How well a large EEG foundation model represents neural activity depends on how the raw signal is tokenized. Existing methods, the authors say, impose arbitrary temporal boundaries that are misaligned with neural state transitions, neglect spatial information across channels, and use fixed segmentation criteria that fail to generalize across heterogeneous EEG paradigms. Their SE-DAGAF Adaptive Tokenizer (SEDAT) integrates four components into a single computationally efficient pipeline: squeeze-and-excitation based spatial aggregation, signal decomposition by data-adaptive Gaussian average filtering, instantaneous-frequency-guided adaptive segmentation, and Fourier-domain resampling.
The evaluation spans 10 heterogeneous EEG datasets covering motor imagery, mental imagery, P300, slow cortical potentials, sleep staging and epilepsy, run through four foundation models, LaBraM, EEGFormer, EEGPT and NeuroGPT, and benchmarked against five baselines: fixed-length windowing, CTXSEG, LiPCoT, TFM-Tokenizer and SiS. SEDAT improved classification by up to 15.3% over fixed-length windowing and by 1.2-4.6% over the next-best method, with all comparisons reaching statistical significance after correction, the researchers report. Token quality analysis showed Silhouette scores of 0.81-0.85 versus 0.33-0.48 for rigid baselines. The authors describe SEDAT as a physiologically grounded and computationally practical tokenization solution for large-scale EEG foundation models.