Neuroscience-Inspired Hierarchical GNN for Grasping Attempt Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-Computer Interfaces (BCI) show potential for upper limb stroke rehabilitation.
• Restoring fine hand functions like grasping remains a significant challenge in current BCI applications.

Purpose of the Study:

• To decode hand grasp attempts using electroencephalography (EEG) for BCI-driven hand rehabilitation.
• To develop novel methods that improve the accuracy of decoding hand movements from EEG signals.

Main Methods:

• Proposed a Small-world Hierarchical Interconnected Graph Neural Network (SHINE) inspired by Small-World Brain Network Theory.
• Implemented SHINE with multiscale convolution, overlapping windows, and learnable variance to capture transient power dynamics.
• Introduced a Progressive Decay Graph (PDG) mechanism to dynamically adjust long-range connection weights based on distance and training progress.

Main Results:

• SHINE demonstrated superior performance over state-of-the-art methods on two EEG datasets (healthy subjects and stroke patients).
• Achieved significant performance improvements in decoding hand opening and closing attempts compared to rest conditions.
• Reported statistically significant accuracy enhancements, including 3.97% for stroke patients attempting to open their hand (p<0.01).

Conclusions:

• The proposed SHINE model effectively decodes hand grasp attempts from EEG signals, outperforming existing methods.
• SHINE's architecture, inspired by brain network principles, offers advancements over traditional Graph Neural Network approaches.
• This work contributes to the development of more effective BCI-driven rehabilitation strategies for restoring hand function after stroke.