Brain-machine interface based on deep learning to control asynchronously a lower-limb robotic exoskeleton: a...

Area of Science:

• Neuroscience
• Robotics
• Machine Learning

Background:

• Traditional brain-machine interfaces (BMIs) face limitations in feature extraction and transfer learning.
• Deep learning offers automated feature extraction and transfer learning for enhanced BMI performance.
• This research focuses on motor imagery (MI) based BMIs for lower-limb robotic exoskeletons.

Purpose of the Study:

• To develop and evaluate a deep learning-based BMI for controlling a lower-limb robotic exoskeleton.
• To investigate the efficacy of transfer learning and model fine-tuning in BMI development.
• To compare different deep learning approaches for decoding neural signals in an asynchronous BMI protocol.

Main Methods:

• Five healthy subjects participated in experimental sessions to collect brain signals.
• A generic deep learning model was developed using transfer learning from initial sessions.
• Three deep learning approaches were compared: no fine-tuning, full fine-tuning, and partial fine-tuning (last three layers).
• Evaluation involved closed-loop control of the exoskeleton using neural activity.

Main Results:

• Deep learning approaches outperformed a traditional spatial features-based method.
• A non-fine-tuned deep learning model showed performance comparable to the features-based approach.
• Fine-tuning all layers of the deep learning model yielded the highest performance.
• Transfer learning demonstrated the potential for generic models across subjects and sessions.

Conclusions:

• The study is a step towards calibration-free BMI methods, reducing training time.
• Complete elimination of calibration was not achieved, but significant progress was made.
• The asynchronous protocol enhanced subject autonomy, mimicking real-world scenarios.
• Findings support the advancement of BMI technology for exoskeleton control with reduced user training.