Unraveling motor imagery brain patterns using explainable artificial intelligence based on Shapley values.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Motor imagery (MI) brain-computer interfaces (BCIs) are crucial for rehabilitation, leveraging the link between MI and motor execution.
• Current MI-BCIs primarily focus on sensorimotor rhythms from M1 and S1, potentially overlooking broader neural networks.
• Understanding the full brain network involved in MI is essential for advancing BCI technology.

Purpose of the Study:

• To investigate the brain mechanisms underlying motor imagery (MI) using deep learning (DL) and explainable artificial intelligence (XAI).
• To determine if DL models achieve higher MI classification accuracy by analyzing a broader range of brain activity.
• To utilize XAI to identify specific brain regions contributing to MI decoding from EEG signals.

Main Methods:

• Applied an adapted Shapley Additive Explanations (SHAP) method to the EEGSym deep learning network.
• Utilized two public datasets with 171 participants performing left and right hand motor imagery tasks.
• Analyzed SHAP values to understand feature importance in inter-subject MI classification.

Main Results:

• Deep learning models primarily utilized frontal electrode signals (F7, F8) and early imagination periods (first 1500 ms).
• Motor imagery involves a distributed network including the prefrontal cortex (PFC) and posterior parietal cortex (PPC), beyond M1 and S1.
• An optimized 8-electrode configuration achieved high inter-subject accuracies (86.5%–88.7%) for MI classification.

Conclusions:

• Combining DL and SHAP-based XAI effectively elucidates the brain networks engaged during MI.
• The findings highlight the potential for optimizing BCI systems for real-world, out-of-laboratory applications.
• This approach enhances our understanding of MI and improves BCI design for rehabilitation and beyond.