A hybrid adaptive EEG signal tokenization and vector embedding framework for interpretable Brain–Computer Interfaces

Despite recent advances, the generalizability of Brain–Computer Interface (BCI) decoding models is constrained by rigid tokenization and entangled embedding strategies. Current foundational models suffer from two major drawbacks: (1) blindly slicing continuous signals into fixed-length pieces destroys the structure of neural events. (2) Forcing diverse brain rhythms into a single embedding space erases the distinction between different biological states. Consequently, the learned representations lack physiological meaning, leading to poor clinical interpretability and reduced classification accuracy. This paper contributes a neurodynamic framework that jointly addresses both challenges. First, we propose the hybrid Dynamic Segment Tokenizer (DyST), which replaces fixed-patch tokenization methods by adaptively partitioning EEG based on its intrinsic dynamics. It generates tokens that preserve the integrity of complete neural events. Second, we propose the Band-Routed Adaptive Vector Embedding (BRAVE), which introduces a multi-codebook architecture that guides each token to a codebook corresponding to a specific neural frequency band (δ, θ, μ, β, γ). This creates a disentangled and interpretable representation. Comprehensive evaluations across seven BCI datasets confirm the framework's superiority, improving classification accuracy by up to 9.07% over the LaBraM benchmark. Our findings suggest that replacing rigid patching with dynamic, event-based tokenization and frequency-routed embeddings is worthy for building BCI systems that are both robust across datasets and clinically meaningful.