EEG foundation models often rely on masked prediction to learn from unlabeled recordings, yet optimizing this objective alone does not guarantee transferable neural representations. Stable positional cues and local correlations allow masked regions to be predicted without integrating distributed neural context, creating a shortcut‑learning problem.
To curb low‑information prediction paths we introduce Neural State Prediction (NSP), a latent‑predictive framework that simultaneously constrains the prediction target and the visible context. NSP employs a Target Encoder updated via an exponential moving average (EMA) to provide latent supervision.
Identity residualization removes additive effects tied to channel identity and relative time from the targets, while topology‑separated context excludes the immediate spatial‑temporal neighborhood of the target from the input, forcing the model to exploit broader signal structure.
We pre‑trained NSP on 2.2 million EEG segments from the TUEG dataset and evaluated it on 30 downstream tasks covering clinical diagnosis, sleep staging, emotion recognition, motor imagery, event‑related potentials, cognitive‑state decoding, and language retrieval.
Under full‑parameter multi‑task fine‑tuning on EEG‑FM‑Bench, NSP achieved a macro balanced accuracy of 63.94 across 14 datasets, surpassing the strongest baseline by 2.35 percentage points. Controlled component ablations quantified each mechanism’s contribution; matched‑context controls and held‑out interventions further clarified the roles of context geometry, signal content, and positional information.
These findings suggest that jointly designing latent targets and their visible context is a promising direction for EEG foundation models to better capture distributed signal structure.
Review