PROGNOSIS IN TRAUMATIC BRAIN INJURY (TBI)

Abstract

Introduction: Prognosis in Traumatic Brain Injury (TBI) can be made using prognostic models (the IMPACT and CRASH models) or brain injury biomarkers (S100B). Current prognostic models are derived from historic datasets recruited from heterogeneous countries in terms of trauma care and for the purpose of clinical trials. Objective: To construct a prognostic model suitable for British trauma care, to compare the prognostic performance of prognostic models with S100B and to assess the combination of prognosticators from the constructed models with S100B. Methods: A dataset of 802 TBI cases from the Trauma Audit and Research Network (TARN), Manchester, UK was used to construct the prognostic models.. During the modelling, criteria for well-developed models as per the literature review were followed such as the dataset being large, the variables being selected from the literature and missing information being imputed. A further dataset of TBI cases was used to validate these models Moreover, the resulting models were run on a dataset of 100 TBI cases who had their serum S100B recorded at 24 hours to compare their performance with S100B. Results: Two prognostic models were constructed (models A and B) to predict the discharge survival. Both models share age, admission Glasgow Coma Scale (GCS), admission pupillary reactivity and presence/absence of hypoxia and lowblood pressure (on admission) and brain stem injury. However, model A includes Injury Severity Score (ISS) which is replaced with cause of injury, extracranial injury, brain swelling and interaction of cause of injury and age in model B. Both models have high performance either on the derivation dataset (Area Under the ROC Curve (AUC) of model A: 0.92 and AUC of model B: 0.93) or the external validation set from a later time period in TARN (AUC of model A: 0.92 and AUC of model B: 0.82). Furthermore, in the S100B dataset, it appears that the performance of prognostic models is not significantly different to that of S100B (for example, AUC of model A in this dataset: 0.64 versus 0.69 of the model just including S100B for survival prediction). A combination of S100B and models prognosticators improved performance and S100 improved the performance of models A and B. Discussion: The proposed prognostic models have very high AUCs and since they have been validated on a different TBI dataset from TARN, they are valid to be used for the purpose of the British trauma care benchmarking. Unfortunately, the results of the analysis on the small S100B dataset are not adequately powerful to be conclusive. However, these findings highlight the importance of future research on this topic in larger datasets. Conclusion: Two prognostic models have been constructed which can be used for the British TBI patients.

Date of Award	1 Aug 2011
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Fiona Lecky (Supervisor), Charmaine Childs (Supervisor) & Sarah O'Brien (Supervisor)