Electroencephalography (EEG) can be instrumental in predicting the neurological recovery of patients six months after cardiac arrest-induced coma. We describe a preliminary convolutional neural network (CNN) using patient descriptors and frequency features, which were extracted from EEG taken over 72 hours following Return of Spontaneous Circulation (ROSC). This model was submitted to the 2023 George B. Moody Physionet/Computing in Cardiology Challenge by team ROSCy Business. We also describe planned variational autoencoder (VAE) and recurrent neural network (RNN) models.
We pre-processed the EEG signals by applying a bandpass filter and extracting Mel frequency cepstrum coefficients (MFCCs) in 10 second windows. Missing data were ignored, and only the most recent 12 hours of data considered, such that features were extracted and concatenated without imputation of unavailable recordings. Zero-padding was used if data were insufficient.
The 18 channels of MFCCs were fed into a CNN with two 2D 2x2 convolutions in the temporal dimension. Demographic information was concatenated in the final layer. In local cross-validation, this model achieved an average challenge score at 72 hours of 0.11. In the unofficial phase, this model performed poorly, with a challenge score at 72 hours of 0.03.
We will significantly change our approach in the challenge's official phase. To date, we have trained a VAE on the EEG, downsampled to 50 Hz in 5 second windows, treating each window independently. We used a mixture of the Kullback-Leibler Divergence, Structural SIMilarity index (SSIM), and L1 loss functions. The model captures macro trends successfully with 0.986 test SSIM after one epoch.
We will use the latent space representation as input features to a bidirectional RNN for classification. We will impute missing EEG data by interpolating within the VAE latent space, which we hope will provide more clinically plausible imputation than common approaches such as multiple imputation.