David Bark: Refining outcome prediction in traumatic brain injury

Abstract

Traumatic brain injury (TBI) remains a leading cause of mortality and a major contributor to global disability-adjusted life years in young and middle-aged adults. While animal studies have shown potential, effective pharmacological treatments have failed to materialize in the clinical setting. A primary challenge in TBI research is the inherent heterogeneity of the condition.

Outcome prediction models can reduce this complexity by better risk stratification for clinical trials, optimizing resource allocation, and improving prognostic communication with patients and families. TBI outcomes are commonly measured by the Glasgow Outcome Scale Extended (GOSE), an 8-point scale ranging from death (1) to upper good recovery (8). However, current prediction models typically dichotomize this scale into favorable/unfavorable outcomes which limits nuanced prognostication.

The primary objective of this thesis was to develop prediction models capable of predicting the full 8-grade GOSE. A secondary objective was to evaluate the integration of machine learning in clinical care, specifically assessing an AI decision support system for detecting intracranial hemorrhage in computed tomography (CT) scans.

We found that advanced machine learning methods performed comparably to standard statistical models when limited to admission variables. However, the incorporation of dynamic physiological data captured prognostic signals that static admission models missed, thereby improving prediction accuracy.

Regarding the AI diagnostic tool, while it successfully identified most hemorrhages, we observed no significant clinical benefit for the patients. This underscores that in clinical settings, the implementation strategy is as critical as the technology itself.

Finally, this thesis emphasizes that predictive models are not static tools. They are sensitive to temporal changes in populations and healthcare protocols. Therefore, future implementation must prioritize the continuous updating and validation of these models.