SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction

Electroencephalography (EEG) provides a non-invasive measure of ongoing neural activity, but building general-purpose EEG models remains challenging due to the heterogeneity of subjects, devices, and electrode montages. Existing EEG foundation models predominantly rely on input-space reconstruction, which can bias the encoder toward memorizing noise and artifacts. We introduce SPERA (Spherical Prior EEG Representation Architecture), an EEG foundation model that predicts in latent space following the joint-embedding predictive architecture (JEPA). SPERA combines three components tailored to EEG: (i) a hybrid attention backbone that interleaves factorized temporal and spatial attention with periodic full-attention layers, (ii) a Legendre-polynomial spatial prior incorporated into attention to encode varying scalp electrode geometries, and (iii) a relational spectral regularizer that aligns latent similarity structure with spectral views, inducing frequency-aware latent representations. Pretrained on approximately 90,213 hours of EEG from 31,771 subjects across 106 datasets, SPERA achieves the highest average performance across nine downstream tasks spanning clinical, cognitive, and BCI applications. SPERA further exhibits strong parameter efficiency and robustness across varying recording conditions, suggesting its potential as a general-purpose backbone for diverse EEG analyses.