Abstract:
Research Interests: Efficient classification of electroencephalogram (EEG) signals is
crucial for the development of brain-computer interface systems. However, the complexity
and variability of EEG signals pose significant challenges for accurate classification. Additionally,
this study has social relevance as it can contribute to the development of assistive
brain-computer interfaces, benefiting individuals with severe motor impairments, such as
those who have experienced a stroke. These interfaces have the potential to improve the
quality of life for these individuals by enabling communication and device control through
brain activity.
Objectives: This study aimed to compare the performance and computational cost of an
artificial neural network using different signal processing techniques for the classification
of resting state and left/right wrist movement imagination states from EEG signals. Three
statistical signal processing techniques, Principal Component Analysis (PCA), Independent
Component Analysis (ICA), and Singular Spectrum Analysis (SSA), were explored
in conjunction with a Convolutional Neural Network (CNN) to enhance the classification
of EEG signals.
Results Obtained: The results revealed that the PCA technique led to a reduction in
training time of up to 63.5% without significantly compromising performance in terms
of classification accuracy. PCA proved to be a promising approach, capturing relevant
information from the EEG signals and improving the CNN’s ability to classify accurately.
On the other hand, both ICA and SSA techniques did not yield promising results. ICA
had negative effects on feature extraction, resulting in decreased classification accuracy
by the CNN. SSA, on the other hand, showed consistently low performance across all
evaluated metrics, indicating challenges in capturing discriminative information from the
EEG-IM signals.