A unified fixed-wavelet transformer decodes eight EEG classes across motor imagery, P300, and SSVEP paradigms

Introduction Motor imagery (MI) can support self-initiated control, whereas brain–computer interfaces (BCIs) based on the P300 or the steady-state visual evoked potential (SSVEP) rely on ongoing visual stimulation. Repetitive visual stimulation can cause visual fatigue; when the stimuli are suspended, however, the corresponding P300 and SSVEP responses can no longer be elicited, and these paradigms alone cannot support a self-initiated command to restart visual stimulation. By contrast, self-initiated MI can provide a voluntary gating signal for switching visual stimulation on or off. Developing a single architecture capable of learning these heterogeneous EEG signatures and decoding multiple user intentions is therefore an important step toward a voluntarily controlled hybrid BCI. Methods To this end, we developed a unified model by extending our previously proposed Analformer, a Transformer-based architecture designed for high classification performance while enabling the analysis of neurophysiologically meaningful EEG representations through AnalNet. We evaluated whether the model could classify EEG trials into eight mutually exclusive classes drawn from MI, P300, and SSVEP paradigms in the OpenBMI dataset. Analformer uses fixed Morlet wavelet kernels to extract frequency-specific EEG features, six Transformer blocks model dependencies across 31 pooled temporal tokens, and a two-stage training procedure to capture both population-level and participant-specific features. Results Across 54 participants, the mean accuracy obtained from the model after participant-specific fine-tuning was 71.00%, compared with 54.53% for EEG-Conformer, 52.30% for EEGNet, and 45.35% for ShallowConvNet under the same evaluation protocol. Ablation analyses showed that replacing the fixed wavelets with learnable convolutions reduced the mean accuracy to 63.15%, whereas omitting the two-stage training procedure reduced it to 47.38%. AnalNet, the built-in analysis module of Analformer, identified class-discriminative patterns, including sensorimotor mu- and beta-band activity during MI and stimulus-locked occipital responses during SSVEP. Discussion Together, these findings indicate that a single Transformer-based architecture can distinguish heterogeneous EEG classes and provide a foundation for unified decoding in hybrid BCIs that combine self-initiated MI gating with P300- and SSVEP-based command selection.