Thursday, September 22nd 2011

Predicting Outcome After Traumatic Brain Injury

The Corticosteroid Randomization After Significant Head injury (CRASH) trail was a huge international double blinded randomized trial which collected a huge cohort of patients suffering traumatic brain injuries with GCS less than 14 on presentation and presenting within 8 hours of injury and randomizing them to receiving a 48 hour course of methylprednisolone versus a placebo. The final results were published in The Lancet in 2005.

As if that was not enough the database collected for the study was the largest, most complete database of patients following head injury in the world. It included more than 10,000 patients from across the world and had a very high rate of follow up through 6 months. This database was used to create a prognostic model for outcome following head injury. Published in the British Medical Journal in 2008 the CRASH model has become one of more widely cited outcome prediction models in clinical practice when dealing with patients with head injury.

MRC CRASH Trial Contributors. “Predicting Outcome after Traumatic Brain Injury: Practical Prognostic Models Based on Large Cohort of International Patients.” Bmj 336.7641 (2008): 425-29

The original paper is available for free as full text on the BMJ website.

This was a retrospective review of outcome of a large cohort of patients.

The database included 10,008 patients originally collected for the CRASH Trial. The database contained information on a large number of variables but the prognostic model focused on 9 initial variables: age, sex, etiology of the trauma, time on presentation, GCS on presentation, pupil reactivity on presentation, results of CT scan, whether the patient had a major extracranial injury, level of per capita income in the country where the injury occured.

They prognosed to two outcomes death within 14 days of injury or outcome at 6 months as measured by the Glasgow Outcome Score which they dicotomized into favorable outcomes (moderate disability or good recovery) and unfavorable outcomes (dead, vegetative state or severe disability).

They developed two models with the above variables: a basic model which excluded the findings on CT imaging and a CT model which included them.

Internally they validated the two models using bootstrap resampling. And then they externally validated the model using the 8000+ patients suffering head injury included in the independent International Mission for Prognosis And Clinical Trial (IMPACT) database. The original description of the IMPACT Trial is here on PubMed.

All nine of the variables included in the final two models independently had strong associations with both outcomes (death at 14 days and poor outcome at 6 months). The table showing the odds ratios for each variable can be found here.

Regression of outcome of model including CT scan findings

After this internal validation they compared their model to outcomes observed in the IMPACT Trial blinded.

CRASH model predictions of outcomes for patients enrolled in IMPACT versus variable outcome

It showed good discrimination in the external validation with a C-score of 0.77 (essentially the area under the reciever operator curve). And for the basic model and the CT model in high income countries there was very good calibration by the Hosmer-Lemeshow test.

The authors have published a calculator with predictive outcomes based on the regression. The calculator gives odds for death at 14 days and poor outcomes at 6 months based on both models.

The sample size of this study is legitimate and one of the stronger points of it. The database appears well maintained and exceptionaly complete considering the number of patients and the challenges of coordinating data collection across continents. Looking at the demographics it appears relatively representative. As well the follow up to 6 months is a legitimate end point for the goals of the study and the sample included a great number of patients to that end point.

The outcomes measured to seem clinically relevant and were set prior to the models being designed. The variables included in the regressions as well have previously been validated at predicting outcomes in other smaller studies and are clinically readily available.

The study found no difference in outcomes based on treatment, including the randomization to steroids or placebo in the trial itself, which is an important consideration.

It is true the CT model in particular showed comparatively poor calibration for patients injured in lower income countries. Even there however the Hosmer-Lemeshow measure isn’t particularly off. The smaller sample size of patients with readily available CT imaging findings may in part explain such.

The major critiques otherwise of the study seem to be those available to be leveled at all published prognostic models. Generalization is a difficult thing especially to individual anecdotal scenarios. However for a study with in vivo data from patients seemingly encountered early following representative head injury and undergoing real world salvage attempts the model shows remarkable, if not perfect, calibration and discrimination.

The models developed from the CRASH trial are very likely the “best” available prognostic models for predicting outcome early after traumatic head injury. The calculator published by the authors appears a valid and useful tool for any health care provider encountering significant neurotrauma. A “better” prognostic model for the prediction of outcome early after head injury seems unlikely.