Game Theory-Based Adaptive Human-Machine Joint Learning for Online MI-BCI Decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Motor imagery brain-computer interfaces (MI-BCI) are crucial for neurorehabilitation and motor assistance.
• Current MI-BCI systems face limitations in online decoding accuracy for motor intentions.

Purpose of the Study:

• To develop an adaptive human-machine joint (AHMJ) learning method for improved online MI-BCI decoding.
• To enhance the performance of multi-class motor imagery (MI) decoding for unilateral upper limb movements.

Main Methods:

• Proposed a novel MI training method using game theory (two-player zero-sum minimax game) to adaptively regulate task difficulty.
• Developed a new online adaptive algorithm integrating knowledge distillation and prototype-guided domain adaptation for stable decoder updating.
• Conducted online MI-BCI experiments with 14 healthy subjects and simulation experiments on a public dataset of 25 healthy subjects.

Main Results:

• The AHMJ learning method significantly improved average decoding accuracy by 9.7% compared to traditional methods (p < 0.01).
• Achieved a 5.6% accuracy improvement over previous human-machine joint learning methods (p < 0.01).
• The online adaptive algorithm demonstrated superior accuracy and stability compared to existing updating approaches.

Conclusions:

• The proposed AHMJ learning method effectively enhances online MI-BCI decoding accuracy.
• This approach holds significant potential for advancing neurorehabilitation and motor assistance technologies.
• The adaptive training and updating strategies are key to improving MI-BCI system performance.