Dual-Branch Attention-Based Frequency Domain Network for Cross-Subject SSVEP-BCIs

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential-based brain-computer interfaces (SSVEP-BCIs) offer high-speed human-computer interaction.
• Current frequency-domain decoding methods inadequately utilize spectral features, limiting cross-subject performance.

Purpose of the Study:

• To develop a novel network, DB-AFDNet, for improved SSVEP-BCI cross-subject classification.
• To independently decode real and imaginary spectral components for more discriminative features.

Main Methods:

• Proposed a Dual-Branch Attention-Based Frequency Domain Network (DB-AFDNet).
• Implemented inter-branch attention similarity and intra-branch orthogonality constraints.
• Evaluated performance on Benchmark and Beta datasets for cross-subject classification.

Main Results:

• DB-AFDNet achieved superior cross-subject classification accuracy compared to state-of-the-art methods.
• Demonstrated relative improvements of 1.36% and 1.45% on the Benchmark and Beta datasets, respectively.
• The proposed constraints effectively learned consensus and discriminative features.

Conclusions:

• DB-AFDNet enhances feature generalizability and classification accuracy in SSVEP-BCIs for cross-subject applications.
• Independent decoding of spectral components and attention mechanisms are crucial for robust BCI performance.