Classification of Motor Imagery Tasks Derived from Unilateral Upper Limb based on a Weight-optimized Learning Model

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Decoding fine motor imagery (MI) tasks is challenging due to dense cerebral cortex activity.
• Existing methods show limited accuracy in classifying subtle MI movements.

Purpose of the Study:

• To enhance the decoding accuracy of unilateral fine motor imagery (MI) tasks.
• To improve the classification of six specific MI types for the right upper limb.

Main Methods:

• A weight-optimized EEGNet model was developed for MI classification.
• Electroencephalography (EEG) data augmentation and a genetic algorithm (GA) were used to optimize model parameters.
• Convolutional kernel parameters were determined by GA, followed by weight optimization via backpropagation.

Main Results:

• The model achieved an average accuracy of 87.97% for three-joint classification.
• Binary classification accuracies for elbow, wrist, and hand joints were 93.92%, 90.2%, and 94.64%, respectively.
• The overall MI classification accuracy reached 81.74%, outperforming traditional algorithms and reducing average error.

Conclusions:

• The proposed algorithm effectively addresses parameter sensitivity and enhances MI classification robustness.
• The method demonstrates superior performance compared to existing neural networks and traditional algorithms.
• The approach is adaptable for other electroencephalography (EEG) classification tasks, including emotion and object recognition.