Related Experiment Video
Updated: Oct 10, 2025
07:37
Assessment and Communication for People with Disorders of Consciousness
Published on: August 1, 2017
9.3K
Investigation on Robustness of EEG-based Brain-Computer Interfaces.
Summary
Deep learning models show better robustness against noise in electroencephalogram (EEG) brain-computer interface (BCI) systems. A new activation function improves performance, enhancing BCI usability.
Area of Science:
- Neuroscience
- Machine Learning
- Signal Processing
Background:
- Electroencephalogram (EEG)-based brain-computer interface (BCI) systems are susceptible to performance degradation caused by noise and artifacts in EEG data.
- Motor imagery (MI) classification is a key application for EEG-BCI, but its accuracy is often compromised by noisy data.
Purpose of the Study:
- To systematically investigate the robustness of machine learning (ML) and deep learning (DL) models for EEG-BCI motor imagery classification against simulated channel-specific noise.
- To compare the performance of different state-of-the-art models under varying low signal-to-noise ratio (SNR) conditions.
- To propose and evaluate a preliminary solution for enhancing the robustness of DL-based EEG-BCI models.
Main Methods:
- Simulated channel-specific noise was introduced into EEG data at various low SNR levels.
- The performance of traditional ML models (e.g., FBCSP) and DL models (e.g., EEGNet, Shallow ConvNet) was evaluated for motor imagery classification.
- A novel approach using saturating nonlinearities in activation functions was proposed to improve DL model robustness.
Main Results:
- Deep learning models demonstrated superior robustness to channel-specific noise compared to traditional ML models.
- EEGNet exhibited greater robustness to channel-specific noise than Shallow ConvNet and FBCSP.
- The proposed activation function modification significantly reduced classification accuracy drop at low SNR (-18 dB), decreasing it from 10.99% to 6.53% for EEGNet and 14.05% to 3.57% for Shallow ConvNet.
- Specific channels highly sensitive to simulated noise were identified.
Conclusions:
- DL-based models offer enhanced robustness for EEG-BCI motor imagery classification in noisy conditions.
- The proposed activation function strategy is a promising preliminary solution for improving DL model robustness in EEG-BCI.
- Further research is needed for more precise solutions to enhance EEG-BCI robustness and overall usability.
