A Transfer Generalization Framework for Improved SSVEP-Based BCI Pattern Recognition

Steady-State Visual Evoked Potential (SSVEP) is a widely adopted paradigm in Brain–Computer Interfaces (BCIs), known for its high information transfer rate (ITR) and minimal training requirements, yet conventional methods often depend on subject-specific calibration, creating a major barrier to scalability and real-world use. To address this, we propose the Transfer Generalization Network (TransGen), a robust calibration-free framework that enhances cross-subject generalization in SSVEP decoding. TransGen incorporates Unified Pattern Templates and Cross-Subject Feature Filters (CSFF) to extract transferable spatial–spectral features across individuals, combined with harmonic component extraction and data augmentation for improved robustness. Experimental evaluations were conducted on the Benchmark dataset (35 subjects) and BETA dataset (70 subjects) using a leave-one-block-out protocol. Results show that TransGen consistently surpasses state-of-the-art approaches: on the 64-channel Benchmark it achieves 98.8% accuracy versus 96.3% for the best baseline, and on the 64-channel BETA it reaches 90.0% accuracy compared to 85.6%. Furthermore, under short time windows (0.2–0.5 s), TransGen improves ITR by 5–10%, achieving a maximum of 261.5 bits/min compared to 257.8 bits/min. These findings demonstrate that TransGen offers a scalable, user-independent, and high-performance solution for practical SSVEP-based BCIs, with strong potential for deployment in communication, rehabilitation, and assistive technologies.