A Method to Extract Task-Related EEG Feature Based on Lightweight Convolutional Neural Network

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Extracting task-related electroencephalogram (EEG) spectra is vital for understanding brain activity.
• Traditional convolutional neural networks (CNNs) show promise but struggle with overfitting on small datasets.

Purpose of the Study:

• To develop a lightweight CNN model for efficient and interpretable EEG spectral feature extraction.
• To evaluate the model's performance on diverse neuroscientific and brain-computer interface tasks.

Main Methods:

• Proposed a lightweight CNN architecture with interpretability assessed via its fully connected layer (FCL).
• Tested the model on tasks including eye-state classification, seizure detection, and hand movement decoding.
• Analyzed spectral features in alpha, theta, and delta bands across different EEG channels.

Main Results:

• The CNN successfully identified task-related spectral features, notably alpha band for eye-state and theta band for seizure detection.
• A significant correlation was found between delta activity and hand movement in brain-computer interface tasks.
• The proposed method achieved comparable interpretability with substantially fewer trainable parameters than existing approaches.

Conclusions:

• The lightweight CNN offers an effective and interpretable solution for EEG spectral analysis.
• This approach demonstrates potential for broader applications in EEG decoding and brain-computer interfaces.
• Reduced model complexity enhances suitability for real-world, resource-constrained scenarios.