Adoption of Deep Learning Methods for SSVEP Classification in XR-Based Wearable Brain-Computer Interfaces

This paper addresses steady-state visually evoked potential (SSVEP) classification in wearable extended-reality (XR) brain-computer interfaces (BCIs), with a threefold objective. First, it investigates the effectiveness of deep learning (DL)-based SSVEP classification under XR stimulation, where platform-dependent rendering, optical see-through visualization, reduced luminance contrast, and interaction with the real environment may degrade the quality of the elicited EEG response. Second, a metrology-based performance assessment is proposed according to the Guide to the Expression of Uncertainty in Measurement (GUM), with classification accuracy and information transfer rate (ITR) expressed as best estimates with associated standard uncertainties. Finally, EEG channel reduction is analyzed toward lightweight XR-SSVEP implementations. As a representative SSVEP-specific DL model, the SSVEP time-frequency fusion network (SSVEP-TFFNet) is evaluated on an open XR benchmark dataset comprising 30 subjects and 1200 trials acquired using Microsoft HoloLens 2. A subject-independent comparison with filter-bank canonical correlation analysis (FBCCA) is performed, while intra- and inter-subject variability are incorporated into the uncertainty evaluation. Results show that SSVEP-TFFNet outperforms FBCCA under the considered XR conditions. Moreover, reduced 6- and 4-channel configurations preserve performance close to the full 8-channel montage. These findings provide evidence of the potential of suitably selected DL models for XR-based SSVEP classification and support uncertainty-aware, reduced-electrode wearable implementations.