EEG-Based Feature Classification Combining 3D-Convolutional Neural Networks with Generative Adversarial Networks for...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for electroencephalogram (EEG)-based motor imagery (MI) decoding in brain-computer interfaces (BCIs).
• Effective MI feature extraction is crucial due to individual and temporal variability in EEG signals.

Purpose of the Study:

• To introduce a novel 3D-convolutional neural network-generative adversarial network (3D-CNN-GAN) for decoding motor imagery.
• To enhance the generalizability and performance of EEG-based BCIs by improving feature extraction and data augmentation.

Main Methods:

• EEG signals were processed using a sliding window to capture temporal, frequency, and phase features, forming a 3D feature map.
• Generative adversarial networks (GANs) synthesized artificial data to augment the dataset and enhance functional connectivity.
• A compact 3D-CNN model was developed to decode the generated temporal-frequency-phase feature (TFPF) maps.

Main Results:

• The 3D-CNN-GAN model achieved improved two-class within-session motor imagery accuracies on both GigaDB (77.03%) and SHU (71.63%) datasets compared to the 3D-CNN model.
• The 3D-CNN-GAN also demonstrated enhanced cross-session motor imagery accuracy (63.04%) on the SHU dataset.
• Results indicate the effectiveness of data augmentation and feature amplification by GANs for EEG decoding.

Conclusions:

• The 3D-CNN-GAN algorithm significantly improves the generalizability of EEG-based motor imagery BCIs.
• This approach offers valuable insights into advancing motor imagery BCI applications and performance.