Finding Neural Biomarkers for Motor Learning and Rehabilitation using an Explainable Graph Neural Network.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Human motor learning is crucial for daily activities but is impaired in neurological disorders like Parkinson's Disease and stroke.
• Identifying neural biomarkers for motor learning is vital for developing effective therapeutic strategies.
• Traditional data analysis methods struggle with the complexity of brain activity, hindering biomarker discovery.

Purpose of the Study:

• To develop a novel Spatial Graph Neural Network (SGNN) model for predicting motor learning outcomes from electroencephalogram (EEG) data.
• To investigate the spatial-temporal dynamics of brain activity during distinct motor learning processes.
• To identify neural biomarkers associated with error-based and reward-based motor learning.

Main Methods:

• Utilized a novel Spatial Graph Neural Network (SGNN) model to analyze electroencephalogram (EEG) data.
• Collected EEG data during a visuomotor rotation (VMR) task to differentiate error-based and reward-based learning.
• Employed spatial, spectral, and temporal explainability methods to interpret SGNN findings and identify key neural features.

Main Results:

• The SGNN model successfully predicted motor learning outcomes from EEG data.
• Identified specific brain regions and temporal dynamics crucial for different types of motor learning (error-based and reward-based).
• Explainability methods validated the identified neural biomarkers against existing literature and ablation studies.

Conclusions:

• The developed SGNN model offers a powerful new method for analyzing complex brain signals like EEG.
• This approach provides significant insights into the neural underpinnings of human motor learning.
• The methodology can be extended to discover biomarkers from diverse brain signals and tasks for various neurological conditions.