A deep learning method for single-trial EEG classification in RSVP task based on spatiotemporal features of ERPs

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Single-trial electroencephalography (EEG) classification is crucial for rapid serial visual presentation (RSVP) tasks.
• Convolutional Neural Networks (CNNs) are effective for extracting RSVP EEG features.
• Existing CNN models often overlook the phase-locked characteristics of event-related potential (ERP) components.

Purpose of the Study:

• To propose a novel CNN model that better leverages the phase-locked characteristic of ERPs for single-trial RSVP EEG classification.
• To extract enhanced spatiotemporal features by learning spatial distributions of ERP components across different time periods.

Main Methods:

• A novel CNN architecture combining standard convolutional, permute, and depthwise convolutional layers is introduced.
• The model performs spatial convolutions separately across different time periods to exploit ERP phase-locking.
• Performance is benchmarked against traditional and deep learning methods, with spatial topography and saliency maps used for feature analysis.

Main Results:

• The proposed CNN model achieved superior classification performance compared to existing methods.
• Spatial topographies revealed typical ERP spatial distributions specific to different time periods.
• Saliency maps identified discriminant electrodes and meaningful temporal features crucial for classification.

Conclusions:

• The novel CNN model effectively incorporates phase-locked ERP characteristics for improved single-trial RSVP EEG classification.
• The model demonstrates excellent performance and provides insights into ERP feature extraction.
• This approach offers a promising advancement for analyzing complex EEG data in RSVP paradigms.