Toward general‐purpose foundation models for electroencephalography: A unified data registry

Le résumé fourni par la source

Abstract Electroencephalography (EEG) is widely used in cognitive neuroscience, clinical diagnosis, and brain–computer interfaces (BCIs). Recent work has begun to explore large‐scale EEG pretraining for transferable representation learning, which requires diverse and well‐organized EEG data across tasks and populations. Currently, public EEG datasets are scattered across platforms and publications and show substantial variation in experimental paradigms, recording settings, and metadata standards. This fragmentation makes large‐scale discovery, integration, and reuse inefficient, particularly for foundation model pretraining. To address this, we systematically screened publicly available EEG datasets updated from 2020 to 2026 and constructed a unified EEG dataset registry tailored to scalable pretraining and benchmarking. The primary output of this work was a registry with structured metadata enabling efficient dataset discovery, filtering, and direct retrieval for EEG foundation model pretraining. We reviewed more than 900 publications and curated 827 eligible datasets, organized under a six‐category taxonomy (cognitive, BCI, naturalistic, clinical, neuromodulation, and methodological). For each dataset, we recorded standardized metadata fields as reported, including task paradigm, device, channels, montage, sampling rate, number of participants, region, age, health status, license, data modalities, and label availability, together with embedded dataset and literature links to support direct retrieval. Using this curated inventory, we present a characterization of EEG resources, including domain imbalance and platform concentration, which highlights the difficulty of assembling corpora from sources. The registry offers centralized access and standardized descriptions, reducing the cost of discovery and cross‐dataset alignment and supporting the pretraining and evaluation of EEG foundation models.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.