A multibranch CNN framework based on time-frequency feature fusion for SSVEP detection

This is the authors' abstract. We don't add key points for this paper.

Steady-state visual evoked potential (SSVEP) has emerged as a prominent paradigm in brain-computer interface (BCI) research due to its non-invasiveness, high signal-to-noise ratio, and robust time-locking/phase-locking properties. Current limitations in both complex frequency-encoded stimulation methods and neural networks constrained to single-domain feature extraction necessitate the development of advanced feature fusion frameworks. Although existing deep learning approaches attempt to integrate multi-domain features, they largely overlook the subject-specific variability in feature-domain contributions to classification outcomes. To comprehensively extract discriminative information while enhancing user adaptability, we proposed a novel filter bank multi-branch convolutional neural network (FB-MultiCNN) architecture for SSVEP signal detection. The methodology simultaneously extracted both time-domain and frequency-domain features with an adaptive linear-weighted feature fusion mechanism. Offline evaluation on public benchmarks achieved an accuracy of 85.51 ± 0.65% and information transfer rate (ITR) of 166.27 ± 3.21 bits/min within a 1.0-s window. Real-time online validation further demonstrated practical efficacy, yielding 96.70 ± 0.92% accuracy and 199.49 ± 4.17 bits/min ITR. By simultaneously extracting time-domain and frequency-domain features, FB-MultiCNN significantly improves the performance of SSVEP-based BCIs.