PMF-CNN: parallel multi-band fusion convolutional neural network for SSVEP-EEG decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Steady-state visual evoked potential (SSVEP) electroencephalography (EEG) is crucial for brain-computer interfaces (BCI) in neurological assessment and rehabilitation.
• Decoding SSVEP-EEG signals faces challenges due to low signal-to-noise ratio and individual variability.
• Existing methods struggle to fully leverage the spatio-temporal-frequency information within EEG signals.

Purpose of the Study:

• To develop an advanced deep learning model for accurate SSVEP-EEG signal classification.
• To address the limitations of low signal-to-noise ratio and individual variability in SSVEP-EEG decoding.
• To enhance the robustness and applicability of BCIs in postoperative rehabilitation.

Main Methods:

• Proposed a Parallel Multi-Band Fusion Convolutional Neural Network (PMF-CNN) integrating multi-frequency band EEG signals.
• Employed parallel spatial self-attention (SAM), temporal self-attention (TAM), and squeeze-excitation (SEM) modules for multi-dimensional feature extraction.
• Developed a novel spatial-temporal-frequency representation and a four-layer CNN classification module.
• Utilized a dual-stage training pattern for model optimization.

Main Results:

• Achieved high classification accuracies of 99.37% and 93.96% on two large public datasets.
• Demonstrated superior performance compared to current state-of-the-art SSVEP-EEG classification algorithms.
• The PMF-CNN exhibited high classification accuracy and good robustness, validated through brain functional connectivity analysis.

Conclusions:

• The proposed PMF-CNN effectively decodes SSVEP-EEG signals by fusing multi-dimensional features.
• The model's high accuracy and robustness show significant potential for applications in postoperative rehabilitation.
• This advancement contributes to more reliable and effective brain-computer interfaces.