A Method with Time-Sensitive Features for the Automated Prognosis Prediction of Cardiac Arrest Patients Based on EEG

Aims: As part of the George B. Moody PhysioNet Challenge 2023, we proposed a method for the neurological recovery of patients following cardiac arrest by a convolutional neural network with time-sensitive features to realize the automated prognosis prediction for these patients after cardiac arrest. Methods: Firstly, we selected EEG records for the first 72 hours to build the input signal. During the data preprocessing, we used 3 strategies to get EEG segments including FFT EEG segments, time EEG segments, and enhanced time EEG segments. In our model, we designed 3 stacked Conv Blocks to extract features in each hour respectively. Then we designed the Time-sensitive Learning Block to learn the time-sensitive weights of these 72 hours. The features extracted by Conv Blocks were scaled by the time-sensitive weights so that we got the time-sensitive features. Then these features were sent into 3 Residual Blocks to get the final features which were used to predict the output of the model. Results: The model with the enhanced time EEG segments as the input and the SE Block in the Time-sensitive Learning Block performed best in the 5-fold cross validation experiments. Our team, Aircas, achieved the Challenge score of 0.45\pm 0.13 on the open training set and 0.475 on the hidden testing set with a rank of 20/36.