Subject-independent EEG classification based on a hybrid neural network

Frontiers in Neuroscience, 2023 · DOI: 10.3389/fnins.2023.1124089 · Published: June 2, 2023

Simple Explanation

This paper introduces a new method for classifying EEG signals in brain-computer interfaces (BCIs) that doesn't require individual calibration. It uses a hybrid neural network approach. The method uses a filter bank GAN (FBGAN) to create more EEG data and a convolutional recurrent network to recognize motor imagery tasks. The proposed hybrid neural network improves subject-independent EEG classification performance through data augmentation and feature enhancement, improving the usability of the BCI system for new users.

Study Duration

Not specified

Participants

9 healthy subjects

Evidence Level

Level 5: Expert Opinion, and/or Case Series or Bench Research

Key Findings

1
The hybrid neural network achieved an average classification accuracy of 72.74 ± 10.44% in four-class tasks of BCI IV-2a.
2
The hybrid neural network framework improves subject-independent EEG classification performance to a conspicuous level through data augmentation and feature enhancement.
3
The proposed framework obtains the best classification accuracy compared to other state-of-the-art subject-independent classification approaches.

Research Summary

This paper presents a novel hybrid neural network framework for subject-independent EEG classification, which combines data augmentation using a filter bank GAN (FBGAN) with a discriminative feature network for motor imagery (MI) task recognition. The FBGAN is designed to acquire high-quality EEG data for augmentation by incorporating sparse CSP features from multiple sub-bands of filtered EEG data into the discriminator. The convolutional recurrent network with discriminative features (CRNN-DF) is proposed to extract distinguishable features from EEG signals with low signal-to-noise ratio to identify MI tasks.

Practical Implications

Improved BCI Usability

The hybrid neural network improves subject-independent EEG classification performance, enhancing the usability of BCI systems for new users.

Practical Application of BCI

The research offers a promising approach for facilitating the practical application of BCI, alleviating the mutual interference between different subject brain patterns and improving the accuracy of the EEG decoding process.

Data Augmentation Potential

The study demonstrates the potential of GANs in generating MI EEG signals and their utility for subject-independent classification.

Study Limitations

1
EEG signals vary greatly from subject to subject, so the standard deviation is also relatively large and the stability is not yet good enough.
2
The quality of the signals generated by FBGAN was not always perfect.
3
The FBGAN model is parallel to each category of each subject, which increases the computational cost.

Your Feedback

Was this summary helpful?

Back to Neurology