Predicting Neurological Outcomes for Cardiac Arrest Patients from Longitudinal EEG Based on Short-Time Fourier Transform and 3-D Deep Residual Network

The George B. Moody PhysioNet Challenge 2023 focused on predicting neurological recovery from coma after cardiac arrest by longitudinal electroencephalogram (EEG). Our team, MetaHeart_YNNU, proposed a novel approach to predict good and poor patient outcomes after cardiac arrest by combing short-time Fourier transform (STFT) and 3-dimensional (3-D) residual neural network. Firstly, a 90-second signal segment was obtained from the 8-channel EEG recording of each hour by truncating the first 15 seconds every 10 minutes, and then these segments from each hour were transformed into the time-frequency spectrograms by STFT to obtain a richer representation of neurological activity. Secondly, a modified 3-D residual neural network was designed to extract the complex physiological patterns of neurological activity from hour, time-frequency dimensions, respectively. Thirdly, a joint loss combing binary focal loss and recall loss was designed in our approach. The binary focal loss alleviated the class imbalance problem by paying more attention to hard-to-distinguish samples. The recall loss was adopted to reduce the probability that patients with the true good label were misclassified into the poor category by maximizing the recall rate of the good category. Finally, our proposed model received a Challenge score of 0.56 (ranked 12th out of 36 teams) on the hidden test set.