FB-EEGNet: A fusion neural network across multi-stimulus for SSVEP target detection

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential (SSVEP) is a key brain-computer interface (BCI) paradigm.
• Deep neural networks (DNNs) are increasingly used for SSVEP target recognition.
• Existing DNNs struggle to fully utilize SSVEP harmonic components and ignore non-target stimuli.

Purpose of the Study:

• To develop a DNN model for SSVEP target detection that incorporates information from multiple sub-bands and non-target stimuli.
• To improve the accuracy and efficiency of SSVEP-based BCIs.

Main Methods:

• Proposed FB-EEGNet, a DNN model that fuses features from multiple neural networks.
• Designed a multi-label approach for each sample.
• Optimized FB-EEGNet parameters across multi-stimulus data to leverage non-target stimulus information.

Main Results:

• FB-EEGNet achieved high classification accuracies and information transfer rates (ITRs) in both subject-specific and cross-subject conditions.
• Subject-specific accuracies reached 89.14% (70.45 bits/min) on a 9-target dataset (0.7s window).
• Cross-subject accuracies reached 92.15% (76.12 bits/min) on the same dataset (1s window).

Conclusions:

• FB-EEGNet demonstrates superior performance compared to existing methods like CCNN, EEGNet, CCA, and FBCCA.
• The model effectively extracts information from multiple sub-bands and cross-stimulus targets.
• FB-EEGNet offers a promising approach for deep feature extraction in SSVEP using neural networks.