aCRF Gen · Proof

Tested on a public study. Scored against its author.

We ran aCRF Gen on an independent, openly licensed test trial and compared every annotation with the SDTM mapping its author published. Nothing was tuned to this study, and the author's answers were removed from the input first.

77 / 77fields read and given a proposed mapping
35 of 42match the author's SDTM domain and variable
0failures against the FDA PDF Specifications v4.1
~5 minfrom upload to a draft aCRF
Side by side

The author's annotations, and ours.

The Adverse Events form. Left: the author's published SDTM annotations. Right: aCRF Gen's draft, drawn on the blank CRF beside each question.

The author's annotated Adverse Events page with SDTM annotations in a column
The author's aCRF
aCRF Gen's draft of the same page, with an annotation box beside each question
aCRF Gen draft

Pre-specified medications: each "Were any … taken?" question becomes CMOCCUR, and each "reason not prescribed" CMREASOC — on a layout the tool had never seen.

The blank Background Heart Failure Maintenance Medications form
Blank CRF
aCRF Gen's draft of the same form with CMOCCUR and CMREASOC annotations
aCRF Gen draft
Scorecard

Every field, counted.

The author annotated 42 of the 77 fields with SDTM; the other 35 have no annotation to compare against. We count a match only when the SDTM domain and variable agree.

ResultFieldsWhat it means
Same domain and variable25Including supplemental qualifiers sent to the same SUPP-- dataset, and Death Details findings, where the QNAM or test code we chose differs from the author's
Same domain and variable; author adds more10Pre-specified heart-failure medications: we give CMOCCUR and CMREASOC; the author also annotates CMCAT and CMPRESP
Different7Listed below — several are defensible alternatives, all are flagged for your reviewer
QuestionAuthoraCRF Gen
Location, laterality, directionality of psoriasis lesions (3)QS (lesion identification)CE.CELOC / CELAT / CEDIR
Has the subject had any psoriasis treatments?CMPRESP / CMOCCURNOT SUBMITTED
Was the ECG performed?NOT SUBMITTEDEG.EGSTAT
Is the adverse event still ongoing?AE.AEENTPTAE.AEENRF
Does the subject meet all inclusion criteria?SUPPDM (IEYN)IE.IEORRES
Not just one study

Three studies, three layouts.

StudyShapeFields mappedFDA PDF check
Independent ODM test trial (shown here)13 pages, 9 forms, numbered table layout77 / 770 failures
GSK Phase I study (Darapladib)51 pages, 12 forms, fillable PDF307 / 3070 failures
Academic oncology trial (HIPEC)One 226-field screening form across 15 pages226 / 2260 failures

The GSK and HIPEC CRFs come from the Portal of Medical Data Models; their licence does not allow commercial reuse, so we report results but do not reproduce them here.

Check it yourself

The files, unedited.

How we ran it

  • Inputs: the author's blank CRF and EDC export (CDISC ODM-XML), with every SDTM annotation the author had embedded in the export removed first.
  • One run on the live service, model Gemini on Google Vertex AI. The draft above is unedited.
  • Scored against the author's SDTM annotations field by field; the author's conventions are one valid choice among several, so some differences are judgement calls, not errors.
  • aCRF Gen produces a draft for qualified review. Your reviewer can change, remove, add or move any annotation in plain English, and every change is recorded in the audit trail.

Test Trial example © 2020 Jørgen Mangor Iversen, used under the MIT License.

Your turn

Try it on your own CRF.

Sign up and request a free trial. Once we switch it on, upload a blank CRF with its EDC export: your first aCRF is free — draft, review workbook, FDA PDF check and audit trail included.