Datasets:

---

trialdesignbench

/

source

like 0

Follow

TrialDesignBench 7

What

Adds the ADVENT trial, absent from the corpus until now:

• documents/10.1056_nejmoa2307291/ — protocol + SAP, each with the full format set
  (.pdf, _mmd.zip, _tex.zip, .html, .lines.json)
• data/tdr.parquet — row 1642

Why it was missing

Not a download failure and not a paywall. The corpus reconciles exactly
(1641 rows <-> 1641 folders, zero *.error.txt), so the pipeline was healthy —
the row was simply never entered in the curation sheet, and sync_dataset.py is
sheet-driven and additive, so a trial nobody types in produces no row, no
folder, and no error.

The protocol and SAP were both posted on ClinicalTrials.gov the whole time, at
the exact URL pattern the pipeline already uses.

Notes for review

• Protocol revision. Prot_002.pdf, taken from the API's
  largeDocumentModule.largeDocs[].filename. Prot_000.pdf also returns HTTP
  200 but is a superseded revision — the CDN keeps serving old ones. Worth a
  broader audit: 1735 existing links use the _000 suffix.
• Trial Design: adaptive_bayesian;noninferiority. From the SAP: Bayesian
  primary analysis with non-informative priors, Goldilocks adaptive sample
  sizing (350-750), non-inferiority margins 0.08 (safety) / 0.15
  (effectiveness); superiority is a secondary endpoint only. Existing
  vocabulary value, used by 6 other rows.
• Lead Sponsor: Boston Scientific Corporation per ClinicalTrials.gov. The
  protocol says FARAPULSE, Inc. (acquired mid-trial); CT.gov was kept for
  consistency with the other 1641 rows.
• Statistician* left null. The posted SAP is unsigned (Author: ____) and
  the protocol names no statistician. 1204/1641 existing rows are null here.
• Schema change: the all-null Unnamed: 22 / Unnamed: 23 columns are
  dropped (24 -> 22). They were artifacts of trailing sheet columns. This
  rewrites every row group, so the parquet shows as a full-file change rather
  than a one-row append — no other row content is modified.

The same row still needs adding to the curation sheet, otherwise the next full
sync_dataset.py run regenerates the parquet from the sheet and drops it.