Mutual Generation for Cross-Domain Challenge in Stroke Patients' Motor Imagery Classification and Functional Recovery...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Stroke is a leading cause of motor disability, necessitating advanced rehabilitation strategies like Motor Imagery Brain-Computer Interfaces (MI-BCIs).
• Existing MI-BCI research predominantly uses data from healthy subjects, posing challenges for classifying Motor Imagery (MI) tasks in stroke patients due to hemispheric lateralization and lesion variations.
• Electroencephalogram (EEG) data from stroke patients exhibit significant variations based on hemiplegic side and lesion location, complicating model development.

Purpose of the Study:

• To explore the efficacy of generative models in mitigating domain differences in EEG data from stroke patients with varying hemiplegic sides.
• To enhance the performance of MI classification models by employing a label softening algorithm to avoid adverse effects from low-quality samples.
• To investigate the potential of the proposed framework for predicting the functional recovery level in stroke patients.

Main Methods:

• Utilized generative models to address domain discrepancies in EEG data from stroke patients.
• Implemented a label softening algorithm to improve model robustness against noisy or low-quality data.
• Validated the approach using two MI-EEG datasets from stroke patients, comparing performance against classical machine learning and state-of-the-art models.
• Introduced a prediction layer to assess the model's capability in forecasting functional recovery.

Main Results:

• The proposed generative model significantly improved MI classification performance across subject-dependent and subject-independent scenarios compared to existing methods.
• Each component of the model, including sub-modules and loss functions, demonstrably contributed to the overall performance enhancement.
• The addition of a prediction layer enabled accurate forecasting of functional recovery levels in stroke patients.

Conclusions:

• Generative models offer a promising solution for overcoming domain differences in stroke patient EEG data for MI-BCI applications.
• The developed MI-BCI classification framework, incorporating generative models and label softening, achieves superior performance and aids in predicting patient recovery.
• The framework's ability to predict functional recovery opens avenues for personalized rehabilitation planning and monitoring in stroke survivors.