Classification of motor imagery EEG using deep learning increases performance in inefficient BCI users.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Motor Imagery Brain-Computer Interfaces (MI-BCIs) traditionally use Machine Learning (ML) requiring extensive signal processing.
• Deep Learning (DL) models offer automatic feature extraction for EEG classification but haven't been evaluated for BCI-inefficient users.
• BCI inefficiency, the inability to produce desired sensorimotor rhythm patterns, remains a significant challenge.

Purpose of the Study:

• To evaluate the effectiveness of DL models in capturing Motor Imagery (MI) features, particularly in BCI-inefficient users.
• To compare the performance of a DL approach against a traditional ML approach for MI classification.
• To investigate if DL models offer advantages for users with lower BCI performance.

Main Methods:

• Recorded EEG signals from 54 subjects performing a left- or right-hand grasp MI task.
• Implemented a ML approach using Common Spatial Patterns (CSP) for feature extraction and Linear Discriminant Analysis (LDA) for classification.
• Developed a Deep Learning (DL) approach using a Convolutional Neural Network (CNN) directly on raw EEG signals.
• Compared classifier performance between high and low performing user groups.

Main Results:

• The CNN model improved classification accuracy for all subjects by 2.37-28.28%.
• This accuracy improvement was significantly greater for the low-performing user group.
• DL models demonstrated superior performance in capturing MI features, especially for users with BCI inefficiency.

Conclusions:

• Deep learning models show promise for enhancing MI-BCI systems by utilizing raw EEG signals.
• DL approaches, particularly CNNs, can significantly benefit BCI-inefficient users who struggle with conventional ML methods.
• This study highlights the potential of DL to overcome BCI inefficiency and improve user accessibility.