BRENNEN

Industries / Health / Pharmaceuticals / Clinical research

Capturing data

Recording every measurement in a trial so that it can be trusted, from paper forms to electronic systems, phone apps, sensors and hospital records, and passing it on in formats regulators can read.

7

Milestones achieved, from 1997 to 2023.

3

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1997

    Electronic records made legal

    The FDA's rule 21 CFR Part 11, published in March 1997, set the conditions under which electronic records and signatures count as trustworthy and equivalent to paper, the basis for replacing the paper case report forms on which each participant's data were written.

    LessonIn 2003 the FDA said the rule was being read so broadly that it held back electronic records and raised costs "without providing a significant public health benefit", and narrowed how it would enforce it.

    law.cornell.edufda.gov

    Credit

    FDA

  2. 2004

    Free data forms for research

    Vanderbilt University built REDCap in 2004, a web tool researchers use to build their own forms for collecting study data, and from 2006 shared it free with non-profit institutions. By REDCap's own counters on 18 November 2026 it was in use at 8,536 institutions in 167 countries, and its papers had 58,900 citations.

    projectredcap.orgapi.crossref.org

    Credit

    Paul Harris and colleagues, Vanderbilt University

  3. 2015

    Phones as research tools

    Apple released ResearchKit, open-source software for running studies on iPhones, in March 2015. The mPower study of Parkinson's disease, built on it, enrolled 9,520 people who agreed to share their data broadly, with no clinic visits.

    LessonSigning up was easy and staying was not: 898 of the 9,520 gave data on at least five days in the first six months.

    apple.compmc.ncbi.nlm.nih.gov

    Credit

    Apple; Sage Bionetworks and the University of Rochester

  4. 2016

    One format for trial data

    For studies starting from 18 December 2016, the FDA required drug applications to send trial data in standard formats, in practice CDISC's SDTM for the data as collected and ADaM for the data as analysed; Japan's regulator also requires CDISC standards. CDISC is the Clinical Data Interchange Standards Consortium, a non-profit standards body.

    fda.govcdisc.org

    Credit

    FDA; CDISC

  5. 2020

    Data no one could audit

    A Lancet study of hydroxychloroquine for Covid-19, built on hospital-record data supplied by the company Surgisphere, was retracted in June 2020 after Surgisphere would not give independent reviewers the full dataset. It was an observational study, not a trial.

    LessonData drawn from health records count as evidence only if someone outside the supplier can trace them to their source.

    pmc.ncbi.nlm.nih.gov

    Credit

    Surgisphere, which supplied the data; the retraction by authors Mandeep Mehra, Frank Ruschitzka and Amit Patel

  6. 2021

    A trial run on records

    The ADAPTABLE trial of two aspirin doses enrolled 15,076 patients at 40 US centres, found many of them by searching hospital records, and took their medical history from those records in a shared format, the PCORnet Common Data Model. Medical history came this way for 99.7% of patients in one group and 99.8% in the other.

    pmc.ncbi.nlm.nih.gov

    Credit

    W. Schuyler Jones, Adrian Hernandez and the ADAPTABLE team; PCORnet; the Patient-Centered Outcomes Research Institute, which paid for it

  7. 2023

    A sensor reading as endpoint

    EMA's CHMP accepted "stride velocity 95th centile" (SV95C), the speed of a child's fastest strides measured at home by a sensor worn on the ankle, as a primary endpoint, the main measure of success, in trials for boys with Duchenne muscular dystrophy who can still walk. It is an alternative to the six-minute walk test done in a clinic, and had been accepted as a secondary measure in 2019.

    ema.europa.eu

    Credit

    EMA's CHMP; Sysnav, named in the opinion as the sensor's supplier

Watch for

  1. AI writes SDTM mappings

    Saama reports that its SDTM Navigator, in pilot use, wrote 75% of the mappings from raw trial data to SDTM automatically, about 70% of them accurately, and cut the time to write mapping specifications and code by 50 to 70% for studies with 20 or more data domains.

    Achieved whenaccuracy is measured on public data with a known answer, or reported by a drug company that is not the vendor.

    saama.com

    Claimed by

    Saama, July 2025

  2. Hospital records fill trial forms

    Mount Sinai's Tisch Cancer Center says IgniteData's Archer system, linked to its Epic hospital records, now fills about 70% of trial case report forms automatically in all 16 of its cancer disease groups, cutting data entry from 5.5 to 2.5 minutes per visit, with "100 percent data accuracy" in the data checked so far.

