M. (Marjolein) Haveman



 Marjolein Edine Haveman

Aim – Reliable prediction of neurological outcome in patients with moderate to severe traumatic brain injury (TBI) at the Intensive Care Unit (ICU) could be useful in informing relatives about the likely outcome of the patient, in clinical decision-making, and for optimal treatment. We aimed to develop a multifactorial model combining quantitative electroencephalography (qEEG) measurements and clinically relevant parameters for prognostication of the individual patient with moderate to severe TBI.

Methods – We included 38 patients with moderate to severe TBI admitted to the ICU of the Medisch Spectrum Twente between 2013 and 2016. Patient outcome was dichotomized based on the Extended Glasgow Outcome Score (GOSE) at one-year follow-up into poor (GOSE 1-2) and good outcome (GOSE 3-8). For visual analysis, we selected 23 qEEG parameters from 19-channel EEG registrations, 4 clinical- and 4 medication parameters over the first week after TBI. Prediction models were created using a Random Forest classifier based on the selected features of 38 patients in the training set at 24, 48, 72 and 96 hours after TBI and combinations of two time intervals. After optimization, we evaluated the ability of the models to predict poor outcome, in terms of area under the operating characteristic curve (AUC), on an external validation set of 19 patients admitted to the ICU of the Medisch Spectrum Twente between 2017 and 2018. We added parameters from the International Mission for Prognosis And Clinical Trial Design (IMPACT) predictor, existing of core, computed tomography and laboratory parameters at admission. Furthermore, we compared the predictive ability of our models to the online IMPACT predictor independently.

Results – Our best models existed of 8 qEEG parameters, mean arterial blood pressure (MAP), age and 9 other IMPACT parameters. These models were able to predict poor neurological outcome after TBI in both the internal validation of the training set (AUC=0.92-0.95, sensitivity 0.75-0.92, specificity 0.88-1.00) and the validation set (AUC=0.79-0.81, sensitivity 0.86-1.00, specificity 0.75). The most relevant features in these models were: amplitude of the EEG, MAP, age and glucose level at admission. In particular, models at 72 hours after TBI (alone or combined) are able to predict poor outcome (GOSE 1-2) in patients after TBI, with a higher sensitivity and specificity than the online IMPACT predictor alone in both the training and the validation set (AUC IMPACT of 0.74 and 0.84 respectively).

Conclusion – A Random Forest model based on qEEG features and MAP at 72 hours after TBI, age and IMPACT parameters has a high predictive ability for poor neurological outcome in moderate to severe TBI patients. This is very promising considering the relatively small training set within an in nature heterogeneous population and the exploratory design of this study. QEEG monitoring should play a role in the ongoing challenge for better decision support tools in moderate to severe TBI patients at the ICU.