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

Abstract

Imbalanced sample sizes in rapid serial visual presentation (RSVP) can substantially compromise the classification accuracy of electroencephalogram (EEG) analyses based on RSVP system. Here, we propose BFRCNet, a specialized neural network architecture designed to enhance classification under imbalanced EEG data conditions. This architecture comprises three stages: feature representation, recombination, and classification. During feature representation, a pyramid structure integrates multiscale spatiotemporal patterns while mimicking visual physiological mechanisms to enhance EEG feature extraction. The recombination stage incorporates anchor samples as auxiliary categories, transforming the imbalanced distribution into a balanced representation. The classification stage leverages a novel focal loss function that integrates class and sample weights, thereby enhancing the reward for minority samples. BFRCNet was evaluated on the THU and CAS datasets and achieved balanced accuracy (BA) scores of 89.53% and 90.15%, respectively, significantly outperforming state-of-the-art methods in addressing class imbalance problems in RSVP tasks.