Transfer Learning with CNN Models for Brain-Machine Interfaces to command lower-limb exoskeletons: A Solution for...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) offer potential for motor rehabilitation in individuals with impairments.
• Motor imagery (MI) based BMIs can translate imagined movements into commands for assistive devices like exoskeletons.
• Limited subject-specific data often hinders the performance of advanced machine learning models in BMIs.

Purpose of the Study:

• To evaluate the efficacy of two convolutional neural networks (CNNs) for motor imagery (MI) based brain-machine interface (BMI) control.
• To investigate the application of transfer learning to overcome data scarcity in MI-based BMI development.
• To assess the potential for creating an automatic neural classification system for commanding lower-limb exoskeletons.

Main Methods:

• Utilized a small dataset from five participants using a lower-limb exoskeleton.
• Employed transfer learning by pre-training CNN models on external EEG datasets and fine-tuning them for individual users.
• Compared CNN performance against a benchmark of Common Spatial Patterns (CSP) and Linear Discriminant Analysis (LDA).

Main Results:

• CNNs, particularly with transfer learning, demonstrated promising performance in classifying motor imagery for BMI control.
• Transfer learning effectively mitigated challenges associated with limited subject-specific EEG data.
• The developed system shows potential for commanding a lower-limb exoskeleton with improved accuracy.

Conclusions:

• CNNs combined with transfer learning represent a viable approach for developing robust MI-based BMIs.
• This methodology can facilitate the creation of user-friendly, automatic neural classification systems for assistive technologies.
• The findings support the clinical relevance of BMIs in enhancing motor recovery and promoting neural plasticity through exoskeleton-assisted rehabilitation.