Subject-independent EEG classification based on a hybrid neural network

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Brain-computer interfaces (BCIs) offer direct communication pathways but often require extensive subject-specific calibration.
• Traditional subject-dependent BCIs pose challenges for users like stroke patients due to lengthy calibration needs.
• Subject-independent BCIs aim to reduce or eliminate pre-calibration, enabling quicker user access and broader applicability.

Purpose of the Study:

• To develop a novel fusion neural network framework for subject-independent electroencephalograph (EEG) classification.
• To enhance motor imagery (MI) task recognition in BCIs by improving EEG data quality and feature extraction.
• To provide a more time-efficient and accessible BCI system, particularly for new users and individuals with motor impairments.

Main Methods:

• A generative adversarial network (GAN), termed filter bank GAN (FBGAN), was designed for high-quality EEG data augmentation.
• Sparse common spatial pattern (CSP) features were extracted from filtered multi-band EEG data to preserve spatial information.
• A convolutional recurrent neural network with discriminative features (CRNN-DF) was employed for MI task recognition via feature enhancement.

Main Results:

• The proposed hybrid neural network achieved an average classification accuracy of 72.74% ± 10.44% on the BCI IV-2a dataset for four-class tasks.
• This performance represents a 4.77% improvement over existing state-of-the-art subject-independent classification methods.
• The FBGAN effectively augmented EEG data, contributing to enhanced recognition accuracy in the CRNN-DF model.

Conclusions:

• The developed fusion neural network framework offers a promising approach for subject-independent BCI systems.
• The novel combination of FBGAN for data augmentation and CRNN-DF for feature extraction significantly improves MI task recognition.
• This research facilitates the practical application of BCIs by reducing calibration time and enhancing user accessibility.