Bi-dimensional representation of EEGs for BCI classification using CNN architectures.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Designing Brain Computer Interfaces (BCI) typically requires extensive parameter tuning for individual users and sessions.
• Convolutional Neural Networks (CNNs) excel at automatic feature extraction from image data, offering potential for handling unknown input data distributions.

Purpose of the Study:

• To develop a BCI pipeline that minimizes parameter adjustments for each subject and run.
• To leverage CNNs for automatic feature extraction from electroencephalography (EEG) signals by converting them into meaningful image representations.

Main Methods:

• Proposed two novel image representations derived from multichannel EEG signals: spectrograms and scalograms.
• Evaluated two classification approaches: a 2D CNN and a combination of 2D CNN with a Long Short-Term Memory (LSTM) network.
• Tested the pipeline on multiple BCI datasets, including BCI IV-2a (4 and 2 classes), BCI IV-2b, BCI II-III, and a private mental calculation database.

Main Results:

• The proposed pipeline demonstrated consistent performance across all tested subjects and datasets.
• Achieved competitive classification accuracies: 71.3 ± 11.9% (BCI IV-2a, 4 classes), 80.7 ± 11.8% (BCI IV-2a, 2 classes), 73.8 ± 12.1% (BCI IV-2b), 83.6 ± 1.0% (BCI II-III), and 82.10% ± 6.9% (private database).

Conclusions:

• The developed BCI pipeline, utilizing CNNs with EEG-derived image representations, effectively automates feature extraction.
• This approach significantly reduces the need for subject-specific parameter tuning, offering a more generalized and efficient BCI system.