Objective:

To evaluate predictive performance of quantitative electroencephalogram (qEEG) features in predicting outcomes among moderate-severe traumatic brain injury (msTBI) patients.

Background:

EEG is utilized among critically ill TBI patients for detection of seizure and non-convulsive status epilepticus; however, role of qEEG in prognostication has not been thoroughly explored.

Design/Methods:

We conducted a retrospective cohort study using the University of California, Irvine TBI and Concussion (NTBIC) database (10/2023-3/2024). Patients with msTBI (GCS <8, requiring neuromonitoring, complex TBI cases per surgical ICU) were included. Poor outcome was defined as discharge Modified Ranking Scale (mRS) ≥5. For select EEG electrode contacts, 5-minute EEG segments at every 4-hour intervals were extracted from first 24 hours of EEG data. Prognostic capacity of qEEG features were assessed via univariate analysis (t-test/chi-square), multivariable logistic-regression, and mean area under the receiver operating characteristic curve (AUCROC) over a fivefold cross validation. Analysis was conducted using MATLAB/Python.

Results:

47 patients (29.8% female, 55.3% non-white) with a mean age of 49.0±22.3 were included. 1 patient had an electrographic seizure. Mean amplitude, delta, beta, and total-power were higher (p<0.05), and there were no differences in use of sedation between patients with good vs. poor outcomes (p>0.05). On multivariate logistic regression, after controlling for age, sex, race, admission GCS, dilated pupil on arrival, and use of any sedation, beta (OR=0.01, 95% CI 0.00-0.54, p=0.023), delta (OR=0.81, 95% CI 0.69-0.96, p=0.013), and total-power (OR=1.16, 95% CI 1.03-1.31, p=0.018) predicted poor outcome. AUCROC using age, admission GCS, and pupils was 0.72±0.17. Adding beta, total-power, and delta, yielded AUCROC of 0.69±0.17, 0.72±0.15, 0.74±0.17 respectively. Combination models of delta/total-power and beta/delta/total-power yielded 0.75±0.18 and 0.77±0.14, respectively.

Conclusions:

qEEG features, notably beta/delta power and total-power, have satisfactory predictive value of ICU-TBI outcome when added to clinical features. Future studies will prospectively validate models and test alternative machine learning methods.