A library of Parkinson’s research, read at machine scale

How the information is treated

Nothing here is read by hand first. An automated pipeline runs every week against PubMed, Europe PMC and ClinicalTrials.gov, including searches written deliberately to surface evidence against the prevailing view — a corpus assembled only from confirming results is a corpus that will confirm anything. What comes back then passes through four stages before any of it can inform a sentence on this site.

1. Classification

   Every record is graded into a nine-level evidence hierarchy — practice guideline, synthesis, randomised trial, other trial, observational study, narrative review, primary report, preprint, trial registration — from the publication types MEDLINE’s own indexers assigned it. The grade is read off a label, not inferred from a title, so a paper cannot be promoted by the confidence of its abstract.

2. Full-text extraction

   For the core corpus the pipeline goes past the abstract: methods, results and limitations are pulled from Europe PMC’s open-access collection and from PDFs Unpaywall finds legally free. An abstract tells you a trial was positive. The methods tell you it was open-label with twenty-four participants.

3. Indexing and retrieval

   Each record is rendered to a structured document and loaded into a Gemini/NotebookLM notebook, where it is embedded and indexed for semantic retrieval. Questions are answered by retrieval-augmented synthesis over that index alone — the model is not asked what it knows, it is asked what these documents say, and every answer comes back carrying the passages it drew on.

4. Continuous re-checking

   A monthly sweep re-examines records the library already holds, oldest confirmation first, and corrects the ones whose post-publication record has moved: retractions, errata, expressions of concern. 79 records have been flagged this way and no longer count as evidence here.

The database

Behind the library is a registry with one row per record, keyed by whatever identifier the record actually has — a DOI for a preprint, a PMID for an indexed paper, an NCT number for a registered trial. Each row carries two fingerprints: a hash of the title and abstract, which catches a revised text, and a hash of the retraction, erratum and comment links, which catches a paper whose standing changed while nothing about its content did.

That registry is what makes the pipeline idempotent. The same searches run every week return mostly the same thousands of records, and the registry is how the fiftieth run adds only what is new instead of a fiftieth copy of everything. It is also partitioned by purpose rather than piled into one index:

• Core3,925guidelines, syntheses and trials — the corpus this site publishes from
• Field45,964everything else peer reviewed, held for breadth
• Watchpreprints and registered trials, rolling — neither is evidence of any strength yet

A retrieval system that cannot tell a preprint from a randomised trial will cite the preprint, so the separation is enforced in storage rather than left to the model to respect at query time. Which partition a record belongs to is decided once, on first sighting, and never re-derived — otherwise a paper re-indexed upstream would migrate between notebooks and leave a copy of itself behind in the old one.

The counts above are read at build time from the registry itself, through a function that can report how many records are held and never which. They are measurements, not marketing.

The meta-studies

Single trials disagree with each other; that is what single trials do. The documents worth the most weight are the ones that have already reconciled the disagreement — systematic reviews and meta-analyses, which pool the trials and quantify the pooled result, and the practice guidelines built on top of them.

The library currently holds 2,130 meta-analyses and systematic reviews and 26 clinical guidelines, alongside 1,381 randomised controlled trials. These sit at the top of the hierarchy and are retrieved first, so an explainer here is built from the syntheses and reaches for an individual trial only where no synthesis covers the question. Where a meta-analysis and a headline-grabbing trial disagree, the meta-analysis is what gets written up — and the disagreement gets written up with it.