Domain-Specific Processing Stage for Estimating Single-Trail Evoked Potential Improves CNN Performance in Detecting...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Conventional Convolutional Neural Networks (CNNs) for Error Potential (ErrP) detection process raw electroencephalography (EEG) signals, including background noise, which can reduce accuracy.
• Existing methods struggle to effectively isolate ErrP signals from background EEG activity, hindering precise prediction of ErrP presence.

Purpose of the Study:

• To develop and evaluate a novel architecture for enhanced Single-Trial (ST) ErrP detection.
• To improve the predictive accuracy of ErrP signals by separating signal enhancement from classification stages.

Main Methods:

• Implemented a two-stage architecture: initial ST ErrP enhancement followed by CNN-based classification.
• Investigated various ST ErrP estimation techniques, including subspace regularization, Continuous Wavelet Transform, and ARX models.
• Evaluated different CNN classifiers such as EEGNet, standard CNN, and Siamese Neural Networks.

Main Results:

• The proposed architecture significantly outperformed the direct application of CNNs to raw EEG signals.
• Subspace regularization methods demonstrated the most substantial improvement in classification metrics.
• Achieved up to a 14% increase in balanced accuracy and a 13.4% increase in F1-score using the enhanced architecture.

Conclusions:

• The novel two-stage architecture effectively enhances ErrP signal detection by preprocessing EEG signals before classification.
• The combination of ST ErrP enhancement techniques, particularly subspace regularization, with CNNs offers a superior approach for ErrP detection tasks.
• This methodology provides a more accurate and reliable method for identifying ErrP signals in complex EEG data.