POTS and dysautonomia literature findings

A public, machine-extracted research dataset and static browser: findings from the POTS, dysautonomia and related-condition literature, with verbatim source quotes, released as five verifiable data packages. Built with condition-findings-builder. Project index: condition-findings. Sister dataset: eds-findings.

Findings packages v1.2 (the full corpus)

The complete harmonised corpus: 148,186 findings from 11,519 papers across Tiers 1–4: ships as five self-contained packages under packages/ (start at packages/README.md and packages/OVERVIEW.md). Each package carries its data (Parquet + CSV), documentation, QA tables, the code that built it and CHECKSUMS.sha256; code/verify_package.py --self-test re-checks a package end to end.

Download and cite

Each package is archived on Zenodo with its own DOI. The "all versions" DOI always resolves to the newest version.

package	contents	size	DOI (v1.2)	DOI (all versions)
1 of 5	POTS and dysautonomia core literature	33.6 MB	10.5281/zenodo.23671720	10.5281/zenodo.23671719
2 of 5	EDS and hypermobility, ME/CFS, MCAS, fibromyalgia, IBS, small fiber neuropathy and RCCX genes, part 1	22.9 MB	10.5281/zenodo.23671722	10.5281/zenodo.23671721
3 of 5	the same conditions, part 2	20.5 MB	10.5281/zenodo.23671724	10.5281/zenodo.23671723
4 of 5	long COVID, post-viral and autism intersections	16.8 MB	10.5281/zenodo.23671726	10.5281/zenodo.23671725
5 of 5	migraine, hypermobility spectrum, Sjögren's, PANS/PANDAS	10.5 MB	10.5281/zenodo.23671728	10.5281/zenodo.23671727

The same five files are attached to the v1.2 GitHub release. Each package's CITATION.cff gives the citation for that package.

Paper flags. Every package carries annotations/paper_flags.csv, which marks retracted papers (10 across the corpus, 139 findings), correction notices extracted as papers, papers with a published correction, and preprints whose published version is also in the dataset. Rows are flagged, never removed: exclude or count them once as the file recommends. Checked against PubMed and Crossref on 2026-10-26.

version	tag	what changed
v1.2	v1.2.0	Paper flags added (retractions, correction notices, duplicate versions; see above). Every model-emitted ontology identifier in Tiers 1–2 was looked up by id at its minting authority (HGNC, ChEBI, MeSH) and compared with the entity name. 14,055 rows are now authority_verified_retrospective (10,879 confirmed, 3,176 replaced by the id the name resolves to); 2,607 ids that denoted a different entity were withheld. The original model id is preserved in ontology_id_v11_model_emitted. No finding text, quote, direction, condition or paper changed.
v1.1	v1.1	Five harmonised packages; Tier 1–2 identifiers flagged unverified_model_emitted.

Join on entity_ontology_id_resolved only where ontology_id_verification is authority_verified or authority_verified_retrospective (19,782 rows). Scientific fidelity: whether each finding states what its paper found: has not yet been checked by a human reader; that audit is open and its result will be posted here and on the site's About page. The browser at pots.erinmagennis.com serves this release.

Browser

The findings browser at pots.erinmagennis.com is a static projection of the packages above (search, entity and condition pages, derived tables, chart builder). It is regenerated from packages/ by site_tools/build_v12_projection.py; every count shown on the site and in this README is read from the generated site/data/meta.json, never typed. Run it locally:

python -m http.server 8000 --directory site      # then open http://localhost:8000/

Naming prefix: pots_. Browser storage keys start with pots_, and every analytics event carries app: 'pots_findings' (Chat page events are ask_provider_selected, ask_key_saved, ask_question_sent; they never include a key, a question or an answer).

Quick start with the data

import pandas as pd
t1 = pd.read_parquet("packages/dysautonomia-core-findings/data/tier1_findings.parquet")
ok = t1.ontology_id_verification.isin(["authority_verified", "authority_verified_retrospective"])
grounded = t1[ok]            # rows whose entity identifier was confirmed at its authority

Each package's README.md has a quickstart, docs/DATA_DICTIONARY.md every column, docs/LIMITATIONS.md what the data cannot support, and code/verify_package.py --self-test re-checks the package end to end.

Verification

python site_tools/verify_site_v12.py --self-test     # site projection vs packages, hand-typed-number scan, checksums
for p in packages/*/; do (cd "$p" && python code/verify_package.py --self-test && shasum -a 256 -c CHECKSUMS.sha256 --quiet); done

Scientific fidelity (does each finding state what its paper found) has not yet been human-checked; see the About page.

Historical checkpoints (archive)

Before the packages existed the repository published an incremental record: data/ (Tier 1, July 2026) and data/tier2/ (Tier 2 at 55%, August 2026), with docs under docs/ and docs/tier2/ and the site generator of that era in site_tools/site_generator.py. They are kept unchanged for reproducibility of the earlier releases (tags tier1-v1, tier2-55pct-2026-08-03); the packages supersede them. The package versions are tags v1.1 and v1.2.0.

Important distinctions

• Co-mention is not interaction or causation.
• Shared biology across literatures is not the same as patient comorbidity.
• A condition label is not the same as the source literature that supplied the paper.
• Entity-based views cover the 60.3% of findings that name an entity, not the whole dataset; the rest are symptom, phenotype and quality-of-life findings.
• A null result does not by itself establish ineffectiveness.

Citation and license

Derived structured data is licensed under CC BY 4.0. Verbatim provenance text remains the copyright of its respective authors and publishers and is not licensed by this project. See NOTICE.md and CITATION.cff.

Credit: Erin Magennis.