EEG decoding with spatiotemporal convolutional neural network for visualization and closed-loop control of...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Closed-loop neurofeedback uses neural signals like electroencephalograms (EEG) to modulate brain activity and behavior.
• Standard EEG preprocessing in brain-computer interface (BCI) paradigms can lack spatial precision.
• High-density EEG signals contain information from remote brain regions, posing a source ambiguity challenge.

Purpose of the Study:

• To develop and evaluate a spatiotemporal filter using a convolutional neural network (CNN) for high-density EEG.
• To improve the spatial precision and specificity of neural signal reconstruction during BCI-based neurofeedback.
• To compare the efficacy of CNN-filtered EEG with standard preprocessing pipelines in capturing sensorimotor network activity.

Main Methods:

• Simultaneous acquisition of EEG and functional magnetic resonance imaging (fMRI) in human participants during BCI neurofeedback training.
• Application of a CNN-based spatiotemporal filter to high-density EEG data.
• Comparison of reconstructed hemodynamic responses from CNN-filtered EEG against modeled responses of the sensorimotor network.
• Analysis of CNN model's middle layers to identify contributing neuronal oscillatory features.

Main Results:

• CNN-constructed filters demonstrated superior spatial precision and specificity in capturing targeted sensorimotor network activities compared to standard pipelines.
• The CNN model successfully reconstructed hemodynamic responses, indicating effective temporal feature extraction.
• Analysis revealed that distributed cortical circuitries, including frontoparietal and sensorimotor areas, contributed to the reconstruction.
• Specific neuronal oscillatory features influencing hemodynamic response reconstruction were identified within the CNN's middle layers.

Conclusions:

• CNN-based spatiotemporal filtering offers a significant advancement over standard methods for EEG signal processing in neurofeedback.
• The developed filter enhances the ability to precisely target and modulate neural activity within specific brain networks.
• Leveraging identified electrophysiological signatures and spatiotemporal filtering can lead to more effective neurofeedback interventions for various applications.