Benefits of deep learning classification of continuous noninvasive brain-computer interface control

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Noninvasive brain-computer interfaces (BCIs) offer a vital communication pathway for individuals with paralysis.
• Deep learning, specifically convolutional neural networks (CNNs), shows promise in enhancing EEG-based motor imagery BCIs.

Purpose of the Study:

• To evaluate the generalization of deep learning improvements in BCIs to continuous control tasks.
• To investigate the utility of neural biomarkers outside the motor cortex for BCI performance.
• To identify challenges in the practical implementation of deep learning-based continuous BCI control.

Main Methods:

• Utilized a large, longitudinal online motor imagery BCI dataset with 4-class continuous 2D feedback.
• Compared deep learning methods (CNNs) against standard methods for decoding accuracy.
• Assessed the impact of full scalp EEG coverage versus motor cortex-only montages.
• Simulated online cursor control to evaluate CNN model output optimization.

Main Results:

• Deep learning methods significantly outperformed standard methods in offline BCI performance.
• Neural biomarkers beyond the motor cortex were identified and leveraged by deep learning to improve performance.
• Tuning CNN output and using full scalp EEG notably reduced average trial length in simulated online control.

Conclusions:

• CNNs offer significant advantages for BCI control, improving accuracy and efficiency.
• Electrode montage selection is crucial, with full scalp coverage showing promise.
• Optimizing the mapping of CNN output to device control is key for practical CNN-based BCIs.