A Parallel Multiscale Filter Bank Convolutional Neural Networks for Motor Imagery EEG Classification

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) based brain-computer interfaces (BCI) are advancing rapidly for motor imagery (MI) tasks.
• Low signal-to-noise ratio (SNR) and time-varying characteristics of EEG signals necessitate robust feature extraction.
• Deep learning methods show promise in EEG signal processing, yet designing effective end-to-end networks for MI remains challenging.

Purpose of the Study:

• To propose a parallel multiscale filter bank convolutional neural network (MSFBCNN) for enhanced MI classification.
• To develop a layered end-to-end network for extracting temporal and spatial EEG features.
• To improve transfer learning capabilities for inter-subject MI classification on limited datasets.

Main Methods:

• A parallel multiscale filter bank convolutional neural network (MSFBCNN) architecture was designed.
• A feature-extraction sub-network was implemented to capture temporal and spatial EEG characteristics.
• A network initialization and fine-tuning strategy was employed to facilitate transfer learning for inter-subject classification.

Main Results:

• The MSFBCNN achieved higher accuracy in intra-subject MI classification compared to baseline methods.
• Transfer learning experiments demonstrated the network's ability to create individual models with acceptable performance for inter-subject classification.
• The proposed network exhibited superior performance, robustness, and transfer learning ability.

Conclusions:

• The MSFBCNN offers a powerful deep learning solution for EEG-based motor imagery classification.
• The developed transfer learning strategy enables effective inter-subject classification even with small datasets.
• This approach advances the development of reliable and robust brain-computer interfaces.