Introduction: Patients surviving cardiac arrest with primary good outcome are in postanoxic coma with uncertain neurological damage. Secondary neurological outcome can be predicted from the electroencephalogram (EEG) to a certain degree. We (team BrAInstorm) aim to improve this prediction by extending the spectrum of considered EEG features.
Methods: We developed a fully-automatic processing pipeline for the prediction of secondary neurological outcome from a longitudinal, 19-channel EEG (100 Hz sampling rate). We started by ensuring adequate signal quality by removing flatlines and clipped amplitudes across all channels. Next, we extracted EEG features adapted from neurosurgical patient care including statistical time domain features, coherence, and power spectra. Additionally, we calculated trends for all features of the consecutive hourly EEG segments. All features served as input for a non-optimized Random Forest classifier provided by the challenge to distinguish patient-wise good and bad outcome and Cerebral Performance Categories.
Results: On provided test data, we yield a true positive rate of 0.21 for predicting poor outcome given a false positive rate of less or equal to 0.05. This puts us in rank 129 of the unofficial phase. Applying 5-fold cross validation, we yield 0.37 ± 0.12 (mean ± standard deviation).
Discussion: We successfully implemented a complete pipeline for secondary neurologic outcome predicting after cardiac arrest based on an input EEG. To further enhance performance, we aim to improve our feature set and optimize our classifier. Additional features will include (1) EEG features from recent publications, e.g. burst suppressions with identical patterns, signs of seizure, and coupling of different power bands and brain regions, (2) features from the field of pattern recognition, and (3) established features from EEG-based sleep staging. To improve the classification, we will implement a Deep Neural Network to replace the Random Forest, and perform extensive hyperparameter optimisation.