A time-frequency feature fusion-based deep learning network for SSVEP frequency recognition.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-state visual evoked potential (SSVEP) is crucial for brain-computer interfaces (BCIs) due to its high signal-to-noise ratio and information transfer rate.
• Inter-subject variability in electroencephalographic (EEG) signals hinders SSVEP frequency recognition, especially in calibration-free settings, demanding extensive calibration data.

Purpose of the Study:

• To develop an improved deep learning network, SSVEP-TFFNet, that mitigates the need for large calibration datasets.
• To enhance cross-subject generalization for SSVEP-based BCIs through dynamic time-frequency feature fusion.

Main Methods:

• Proposed SSVEP time-frequency fusion network (SSVEP-TFFNet) with parallel time-domain and frequency-domain branches.
• Dynamic weighting mechanism to fuse extracted features, strengthening expression ability and generalization.
• Cross-subject classification performed on 12-class and 40-class SSVEP datasets.

Main Results:

• SSVEP-TFFNet achieved 89.72% accuracy on a 12-class SSVEP dataset, outperforming baseline methods by 1.83%.
• Achieved 72.11% and 82.50% accuracy on 40-class SSVEP datasets, surpassing controlled methods by 7.40% and 6.89% respectively.
• Demonstrated superior performance compared to traditional and principal deep learning approaches.

Conclusions:

• The dynamic time-frequency feature fusion strategy is effective in improving SSVEP classification.
• SSVEP-TFFNet offers a new paradigm for calibration-free SSVEP-based BCI systems, enhancing usability and performance.