Improving the Performance of Electrotactile Brain-Computer Interface Using Machine Learning Methods on Multi-Channel...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Traditional tactile brain-computer interfaces (BCIs) using steady-state somatosensory-evoked potentials face limitations in accuracy and data transfer rates.
• Electrotactile stimulation generating somatosensory event-related potentials (sERPs) presents a promising alternative for tactile BCI control signals.

Purpose of the Study:

• To optimize the performance of a novel electrotactile BCI by applying advanced feature extraction and machine learning to sERP signals.
• To classify users' selective tactile attention based on EEG signals recorded over the sensory-motor cortex.

Main Methods:

• Utilized sequential forward selection (SFS) for feature extraction from temporal sERP waveforms across five EEG channels.
• Systematically evaluated classification performance using logistic regression, k-nearest neighbors, support vector machines, random forests, and artificial neural networks.
• Investigated the impact of the number of stimuli on classification accuracy and information transfer rate.

Main Results:

• Achieved significant improvements in classification accuracy compared to previous studies.
• Demonstrated that reducing the number of stimuli for sERP generation increased the information transfer rate without a significant accuracy drop.
• Attained over 90% accuracy with 10 stimuli and a less than 7% accuracy decrease with 6 stimuli, alongside a 60% increase in information transfer rate using a support vector machine classifier.

Conclusions:

• Advanced methods for tactile BCI control based on event-related potentials were developed.
• Optimizing sERP elicitation, feature extraction, and classification is crucial for balancing accuracy and speed in assistive BCI applications.
• Electrically induced sERPs represent an understudied but valuable control signal modality for reactive BCIs.