Develop an automated algorithm using quantitative EEG (qEEG) to predict discharge Glasgow outcome scale (GOS) after subarachnoid hemorrhage (SAH).

Continuous electroencephalography (EEG) is used for monitoring after SAH. Seizures, epileptiform abnormality burden, and quantitative EEG measures such as alpha power predict outcome after SAH (De Marchis 2015, Zafar 2018, Gollwitzer 2019). We hypothesized that machine learning combining multiple qEEG features improves outcome prediction after SAH.

We retrospectively analyzed patients with moderate to severe SAH (HH3-5 or FS3-4; 2011-2015) who were monitored with continuous EEG. We dichotomized GOS at discharge as poor (GOS=1-3) or good (GOS=4-5). We extracted eight qEEG features: total power, delta power, theta power, alpha power, alpha delta ratio (ADR), percent alpha variability (PAV), Shannon entropy and epileptiform discharge (ED) burden (defined as the number of sporadic or periodic discharges detected per hour). We used Persyst to identify EDs and used Matlab to calculate all features. We trained random forest models to predict discharge outcome and tested performance using 5-fold cross validation. 

113 patients met study criteria, and 91 (80.53%) had poor GOS at discharge. Our best model, a random forest with 60 trees, included global and cumulative ADR, total power, delta power, alpha power, percent alpha variability, and ED burden. The model predicted discharge outcome with area under the receiver operating characteristic curve (AUC)=0.707-0.769 2-10 days post-SAH. The model performed the best 4.5 days post-SAH (AUC=0.769; 95% Confidence Interval=[0.720, 0.812]). The three features with the highest predictor importance estimates were cumulative ED burden, PAV, and total power. We compared the results of our algorithm with a logistic regression of Hunt-Hess at admission as a predictor (AUC=0.686, 95% CI=[0.571-0.800]).

Automated calculations of qEEG features can be used in a random forest model to predict SAH outcome.