Electroencephalographic Signal Data Augmentation Based on Improved Generative Adversarial Network.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning methods are crucial for analyzing electroencephalogram (EEG) signals in human-computer interaction.
• Limited EEG datasets pose a significant challenge for training effective deep learning models.
• Existing data augmentation techniques may not fully capture the complexity of EEG signals.

Purpose of the Study:

• To develop an improved generative adversarial network (GAN) model, L-C-WGAN-GP, for generating synthetic EEG data.
• To augment limited EEG training datasets to enhance the performance of deep learning applications, particularly in Brain-Computer Interfaces (BCI).
• To improve the accuracy of compressed sensing reconstruction for EEG signals using generated data.

Main Methods:

• The proposed L-C-WGAN-GP model utilizes a generator with a Long Short-Term Memory (LSTM) network and a discriminator with a Convolutional Neural Network (CNN).
• The model employs a gradient penalty-based Wasserstein distance as the loss function for training.
• The generated EEG data is used to augment existing datasets for model training and compressed sensing reconstruction.

Main Results:

• The L-C-WGAN-GP model successfully learns statistical features of EEG signals, generating synthetic data that closely approximates real samples.
• Quantitative evaluation using RMSE, FD, and WTD indicators shows the generated EEG signals are superior to those from existing advanced amplification techniques.
• Incorporating the augmented dataset reduced the loss in compressed sensing reconstruction of EEG signals by approximately 15%.

Conclusions:

• The L-C-WGAN-GP model provides an effective method for generating high-fidelity artificial EEG data.
• Data augmentation using L-C-WGAN-GP significantly improves the performance and accuracy of deep learning models in BCI applications.
• The study demonstrates a substantial improvement in EEG signal compressed sensing reconstruction accuracy through the use of synthetically generated data.