The GLP-1 Data API

GLP-1 Real-World Evidence Brief — August 2026

A PDF brief plus the full dataset CSV — a dated corpus snapshot, currently 2026-09-15 with 2,109 rows, each carrying its source, date and extraction confidence, the drug where the source names one (552 rows name none) and, where a rate is a registry serious-adverse-event figure, a rateKind flag. 150 of those rows are April-2026 seed points, flagged in a "provenance" column and excluded from every published rate. Covers 15 side effects across 9 drugs, plus the 13 community rows whose own effect falls outside those 15 (kept in so the brief’s community reporting-frequency denominator is reproducible from the file — see the 17 September correction in the brief), drawn from PubMed, FDA FAERS, ClinicalTrials.gov, news coverage of EMA/MHRA/WHO safety actions, and public community reports, plus the full methodology annex. The CSV is a point-in-time snapshot, not a live feed; the free API stays current.

What this is not:incidence rates. Only a minority of sources state a numeric rate, so our per-effect estimates rest on small samples — as of 14 September 2026 no effect rests on a single rate-bearing source any more (hair loss did until that morning, and has since gained a third distinct source), one (emotional blunting) has no citable corpus rate at all and publishes no clinical number, and two rest partly or mostly on registry SERIOUS adverse-event rates (pancreatitis, 2 of its 3 sources; gallstones, 3 of its 10, as of 3 October 2026 — corrected 4 October 2026 from “gallstones, pancreatitis rest wholly or mostly”, which overstated gallstones: three of its seven sources on 21 September was already a minority), which are a floor on all-cause incidence and are labelled as such on every row (rateKind) — these memberships are recomputed daily, so read them from the API rather than from this sentence — and every estimate is published with its source count and a confidence grade, not as prevalence. The clinical and community tracks are kept separate and never blended; community reports carry self-selection bias. Treat the dataset as a sourced corpus and a discordance signal, not as an epidemiological estimate.

Known limitation of this snapshot: until 10 September 2026, our ClinicalTrials.gov registry collector kept only the five highest-percentage arm rows per trial — a selection on the very figure it measures. The collector was fixed that day and every affected trial is being re-collected in full. This 2026-09-15 snapshot postdates the fix and contains the re-collection landed to date, but the re-collection was still in progress when it was taken, so registry-derived figures in the file lag the live API. Current status: /data, and every /api/data response carries a samplingNote field until it is resolved. The previous 2026-09-02 file predated the fix and also carried 379 duplicate rows and 132 rows outside its stated 15-effect scope; it was superseded on 15 September and its defects are listed in the export errata shipped with every order.

Look before you buy: download a 31-row sample — every 70th row of the exact file, unmodified, with the same 14 columns. It is a systematic slice rather than a curated one, so 20 of its 31 rows carry a numeric rate, close to the 1,214 of 2,109 in the full file.