Convolutional neural network architecture and input volume matrix design for ERP classifications in a tactile...

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCI) enable communication and control through brain signals.
• P300-based BCIs utilize event-related potentials (ERPs) for signal detection.
• Current BCI systems often require personalized training for optimal performance.

Purpose of the Study:

• To enhance classification accuracy in full-body tactile P300-based BCIs (fbBCI).
• To investigate the efficacy of Convolutional Neural Networks (CNNs) for ERP classification in fbBCI.
• To achieve non-personalized ERP classification for improved BCI usability.

Main Methods:

• Offline ERP classification using a CNN classifier.
• Input data: 1D somatosensory ERP intervals from electrode channels, formatted as a 60x60 input volume.
• Classification architecture: CNN for filter training followed by a multilayer perceptron with one hidden layer.

Main Results:

• The CNN classifier achieved 100% classification accuracy for all ten participants.
• The proposed method demonstrated successful non-personalized ERP classification.
• The CNN model effectively processed somatosensory ERP intervals for fbBCI.

Conclusions:

• CNNs are highly effective for classifying ERPs in fbBCI paradigms.
• Non-personalized classification with 100% accuracy is achievable using CNNs.
• This approach significantly advances the potential for practical and user-friendly BCIs.