EEG masked autoencoder pretrained twice, on EU/US archives and on a pool adding South Asian clinics, to test how pretraining data shapes transfer.
No providers recorded yet. Browse all providers
Nearly every EEG foundation model is pretrained on the same few public archives — the Temple University Hospital corpus, PhysioNet, MOABB — collected mostly in North American and European centers on a narrow range of amplifiers, then scored on benchmarks from the same ecosystem. A representation that has partly memorized one hospital system's recording characteristics can look excellent under that arrangement, and nothing in a standard benchmark table can tell it apart from one that has learned neural physiology.
PRISM (Population Representative Invariant Signal Model), from the EEG team at NeuroDx — the company whose later work appears under Mannas AI — and the Indian Institute of Technology Mandi, attacks that confound with a controlled ablation rather than a bigger corpus. Architecture, preprocessing and hyperparameters are fixed; only the pretraining population changes. One checkpoint sees TUH plus PhysioNet Motor Imagery, the other a pool that adds multi-center South Asian clinical recordings from several EEG systems.
The answer is a trade-off. Narrow-source pretraining gives stronger linear probes on benchmarks drawn from the same distribution, while diverse pretraining gives representations that adapt better once fine-tuned. The gap is widest on a clinical task the paper introduces: telling epilepsy apart from its mimickers, such as psychogenic non-epileptic seizures and syncope, from routine interictal EEG.
PRISM follows the masked-autoencoder design of REVE. Each channel is resampled to 200 Hz and cut into one-second patches with a 20-sample overlap, embedded by a convolutional projection to width 512. The encoder is a 12-layer pre-norm transformer with 8 heads and a GELU feedforward of expansion factor 4, processing only unmasked tokens; a 4-layer decoder of the same configuration fills masked positions with learnable tokens and reconstructs them through a linear head. Spatio-temporal block masking hides 55% of tokens within a 3 cm and 3 s radius of random seeds. The objective is an L1 reconstruction loss plus the auxiliary path's L1 loss at weight 0.1. All recordings are bandpassed to 0.5–99.5 Hz, notch-filtered, Z-scored per channel and mapped to the 19-channel 10–20 montage. The multi-source pool adds 9,663 subjects and 4,170 hours recorded on Natus, Nihon Kohden and BPL systems.
Across six benchmark tasks — ADFTD dementia, BCI-IV-2a and PhysioNet motor imagery, HMC sleep staging, Siena Scalp pathology and EEGMAT cognitive load — the narrow-source checkpoint leads under linear probing on four, and the multi-source checkpoint matches or exceeds it on five under full fine-tuning. On epilepsy versus mimickers, the multi-source checkpoint reaches 0.607 balanced accuracy against 0.484 for the narrow-source one, a 12.3-point gap. On motor imagery, PRISM tops LaBraM, CBraMod and REVE on three of four protocol–task combinations despite pretraining on three source corpora against REVE's 92 datasets. The OSF release, under CC BY 4.0, contains training, preprocessing and probing code and four checkpoint files: one narrow-source checkpoint and three multi-source variants.
The model is aimed at clinical neurophysiology, where the question is less "which benchmark" than "whose patients". The epilepsy-versus-mimicker task stands in for a real diagnostic bottleneck: up to a quarter of refractory epilepsy diagnoses are later found to be wrong. Groups building EEG classifiers for hospitals outside the TUH ecosystem can fine-tune the multi-source checkpoint, and teams comparing EEG models can use the protocol audit before trusting a ranking.
PRISM's lasting contribution is methodological. By holding architecture and objective fixed and varying only the pretraining population, it shows that corpus composition changes what a representation is good for, and that dataset count is a confounding variable whenever published models differ in architecture and data at once. Its warning about linear-probe-only evaluation, and its itemized account of why EEG-Bench and EEG-FM-Bench disagree, apply to any EEG foundation model. The work is an unreviewed preprint at a single model size; its multi-source pool confounds geography with acquisition hardware, and the epilepsy cohort comes from the same South Asian population the diverse pool adds, which may account for part of that checkpoint's advantage. The South Asian pretraining recordings are not part of the release, and the authors plan to release the epilepsy-mimicker dataset on acceptance.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.