Exploring EEG-based motor imagery decoding: a dual approach using spatial features and spectro-spatial Deep Learning...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Decoding motor imagery (MI) from electroencephalography (EEG) signals is crucial for brain-machine interfaces (BMIs) and neurorehabilitation.
• EEG signals are challenging for decoding due to non-stationarity and noise, hindering effective algorithm development.
• Accurate decoding algorithms are vital for controlling neurorehabilitation devices and stimulating motor cortex recovery.

Purpose of the Study:

• To propose and evaluate novel approaches for decoding MI during pedalling tasks using EEG signals.
• To investigate the efficacy of Common Spatial Patterns (CSP) with Linear Discriminant Analysis (LDA).
• To explore the performance enhancement offered by a spectro-spatial Convolutional Neural Network (CNN) compared to CSP-LDA.

Main Methods:

• EEG data were recorded from users pedalling on a cycle ergometer.
• Two decoding approaches were evaluated: CSP feature extraction with LDA classification, and a spectro-spatial CNN architecture.
• The CNN approach incorporated a filter bank preprocessing pipeline for spectro-temporal and spectro-spatial feature extraction.

Main Results:

• The CSP-LDA approach provided a baseline for motor imagery decoding.
• The proposed spectro-spatial CNN achieved higher decoding accuracy, reaching up to 80% in some instances.
• The CNN approach demonstrated greater accuracy but also exhibited higher instability compared to CSP-LDA.

Conclusions:

• The spectro-spatial CNN shows significant potential for improving motor imagery decoding from EEG signals.
• This advanced approach could enhance the effectiveness of BMIs and neurorehabilitation systems.
• Further research may focus on mitigating the instability observed in the CNN approach for more robust applications.