The analysis of EEG signals is essential for assessing the mental state of a patient during coma. In fact, some EEG patterns are associated with milder situations with a good prognosis, while others with more complicated or irreversible conditions. Deep learning models can automate the analysis of EEG signals and make objective predictions about a patient's neurological outcome, by learning the outcome associated to each of these EEG patterns.
This work aims to predict the neurological recovery of a patient from coma after a cardiac arrest, by using only the most recent five-minute EEG recording available. Data provided by the 2023 PhysioNet Challenge were used. This approach is based on the idea that EEG recordings become more informative as the comatose state progresses.
We modeled the problem as a binary classification task over the two possible outcomes (recovery or no-recovery). Since only one recording per patient was considered, we chose to develop a classifier composed by a one-dimensional three-layer convolutional neural network (CNN), followed by three fully connected layers. The EEG signals were resampled from 100 Hz to 30 Hz and provided as input to the network, so to force the CNN to extract features from the lower band of the spectrum of the cerebral activity. Different custom loss functions were tested, considering the value of the cerebral performance category scale of the patients or on the timing of the recordings. The model was evaluated using a 5-fold cross-validation on the provided data, or on the independent test set of the challenge. On cross-validation, our model achieved a challenge score (the highest true positive rate at a false positive rate lower than 0.05) of 0.231±0.069 (mean±std) and AUC of 0.767±0.045. On the competition test set, it scored 0.224 (team name: unimi_bisp_squad).