EEGDash: An open-source platform for machine learning on public neurophysiological data
Public recordings of electroencephalography (EEG), magnetoencephalography (MEG), intracranial EEG (iEEG), functional near-infrared spectroscopy (fNIRS) and electromyography (EMG) are easy to find and hard to train on. Turning a deposit into model input still means writing download, parsing, repair and windowing code for each dataset, and a dataset that meets a metadata standard can still fail to load. EEGDash is an open-source registry over public Brain Imaging Data Structure (BIDS) datasets from the OpenNeuro and NEMAR archives and from local deposits. Each dataset becomes a Python class; format defects are repaired lazily; signal handling, windowing and format checking are delegated to MNE-Python, Braindecode and the official BIDS validator. Semantic search, a web browser, publication-derived dataset tags and a feature-extraction module share the registry. EEGDash indexes the metadata of 892 public BIDS datasets (43,911 participants and 126,672 hours across the five modalities), loads any of them into a PyTorch dataset without custom data-loading code, and shows on 503 OpenNeuro EEG datasets that passing the BIDS validator does not predict whether a dataset loads (Spearman $ρ= -0.05$), so an archive built for machine learning has to track both.