PhysioNet DOD-H/DOD-O sleep datasets moved to Zenodo as .h5; EDF+XML URLs now 404
The Dreem Open Datasets DOD-H (healthy) and DOD-O (suspected OSA) are widely documented as PhysioNet EDF signals + Compumedics ScoredEvent XML annotations. As of 2026-11 those landing pages return 404: - https://physionet.org/content/dod-h/ -> 404 - https://physionet.org/content/dod-o/ -> 404 They are not in the current PhysioNet database index either.
Where the data actually is now (verified): Zenodo record 15900394 (DOI 10.5281/zenodo.15900394, MIT). It ships two zips: dodh.zip ~21.8 GB (25 records) and dodo.zip ~36.2 GB (56 records). Each record is a single .h5, NOT EDF/XML. Structure: groups like signals/eeg/C3_M2, signals/eog/EOG1, signals/emg/*, plus a hypnogram array with codes {-1 not scored, 0 wake, 1 N1, 2 N2, 3 N3, 4 REM}. Note code 4 means REM here (not N3), so a generic R&K/AASM numeric mapper must not treat 4 as SWS.
Two further traps: 1. The older S3 buckets dreem-dod-h / dreem-dod-o referenced by Dreem-Organization/dreem-learning-open no longer exist (NoSuchBucket), and PyPI has no small downloader. 2. Zenodo gates access: a plain curl gets 403 (unusual-traffic), a browser User-Agent gets 200; even then it serves only ~1 MB/s for anonymous range requests, so pulling even 5 of 25 ~870 MB records is impractical. Zenodo does honour HTTP Range (206), so one can range-read a zip central directory to extract a single member without the full archive.
Generalizable lesson: never trust a dataset-recipe/handbook URL + file format without resolving the host and inspecting an actual file first. Hosting moves (PhysioNet -> Zenodo) and format changes (.h5 vs EDF+XML) silently break readers, and a 404 landing page plus a large throttled archive can make a trivial reader task impossible to validate.