Geometric neural network based on phase space for BCI-EEG decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep Learning (DL) integration in Brain-computer interface (BCI) lags behind other fields.
• Electroencephalography (EEG) is common for BCI but suffers from limited data and noise.
• Current BCI systems with many electrodes are cumbersome, hindering widespread adoption.

Purpose of the Study:

• To develop a DL algorithm for effective BCI with minimal electrodes.
• To improve user comfort and BCI system adoption.
• To enhance motor imagery (MI) decoding performance.

Main Methods:

• Proposed the Phase-SPDNet architecture, combining the Augmented Covariance Method and SPDNet.
• Evaluated performance using 5-fold cross-validation on 3 electrodes over the Motor Cortex.
• Tested on nearly 100 subjects across open-source datasets using the Mother Of All BCI Benchmark framework.

Main Results:

• Phase-SPDNet significantly outperformed state-of-the-art DL architectures in MI decoding.
• The augmented approach combined with SPDNet demonstrated superior performance.
• The proposed architecture achieved high accuracy with limited electrodes.

Conclusions:

• Phase-SPDNet offers an explainable DL solution for BCI.
• The architecture requires a low number of trainable parameters.
• This advancement facilitates more comfortable and reliable BCI systems.