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

초록

Brain-computer interfaces (BCIs) using electroencephalography (EEG) enable non-invasive, real-time interaction for individuals with motor impairments by decoding neural signals associated with movement intention. Although traditional classification-based motor imagery (MI) BCIs translate discrete EEG patterns into predefined commands, such approaches inherently lack the flexibility required for continuous and naturalistic movement control. To overcome these limitations, we propose a novel EEG-based decoding framework that directly predicts continuous 3D upper limb trajectories using a hybrid deep learning architecture. The model integrates frequency-adaptive Sinc convolutional filters, a Transformer-based temporal encoder, and ensemble learning to extract meaningful spatiotemporal features from raw EEG signals. To ensure balanced spatial coverage, we developed a task-specific experimental paradigm in which each trial comprised sequential motor execution (ME) and MI phases based on one of four predefined 3D directions: forward-vertical, forward-leftward, forward-rightward, and lateral-vertical. The MI phase was placed after the ME phase to reinforce vivid internal representations of movement. This approach allowed the model to capture continuous 3D trajectories across the X, Y, and Z axes without directional bias. Comprehensive evaluation under both ME and MI conditions demonstrates robust decoding performance, achieving average correlation coefficients of 0.7728 (ME) and 0.7110 (MI), significantly outperforming existing baseline models. Furthermore, the framework generalizes well to unseen MI tasks, demonstrating its feasibility for real-world deployment in assistive neuroprosthetic systems with a reduced calibration session.

키워드