From Frequency to Temporal: Three Simple Steps Achieve Lightweight High-Performance Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) based motor imagery decoding faces challenges with high data noise and computational complexity.
• Existing deep learning models often struggle to balance accuracy and efficiency.

Purpose of the Study:

• To develop a novel EEG motor imagery decoding method that achieves high accuracy with low computational cost.
• To address limitations of current deep learning approaches in EEG signal processing.

Main Methods:

• Frequency domain analysis was used to optimize EEGNet by adjusting convolution kernels and pooling sizes based on brain science knowledge of key frequency bands.
• A residual network was incorporated to preserve high-frequency detailed features.
• Temporal convolution modules were employed to enhance feature discriminability by capturing temporal dependencies.

Main Results:

• The proposed method achieved high average classification accuracies of 86.23% (BCI Competition IV 2a) and 86.75% (BCI Competition IV 2b).
• Computational cost was significantly reduced, with Multiply-accumulate operations (MACs) over 50% lower than advanced models (27.16M) and a Forward/Backward Pass Size of 14.33MB.

Conclusions:

• Integrating brain science knowledge with deep learning techniques effectively enhances EEG motor imagery decoding accuracy and reduces computational complexity.
• The study highlights the importance of incorporating neuroscience principles into AI model development for practical applications.