E2T: EEG-to-Trajectory Transformer for Motor Imagery-Based Fully-DoF Motion Prediction.

Area of Science:

• Neuroscience and Biomedical Engineering
• Brain-Computer Interfaces (BCIs)
• Signal Processing

Background:

• Traditional electroencephalography (EEG)-based BCIs for motor imagery (MI) lack flexibility for continuous movement control.
• Current classification-based BCIs translate discrete EEG patterns into limited commands, hindering naturalistic interaction.
• Individuals with motor impairments require advanced BCIs for seamless real-time control of assistive devices.

Purpose of the Study:

• To develop a novel EEG-based decoding framework for predicting continuous 3D upper limb trajectories.
• To overcome the limitations of discrete command translation in traditional motor imagery BCIs.
• To enable more naturalistic and flexible control for assistive neuroprosthetic systems.

Main Methods:

• A hybrid deep learning architecture integrating frequency-adaptive Sinc convolutional filters and a Transformer-based temporal encoder.
• A task-specific experimental paradigm with sequential motor execution (ME) and motor imagery (MI) phases.
• Ensemble learning to extract spatiotemporal features from raw EEG signals for 3D trajectory decoding.

Main Results:

• The framework successfully decoded continuous 3D upper limb trajectories across X, Y, and Z axes without directional bias.
• Achieved high decoding performance with average correlation coefficients of 0.7728 (ME) and 0.7110 (MI), outperforming baselines.
• Demonstrated generalization to unseen MI tasks, indicating feasibility for real-world applications with reduced calibration.

Conclusions:

• The proposed hybrid deep learning framework enables direct prediction of continuous 3D trajectories from EEG.
• This approach offers a significant advancement over traditional BCIs, providing more naturalistic motor control.
• The framework shows promise for practical deployment in assistive neuroprosthetic systems, enhancing user independence.