions can be most effective, allows audit between different units or time points and provides objective data with which to communicate with patients' relatives. Several established risk prediction models exist in the ICU; however, they have been shown to have suboptimal discrimination and calibration in this patient group [1]. Our aim was therefore to develop a novel model to predict mortality specifically for head injury.

Methods

A literature review was undertaken to identify variables predictive for mortality after severe head injury. The ICNARC Case Mix Programme, containing multiple data from 374,594 admissions to 171 critical care units in England, Wales and Northern Ireland from 1995 to 2005, was searched for head injury patients with a primary diagnosis of 'primary brain injury', 'subdural haematoma, or 'extradural haematoma'. Each variable that could be supported by the database was entered into a stepwise logistic regression model with mortality as the outcome. Calibration of the risk prediction model was assessed by the area under the receiver operating characteristic curve, discrimination by the Hosmer–Lemeshow C statistic and overall fit by Brier's score.

Results

A total of 10,937 admissions with head injury were identified. A prediction model was constructed using 14 variables and shown to have a superior discrimination and calibration to APACHE II, SAPS II and MPM II. A simplified model consisting of only three variables also performed better than existing models.

Conclusion

We present two novel prediction models for mortality after head injury requiring intensive care. Both models, even the simplified model of only three variables, had superior discrimination and calibration to existing ICU risk-prediction models.