Current Approaches Fall Short in Capturing the Nuances of Traumatic Brain Injury Patient Care
The current methodologies employed to assess the evolving condition of patients with traumatic brain injuries (TBI) within the confines of the intensive care unit (ICU) lack the ability to encompass the crucial contextual factors required for tailoring personalized treatment. In this context, we have undertaken the task of amalgamating the extensive and diverse array of data stored within medical records – totaling 1166 variables spanning both pre-ICU and ICU periods. Our aim is to construct a comprehensive model that delineates the distinct impact of the clinical trajectory on the 6-month functional prognosis, measured through the Glasgow Outcome Scale-Extended (GOSE).
Through meticulous analysis of a prospective cohort consisting of 1550 TBI patients spread across 65 medical centers, we have harnessed the power of recurrent neural network models. These models are designed to transform the time series representation of all variables – even in the presence of missing data – into an ordinal GOSE projection at 2-hour intervals. This intricate interplay of variables, encompassing the entire spectrum of parameters, elucidates a remarkable explanatory capacity of up to 52% (95% CI: 50–54%) in relation to the ordinal variance observed in the functional outcome. Notably, a substantial portion – reaching up to 91% (95% CI: 90–91%) – of this explanatory prowess emanates from the pre-ICU and admission data, representing the static aspects of patient information.
The dynamic facet of patient data collected within the ICU – referred to as dynamic variables – contributes an additional 5% (95% CI: 4–6%) to the overall explanation. However, this augmentation is not sufficient to offset the relatively diminished performance observed among patients with longer ICU stays (exceeding 5.75 days). Among the pivotal contributors to this model are prognoses articulated by attending physicians, distinctive features identified through CT scans, and indicators reflecting neurological function.
While it is evident that static data presently accounts for the lion’s share of explanatory power in the context of functional outcomes post-TBI, our data-driven scrutiny uncovers avenues for further exploration in characterizing the dynamics of extended ICU stays. Moreover, our innovative modeling strategy holds promise in converting voluminous patient records into coherent time series representations, effectively handling the intricacies of missing data while requiring minimal processing.