Factors related to outcome in patients with traumatic brain injury : With special reference to analysis with machine learning methods

Sammanfattning: Traumatic brain injury (TBI) is a leading cause of death in trauma. Neurointensive care (NICU) units have emerged to optimize treatments of the injured brain. The main objective is to identify and avoid conditions that lead to further insult and potential cell death. Treatments are based on multimodal monitoring, which includes imaging techniques and monitoring of physiological and biochemical processes, from whole brain to the cellular level. The relationships between the noisy, complex and non-linear data from TBI patients are not fully understood, and may require special methods of analysis. Machine learning (ML) is a group of computer based analysis methods used in data mining and pattern recognition that may be suitable for this purpose. The studies in this thesis were performed in order to analyze patterns of information from cerebral microdialysis (MD), computed tomography (CT) scans and central hematological parameters in patients with TBI, and to investigate their relationships towards parameters known to relate to outcome and to outcome itself. MD is a method to monitor molecules in the extracellular fluid. It has been advocated primarily as a sensitive monitor of ischemic events. Our results (study I and II) indicated that the dominant patterns of MD in TBI patients were long term and therefore delineated individuals. This included patterns that have primarily been interpreted as an ischemic response. In addition, correlations to the pressure and surrogate flow variables, intracranial and cerebral perfusion pressure, were weak within their constraints in the NICU. Together, the results suggest that much of the data that indicates highly perturbed metabolism may have other causes than pressure/flow-related tissue hypoxia. This complies with the emerging awareness that MD, based on commonly used markers, may not be specific for ischemia. CT scans were reviewed following an extensive protocol in patients with severe-to-mild TBI (study III). Midline shift as a continuous variable was found to be the best predictor of favorable/unfavorable outcome. A score using locations and grading of traumatic subarachnoidal blood retained more information than conventional thickness grading. Analyses identified a set of few variables that retain most of the information content. These variables were found to be better predictors of favorable/unfavorable outcome than the reweighted variables of the Marshall classification. Interdependencies limited the contribution of CTs to at best 6 % added estimated explained variance in excess of clinical variables, age, Glasgow Coma Scale (GCS) and pupil responses. A CT-score is suggested, but will require external validation and adjustment. A simple rule-of-thumb predicting TBI outcome is given As with CT variables, complex interdependencies limited the additional contribution of laboratory variables towards prediction, adding 1-2 % in severe-to-mild TBI (study IV). Fifteen of eighteen parameters were correlated to favorable/unfavorable outcome in univariate analyses, but few remained so after adjusting for clinical variables and CT-score. Four variables (creatinine, glucose, osmolarity and albumin) were then identified as independent predictors, although their levels may be related to unmeasured parameters such as premorbidity, trauma severity and resuscitation treatments. The observation that higher osmolarity levels were related to worse outcome, even at values within normal ranges, may warrant further investigation. Notably hemoglobin levels were not related to favorable/unfavorable outcome adjusting with a post-resuscitation GCS vs. an admission GCS. Finally, ML methods proved powerful, despite that their non-linear capabilities did not translate to substantial gains in predictions when compared to conventional regression methods. This could reflect the nature of the investigated problems and the methods themselves. In fact, selection of study populations by levels and choices of GCS will have much larger influence on results. This indicates a need for standardized and well-defined study populations. In contrast to regression analyses, the use of ML for mapping complex data to visual representations of time-series data is most intriguing, and could hold promise for complex clinical monitoring situations.