INCEPT

Mask a second of scalp EEG and train a model to fill it back in, and much of what it learns to predict is not brain activity: the reference scheme, the amplifier's noise floor, a drifting electrode, the artifact signature of one recording room. Reconstruction-centered EEG foundation models such as CBraMod, CSBrain and CodeBrain treat all of that locally predictable content as equally worth representing. INCEPT inverts the premise: what deserves a place in the representation is whatever stays stable when the same recording is observed twice, at different spatial and temporal scales.

INCEPT — INvariance-oriented Contextual EEG Pre-Training — is a self-supervised transformer encoder from the School of Biomedical Engineering at ShanghaiTech University, with Shanghai United Imaging Intelligence and the Shanghai Clinical Research and Trial Center, posted as a preprint in August 2026. It is pretrained on roughly 11,000 hours of unlabelled routine clinical EEG retained from the Temple University Hospital EEG Corpus, then evaluated on ten downstream datasets grouped into three levels of post-acquisition analysis: signal-level assessment, brain-state decoding and brain-health evaluation.

The paper's second argument is that a transferable EEG encoder should be subject-sensitive rather than subject-agnostic. State and disease effects are often expressed as deviations from a subject's own baseline, so a representation that scrubs away individual physiology erases the signal with it.

#Key Features

• Cross-view invariance objective: A teacher encoder sees macro-level views of a 30-second segment while a student sees micro-level and masked macro-level views, and the two are aligned through an EMA teacher with Sinkhorn-Knopp centered targets. Masked contextual reconstruction is kept alongside it at equal weight rather than discarded.
• Spherical-harmonic electrode embeddings: Scalp positions are encoded as continuous spherical-harmonic basis functions plus a residual Cartesian pathway, instead of a learned embedding per channel identity. A montage the model never saw still has a well-defined position.
• Alternating channel and global attention: Transformer layers restricted to a single electrode's time sequence alternate with layers attending over all electrodes, building local temporal dynamics before integrating distributed context.
• Designed for frozen readout: A pooled whole-brain representation is meant to be usable by a single linear head, so linear probing is a primary evaluation protocol rather than a diagnostic afterthought.

#Technical Details

The encoder is 12 transformer layers over 768-dimensional tokens; each token is a one-second patch at 250 Hz, fused from a time-domain projection and an FFT amplitude spectrum through a learnable coefficient vector. Every 30-second segment yields two macro-level views and eight micro-level views, with masked macro views built by block-wise masking at ratios of 0.1 to 0.3, and the prediction head maps to an 8,192-dimensional distribution. Pretraining ran 250,000 AdamW iterations at a peak learning rate of 8e-4 on two 80 GB NVIDIA A100 GPUs. The corpus is filtered to the 19 common 10-20 electrodes, notch- and band-pass filtered, ICA-cleaned with ICLabel, and cut into non-overlapping 30-second windows.

Across the ten-dataset benchmark, INCEPT ranks first among the evaluated EEG foundation models on 26 of 30 linear-probing metrics and 24 of 30 fine-tuning metrics. Frozen, it reaches 81.71% balanced accuracy on TUAB abnormality detection, 58.30% balanced accuracy and 48.06 kappa on TUAR artifact recognition, 65.20% balanced accuracy on ADFTD dementia classification and 90.42% AUROC on Siena seizure assessment; fine-tuned, 84.75% balanced accuracy on FACED emotion recognition and 79.88% on ISRUC-S1 sleep staging. Seed-to-seed variance is markedly lower than for task-specific supervised encoders. A matched-architecture ablation isolates the objective: invariance learning alone improves on masked modelling alone by 15.5% on average across 12 metric-dataset pairs, with the largest gains on montage-shifted and multi-class transfer.

#Applications

The intended use is as a reusable backbone for the analysis that happens after an EEG is recorded — screening recordings for abnormality, flagging eye-movement and muscle artifacts for quality control, staging sleep from sparse derivations, decoding motor imagery for brain-computer interfaces, and assessing depression, mental stress, dementia and seizure activity. Because much of the reported performance survives freezing the backbone, groups with a modest labelled collection can fit a linear head rather than train an encoder from scratch, and the spherical-harmonic positional scheme means a new montage does not require a new channel vocabulary.

#Impact

INCEPT is a preprint awaiting peer review, and no code, weights or preprocessing scripts have been released — the paper states that code will be published upon acceptance, so the results cannot yet be reproduced and the encoder cannot yet be reused. Its contribution is an argument as much as an artifact: that reconstruction is an incomplete organizing principle for EEG pretraining, and that cross-view invariance supplies what it misses. The authors are explicit about the limits of the evidence — the benchmarks are retrospective public datasets rather than prospective clinical workflows.