The goal of the 2023 Physionet challenge is to develop a model that can predict the neurological recovery of comatose patients (following cardiac arrest) using EEG data and patient metadata.
To address this challenge, we initially used two deep networks, Deep4 and EE(G)-SPDNet, which have previously achieved state-of-the-art performance in various EEG decoding tasks. While Deep4 is a regular deep convolutional network, EE(G)-SPDNet mixes Riemannian-based and convolutional layers.
In the unofficial phase of the challenge, we achieved our highest score of 0.57 on the validation set by adapting EE(G)-SPDNet to receive categorical information from patient metadata (learned embeddings of the metadata were compressed via an MLP before being fused with the EE(G)-SPDNet output). We will continue to optimize this approach, as we seek to determine the most effective way to fuse EEG data with patient metadata in this architecture. In a 10-fold cross-validation (CV) on the training data, our Deep4-network (not using any metadata) achieved an AUC of 0.79 and a challenge score of 0.29. In a 5-fold CV the EE(G)-SPDNet achieved an AUC of 0.81 and a challenge score of 0.33.
We plan to explore other established convolutional networks, such as EEGNet, TCNNet, and ShallowNet, as well as invertible networks and transformer-based networks methods. For those networks, we intend to investigate how EEG and categorical data can be effectively encoded and used to maximize the decoding performance. Finally, if time permits, we also plan to explore the usage of invertible networks that may help better understand the discriminative features in the EEG data.