Abstract
Deep learning has significantly advanced brain-computer interface (BCI) technology. However, most deep learning models operate as black boxes, limiting their clinical applicability and scientific interpretability. This lack of transparency makes it difficult to determine whether predictions are driven by genuine neural activity or artifacts. To address this limitation, we propose Analformer, a novel Transformer-based architecture designed to achieve both high predictive performance and neuroscientific interpretability. The core component of Analformer is an Analytical Patch Embedding module, which employs fixed, non-trainable Morlet wavelet kernels to extract explainable spatio-temporal-frequency features from raw EEG signals. This structure enables standard neurophysiological analyses-including time-frequency analysis, topography, and F-value time-frequency (FTF) analysis-to be derived directly from the model's internal representations. Furthermore, by analyzing attention weights over these interpretable features, Analformer provides attention-based connectivity that may reflect functional relationships between brain regions. We evaluated Analformer on two large public datasets covering three representative BCI paradigms: Motor Imagery (MI), event-related potentials (ERP), and steady-state visually evoked potentials (SSVEP). Experimental results demonstrate that Analformer achieves competitive performance across all paradigms while producing analysis outputs consistent with established neuroscientific findings. These results suggest that Analformer provides a unified framework that bridges high-performance BCI decoding with interpretable, data-driven scientific analysis.