O77 Deep learning for EEG classification for outcome prediction of postanoxic coma

Objectives Electroencephalography is increasingly used for outcome prediction of comatose patients after cardiac arrest. Here, we explore a deep learning approach using convolutional neural networks. Methods Using epochs from continuous EEGs from 287 patients at 12 h and 399 patients at 24 h after cardiac arrest, we trained and validated a convolutional neural network with raw EEG data (longitudinal bipolar montage). The network consisted of one convolutional layer and two fully-connected neuronal layers. Clinical outcome at six months was classified by the Cerebral Performance Category scale (CPC), dichotomized as good (CPC 1–2) or poor (CPC 3–5). In 5 min artifact-free epochs, partitioned into 10 s snippets, we trained the network using data from 80% of the patients. Validation was performed in the remaining patients. The network was implemented in Python using Keras and Theano and a CUDA-enabled GPU. Results Outcome prediction was most accurate at 12 h after cardiac arrest, with a sensitivity of 58% at 100% specificity for the prediction of poor outcome. Good outcome could be predicted at 12 h with a sensitivity of 58% at a specificity of 97%. Training of the network took 50 min, classification was realized within 1 s. Discussion Deep learning allows outcome prediction with higher accuracies than visual EEG assessment, however it does not reveal which EEG features are the most discriminative. Conclusions We present a convolutional neural network for objective and reliable outcome prediction after cardiac arrest. Significance Deep learning might be an important future tool for neuro-monitoring in the ICU.