A motor imagery classification model based on hybrid brain-computer interface and multitask learning of...

Area of Science:

• Neuroscience and Biomedical Engineering
• Brain-Computer Interfaces (BCIs)
• Machine Learning for Signal Processing

Background:

• Motor imagery (MI) classification relies on extracting deep features from bioelectric signals like EEG and EMG.
• Developing effective models for MI tasks is a significant research challenge.

Purpose of the Study:

• To develop a multimodal multitask hybrid brain-computer interface network (2M-hBCINet) for enhanced motor imagery classification.
• To leverage deep features from both EEG and EMG signals for improved BCI performance.

Main Methods:

• A variational autoencoder (VAE) was used for unsupervised deep feature extraction from EEG and EMG signals.
• A channel attention mechanism (CAM) selected salient deep features.
• Multitask learning (MTL) trained the model on MI classification, EEG/EMG reconstruction, and feature metric learning tasks.
• Leave-one-out cross-validation (LOOCV) validated the model on custom and public datasets.

Main Results:

• The 2M-hBCINet model demonstrated superior performance compared to existing models.
• The model achieved optimal results across various frequency bands.
• Effectiveness was confirmed under muscle fatigue conditions.

Conclusions:

• The 2M-hBCINet model shows excellent performance and generalization capabilities for motor imagery classification using EEG and EMG data.
• This end-to-end model has potential applications in other EEG-related fields like anomaly detection and emotion analysis.