BFRCNet: Addressing the class imbalance problem in the rapid serial visual presentation paradigm for decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Imbalanced sample sizes in electroencephalogram (EEG) analyses for rapid serial visual presentation (RSVP) tasks significantly reduce classification accuracy.
• Existing methods struggle to effectively handle the disparities in data distribution inherent in RSVP EEG datasets.

Purpose of the Study:

• To introduce BFRCNet, a specialized neural network architecture designed to improve EEG classification accuracy under imbalanced data conditions.
• To enhance the analysis of RSVP tasks by developing a robust method for handling class imbalance in EEG signals.

Main Methods:

• BFRCNet employs a three-stage architecture: feature representation, recombination, and classification.
• The feature representation stage utilizes a pyramid structure for multiscale spatiotemporal pattern integration, inspired by visual physiology.
• The recombination stage uses anchor samples to rebalance the data distribution, and the classification stage incorporates a novel focal loss function with class and sample weights.

Main Results:

• BFRCNet achieved balanced accuracy (BA) scores of 89.53% on the THU dataset and 90.15% on the CAS dataset.
• The proposed method significantly outperformed existing state-of-the-art techniques in addressing class imbalance for RSVP tasks.
• The novel focal loss function effectively prioritized minority samples, improving overall classification performance.

Conclusions:

• BFRCNet demonstrates superior performance in classifying imbalanced EEG data for RSVP tasks.
• The architecture provides a robust solution for enhancing classification accuracy in scenarios with significant class disparities.
• This work offers a promising advancement for EEG-based brain-computer interfaces and cognitive state analysis in RSVP paradigms.