A Graph-Based Hierarchical Attention Model for Movement Intention Detection from EEG Signals

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCI) enable device control via brain signals.
• Current BCIs require subject-specific calibration, limiting widespread adoption.
• Subject-independent BCIs, applicable without pre-calibration, are highly desirable but underexplored.

Purpose of the Study:

• To address the challenge of subject-independent EEG-based human intention recognition.
• To develop a novel model that generalizes across users without individual training.
• To improve the direct applicability of BCI systems to new users.

Main Methods:

• Proposed a Graph-based Hierarchical Attention Model (G-HAM).
• Utilized graph structures to represent spatial EEG sensor information.
• Employed a hierarchical attention mechanism to focus on discriminative temporal periods and EEG nodes.

Main Results:

• G-HAM demonstrated superior performance in subject-independent EEG intention recognition.
• The model successfully exploited invariant EEG patterns across 105 subjects.
• Achieved better generalization to new users compared to state-of-the-art and baseline methods.

Conclusions:

• The G-HAM model offers a promising solution for subject-independent BCIs.
• This approach significantly advances the feasibility of direct BCI application.
• Highlights the potential of exploiting cross-subject EEG patterns for robust intention recognition.