NexusNet: Leveraging Nexuses in Graph Neural Network for Enhanced Motor Imagery Decoding

Core code will be made available upon acceptance.

This is a PyTorch implementation of NexusNet for MI decoding.

Abstract

• We propose a lightweight GNN, NexusNet, designed to capture complex relationships beyond pairwise connections.

• We conduct thorough experiments on two public datasets to validate NexusNet. Specifically, it achieves an average accuracy of 79.31% (hold-out) on the BCIC-IV-2a dataset and 87.70% (hold-out) on the BCIC-IV-2b dataset.

• We visualize the primary Nexuses to quantitatively analyze the relationships reconstructed by NexusNet. This visualization enables a detailed examination of how different Nexuses contribute to the decoding process.

Requirements

Please refer to requirements.txt

Model Zoos

Pretrained checkpoints are available in

• bciciv2a_checkpoint
• bciciv2b_checkpoint

License

This project is licensed under the MIT License - see the LICENSE file for details.