IMH-Net: a convolutional neural network for end-to-end EEG motor imagery classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) technology relies heavily on electroencephalography (EEG) classification algorithms.
• Existing algorithms often require subject-specific calibration, limiting their practical application for new users.

Purpose of the Study:

• To propose IMH-Net, an end-to-end subject-independent model for EEG-based BCI.
• To improve the accuracy and robustness of motor imagery classification in a subject-independent manner.

Main Methods:

• Utilizes Inception blocks for extracting frequency domain features from EEG data.
• Employs feature vector compression to extract spatial domain features.
• Applies a Multi-Head Attention mechanism for learning global information and classification.

Main Results:

• Achieved 73.90±13.10% accuracy and 73.09±14.99% F1-score on the OpenBMI dataset in a subject-independent setting.
• Demonstrated a 1.96% accuracy improvement over comparison models on the OpenBMI dataset.
• Obtained state-of-the-art accuracy and F1-score on the BCI competition IV dataset 2a in a subject-dependent manner.

Conclusions:

• The proposed IMH-Net model significantly enhances the accuracy of subject-independent motor imagery (MI) classification.
• The algorithm exhibits high robustness, indicating strong practical value for BCI applications.
• IMH-Net reduces the need for extensive user-specific calibration, facilitating broader BCI adoption.