Deep Learning-Based Decoding and Feature Visualization of Motor Imagery Speeds From EEG Signals.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Decoding motor imagery (MI) speed is crucial for advanced brain-computer interfaces (BCIs).
• Understanding the neurodynamics underlying MI speed is essential for improving BCI performance.
• Deep learning models offer potential for analyzing complex electroencephalography (EEG) data in MI tasks.

Purpose of the Study:

• To investigate the neurodynamics of motor imagery speed decoding using deep learning.
• To identify temporal and spatial EEG patterns associated with different imagined movement speeds.
• To explore the role of specific frequency bands and cortical regions in MI speed decoding.

Main Methods:

• Utilized the EEGConformer deep learning model for EEG signal analysis.
• Applied explainable artificial intelligence (XAI) techniques to interpret model findings.
• Focused on identifying patterns in alpha and beta oscillations and key cortical areas.

Main Results:

• Successfully decoded EEG patterns related to different motor imagery speeds.
• Classification accuracy was limited and participant-specific.
• Highlighted the importance of alpha and beta oscillations and frontal, motor, and occipital cortices.
• Observed steady-state movement-related potentials at the fundamental frequency during repeated MI.

Conclusions:

• Motor imagery speed is decodable from EEG signals, but current classification performance is limited.
• Specific frequency bands (alpha, beta) and cortical regions are involved in encoding MI speed.
• Steady-state responses provide insights into the neural encoding of movement intention speed.