    Achieved whena peer-reviewed or independently audited report gives the counts behind the percentages and checks accuracy against source records.

    newswise.com

    Claimed by

    Mount Sinai Tisch Cancer Center, October 2026

  3. No copying errors, one study

    In six trials at three sites outside the US (66 participants, 15,679 form fields), staff copying data by hand made errors in 8.23% of fields, and data sent directly from hospital records to the trial database had none. Two authors work for the vendor, Yonalink, which also mapped the fields.

    Achieved whena group with no ties to the vendor repeats it at a larger scale.

    pmc.ncbi.nlm.nih.gov

    Claimed by

    UT Health San Antonio and Yonalink, in the AMIA Joint Summits proceedings, June 2026

Open

  1. 01

    Score AI on SDTM conversion

    A public test that gives an AI system the raw example datasets in pharmaverseraw (adverse events, demographics, disposition, drug exposure and vital signs) and the SDTM rules, compares its output cell by cell with the matching pharmaversesdtm data, and runs CDISC's open rules engine (CORE) on the result.

    The brief
    Credit · unclaimed
  2. 02

    Open hospital-to-SDTM converter

    Open-source code that turns hospital-record data in the FHIR R4 format, the common standard for exchanging health records, into five SDTM domains (demographics, vital signs, lab results, medications and adverse events), following the HL7 and CDISC mapping tables, tested on synthetic patients from Synthea and checked with CORE.

    The brief
    Credit · unclaimed
  3. 03

    Which form fields records hold

    Label every field on public case report forms, starting with the annotated blank form of the CDISC pilot study, as held in hospital records as structured data (with its FHIR path), held only in free-text notes, or not held at all. Then test the structured ones by filling them from Synthea patients.

    The brief
    Credit · unclaimed

Open 01 · the brief

Score AI on SDTM conversion

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test that gives an AI system the raw example datasets in pharmaverseraw (adverse events, demographics, disposition, drug exposure and vital signs) and the SDTM rules, compares its output cell by cell with the matching pharmaversesdtm data, and runs CDISC's open rules engine (CORE) on the result.

Why it is still open

The only accuracy figures come from a vendor (Saama, about 70%) or from one study on data that are not public (ClinAgent, 72.1%). A PharmaSUG 2026 paper built an AI loop checked by CORE but reported no accuracy against public data. I found no public leaderboard.

europepmc.orgpharmasug.org

Built with what exists

pharmaverseraw and pharmaversesdtm; sdtm.oak, whose hand-written programs give a baseline; CDISC CORE; Claude or Gemini run as a coding agent that reruns its code and fixes it from the validator's errors.

raw.githubusercontent.com

  1. First step

    Score one model on demographics and vital signs.

  2. Done looks like

    A scoring script that reports, for each data domain, the share of variables exactly right, the share of records matched and the CORE error count. Baselines for at least two AI models and for sdtm.oak.

Open 02 · the brief

Open hospital-to-SDTM converter

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Open-source code that turns hospital-record data in the FHIR R4 format, the common standard for exchanging health records, into five SDTM domains (demographics, vital signs, lab results, medications and adverse events), following the HL7 and CDISC mapping tables, tested on synthetic patients from Synthea and checked with CORE.

Why it is still open

The joint HL7 and CDISC guide calls itself "purely a 'descriptive' guide", with mapping tables only. Commercial systems are closed, and in the Yonalink study the vendor mapped the fields for each trial at its start. I found no open-source converter.

hl7.orgpmc.ncbi.nlm.nih.gov

Built with what exists

Synthea, an open-source generator of synthetic patient records; the HL7 mapping tables; Python or R; CDISC CORE; Claude.

  1. First step

    Convert Synthea's patient records to demographics and its vital-sign observations to vital signs.

  2. Done looks like

    At least 1,000 synthetic patients converted with no CORE errors in the five domains, and 100 random values traced back to their FHIR source with a 100% match.

Open 03 · the brief

Which form fields records hold

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Label every field on public case report forms, starting with the annotated blank form of the CDISC pilot study, as held in hospital records as structured data (with its FHIR path), held only in free-text notes, or not held at all. Then test the structured ones by filling them from Synthea patients.

Why it is still open

Coverage figures exist only for private forms. Sanofi and Flatiron found 86% of collected data points eligible for transfer across five cancer trials whose forms are not public, and the TransFAIR study at three European hospitals transferred 39.6% of the data in its scope. I found no public field-by-field table.

resources.flatiron.comeuropepmc.org

Built with what exists

the CDISC pilot project on GitHub; the HL7 mapping tables; Synthea; Claude for first-pass labels.

  1. First step

    Label the CDISC pilot form.

  2. Done looks like

    At least five forms fully labelled. 200 fields coded separately by two people, agreeing at least 90% of the time. The share of fields that can be filled, per form, confirmed by a Synthea fill test.