BandFocusNet: A Lightweight Model for Motor Imagery Classification of a Supernumerary Thumb in Virtual Reality

Area of Science:

• Neuroscience and Human-Computer Interaction
• Robotics and Human Augmentation

Background:

• Controlling supernumerary effectors for human movement augmentation is challenging due to issues with agency, control, and synchronization.
• Electroencephalography (EEG) via motor imagery (MI) presents a promising control strategy for these artificial limbs.

Purpose of the Study:

• To investigate if MI activity from a supernumerary effector can be reliably differentiated from that of a natural effector.
• To address concerns regarding the concurrency and distinguishability of brain signals for simultaneous natural and augmented limb control.

Main Methods:

• Twenty subjects participated in a virtual reality experiment observing and performing MI of natural and supernumerary thumb movements.
• A novel deep learning model, BandFocusNet, was developed to analyze temporal, spatial, and spectral EEG data.
• Leave-one-subject-out cross-validation was employed for model performance assessment, alongside explainability and event-related spectral perturbation (ERSP) analyses.

Main Results:

• BandFocusNet achieved an average classification accuracy of 70.9% in differentiating MI between natural and supernumerary thumbs.
• Explainability analysis highlighted the importance of frontal cortical regions.
• ERSP analysis revealed increased delta and theta power in these regions during natural thumb MI, but not during supernumerary thumb MI.

Conclusions:

• The findings suggest that distinct neural activation patterns exist for MI of natural versus supernumerary effectors.
• The observed absence of typical natural effector activation during supernumerary effector MI may indicate a lack of embodiment.
• This research provides a foundation for developing more intuitive and effective control systems for human movement augmentation.