EEG motor imagery classification using deep learning approaches in naïve BCI users.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Motor Imagery (MI)-Brain Computer-Interfaces (BCI) illiteracy affects user performance due to factors like fatigue and lack of experience.
• Naïve BCI users often struggle to achieve optimal system performance.

Purpose of the Study:

• To investigate the effectiveness of three Deep Learning (DL) methods in improving BCI performance for naïve users.
• To compare DL methods against traditional baseline methods for MI signal discrimination.

Main Methods:

• Implementation of Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM)/Bidirectional Long Short-Term Memory (BiLSTM), and CNN-LSTM models.
• Evaluation on a dataset of 25 naïve BCI users for upper limb MI signal discrimination.
• Comparison with baseline methods: Common Spatial Pattern (CSP), Filter Bank Common Spatial Pattern (FBCSP), and Filter Bank Common Spatial-Spectral Pattern (FBCSSP).

Main Results:

• The LSTM-BiLSTM approach demonstrated superior performance, achieving a mean accuracy of 80% (up to 95%) and an Information Transfer Rate (ITR) of 10 bits/min with a 1.5s temporal window.
• DL methods showed a significant 32% performance increase compared to baseline methods (p<0.05).
• Key performance metrics included Accuracy, F-score, Recall, Specificity, Precision, and ITR.

Conclusions:

• Deep Learning methods, particularly LSTM-BiLSTM, substantially enhance BCI system performance for users lacking experience.
• The findings suggest increased controllability, usability, and reliability of robotic devices for naïve BCI users.
• This research paves the way for more accessible and effective BCI applications.