Research on a soft saturation nonlinear SSVEP signal feature extraction algorithm.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) using steady-state visual evoked potentials (SSVEP) offer high performance but depend heavily on calibration data.
• Current deep learning (DL) methods for inter-subject SSVEP classification show potential but require performance improvements compared to intra-subject methods.

Purpose of the Study:

• To develop an efficient deep learning network, e-SSVEPNet, for improved SSVEP signal recognition.
• To enhance the robustness and accuracy of SSVEP-based BCIs, particularly in scenarios with limited calibration data.

Main Methods:

• Proposed e-SSVEPNet, a novel deep learning model incorporating a soft saturation nonlinear module designed for noise robustness.
• Evaluated e-SSVEPNet on an SSVEP dataset using varying sliding time window lengths (1s, 0.5s) and training data sizes.
• Compared e-SSVEPNet against traditional and other DL baseline methods.

Main Results:

• The proposed e-SSVEPNet achieved the highest average accuracy for intra-subject classification on the SSVEP dataset.
• Demonstrated improved performance in SSVEP signal classification and recognition.
• Showcased higher decoding accuracy with shorter signal durations.

Conclusions:

• e-SSVEPNet significantly enhances SSVEP signal classification and recognition performance.
• The model exhibits strong potential for realizing high-speed SSVEP-based BCIs, even with limited data.
• The soft saturation nonlinear module contributes to improved noise robustness and overall accuracy.