KAST-BAR bridges EEG signals and medical language across 21 datasets
Researchers propose a brain autoregressive model that bridges low-level EEG signals and high-level medical semantics through dynamic topology modeling and expert-anchored text profiles.
KAST-BAR is a foundation model that aligns EEG signal topology with expert-level medical language. The model addresses two persistent gaps in neural decoding: the difficulty of modeling brain signals' non-Euclidean spatial structure over time, and the semantic distance between raw physiological data and the clinical text practitioners use to describe it. Pre-trained on 21 EEG datasets, KAST-BAR uses a Dual-Stream Hierarchical Attention encoder to capture local temporal dynamics alongside global spatial context, then synthesizes instance-level textual profiles through a Knowledge-Anchored Semantic Profiler that grounds descriptions in medical knowledge. A Semantic Text-Aware Refiner reconstructs EEG representations using Latent Expert Queries, dynamically updating signal embeddings to match the semantic space of clinical language.
The architecture's three-stage pipeline—topology encoding, knowledge-anchored profiling, and text-aware refinement—lets the model generalize across tasks without task-specific tuning. In evaluations, KAST-BAR outperformed existing EEG foundation models on six downstream benchmarks, including emotion recognition, motor imagery classification, and seizure detection. Code and model weights are available at github.com/KAST-BAR/KAST-BAR.