EEGDash opens 86,000+ hours of brain data for ML with no custom code

EEGDash, now available on arXiv, tackles one of the biggest bottlenecks in computational neuroscience: turning raw public datasets into machine-learning-ready formats. The platform catalogs 791 recordings from 39,778 participants totaling over 86,051 hours, spanning five modalities (EEG, MEG, iEEG, EMG, fNIRS) from the OpenNeuro and NEMAR archives. Each dataset is exposed as a Python class that preserves signal attributes and integrates directly with MNE-Python for signal handling, Braindecode for windowing, and the official BIDS validator for format compliance. This eliminates the thousands of lines of custom download, repair, and pre-processing code typically required.

The system includes a metadata-first registry with semantic search, an automatic format-repair layer, and dataset-level tags extracted from the source publication. A feature-extraction framework enables rapid benchmarking and cross-dataset analysis. By providing per-record loadability and compliance metadata, EEGDash allows researchers to focus on model development rather than data wrangling. The project is open-source and aims to lower the barrier for applying machine learning to neurophysiological data at scale.