Industries / Health / Pharmaceuticals / Clinical research
Finding and enrolling participants
Finding people who qualify for a trial, telling them about it, getting their freely given consent, and making sure the people enrolled resemble the people who will use the treatment.
7
Milestones achieved, from 1947 to 2025.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1947
Consent becomes rule one
At the end of the Nuremberg Doctors' Trial of Nazi physicians, the American judges' verdict of August 1947 set out ten rules for medical experiments on people, later known as the Nuremberg Code. The first reads: "The voluntary consent of the human subject is absolutely essential."
Credit
the American judges of the tribunal, working from a memorandum by the physician Leo Alexander
1974
Ethics boards required by law
In 1972 an Associated Press story exposed the US Public Health Service's Tuskegee study, which since 1932 had followed 399 Black men with syphilis and 201 without, told them they were being treated for "bad blood", collected no informed consent and did not offer them penicillin once it became the standard treatment. The National Research Act of 12 July 1974 then required institutions with federal health grants to have an Institutional Review Board, an ethics committee that reviews research on people before it starts.
LessonThe study ran for 40 years before people outside it stopped it.
Credit
US Congress
2014
A trial run wholly online
Pfizer's REMOTE trial of a drug for overactive bladder ran over the web from a single site; its authors call it the first entirely web-based trial run under an FDA investigational drug application. Of 5,157 people who registered on the trial website, 18 were randomised, against a goal of 283.
LessonPutting every step online did not by itself fill a trial; almost everyone dropped out during screening.
Credit
Miguel Orri and Craig H. Lipset of Pfizer, with Steven R. Cummings and colleagues at the University of California, San Francisco
2015
Trials that cannot fill
Of 2,579 phase 2 and phase 3 trials that closed in 2011, 481 (19%) either stopped for poor enrolment or finished with less than 85% of the people they planned, and 48,027 patients had joined trials that could not answer their main question. Trials with more eligibility criteria, fewer sites or no industry funding fell short more often.
LessonA trial that cannot recruit enough people asks those who did join to bear its risks for no answer.
Credit
Benjamin Carlisle, Jonathan Kimmelman and colleagues (McGill University)
2019
Key information goes first
The revised US Common Rule, the federal rule for research on people first issued in 1991, requires every consent form to open with a short summary of the main reasons someone might or might not want to join, and requires one approved consent form from each federally funded trial to be posted on a public federal website. Most institutions had to comply from 21 January 2019.
LessonA 2023 review of 102 posted forms found a key-information section in 76, but those sections were no easier to read than the rest of the form.
Credit
the US Department of Health and Human Services and the other federal agencies that follow the Common Rule
2022
Trials should match patients
The EU Clinical Trials Regulation, applied from 31 January 2022, says the people in a trial should represent the groups, for example by sex and age, likely to use the medicine, and a protocol must explain any group it leaves out. Its public portal, the Clinical Trials Information System (CTIS), added a map of trials in March 2025 that can be searched in plain words in the official languages of the EU and EEA.
Credit
the European Parliament and Council; the European Medicines Agency and EU member states, which run CTIS
2025
Diversity plans stall
A US law of December 2022 told the FDA to require "Diversity Action Plans" for pivotal trials, setting out how sponsors will enrol people of different ages, sexes, races and ethnic groups, and the FDA issued draft guidance in June 2024. In January 2025, after an executive order against diversity programmes, the FDA took the guidance off its website; a court order made the agency restore the page, which on 18 November 2026 still showed the June 2024 document as a draft.
LessonThe plans become binding only once the guidance is final, so a requirement written into law stalled when policy changed.
Credit
US Congress, in the Food and Drug Omnibus Reform Act (FDORA); FDA
Watch for
AI chart review screens patients
Cleveland Clinic reports that Dyania Health's Synapsis AI, a system that reads medical records, prescreened patients for a phase 3 heart trial: seven enrolled in six days through AI screening, against 10 in 90 days through usual screening, and 36.6% of the 30 patients it found were Black, against 7.1% of those found by routine screening. Cleveland Clinic has invested in Dyania and may benefit if the system sells.
Achieved whena group with no financial tie to Dyania runs a randomised or multi-site test and reports how many people enrolled, not only how many were flagged.
Claimed by
Cleveland Clinic and Dyania Health, March 2026
Trial matching by NHS App
The UK government said millions of people can now search for and sign up to trials through the NIHR's Be Part of Research service in the NHS App, and that the service will later match patients to studies from their own health data and send them alerts.
Achieved whenautomatic matching is live and NIHR or an independent body publishes how many people were matched and then enrolled, by ethnic group.
Claimed by
the UK Department of Health and Social Care and the National Institute for Health and Care Research (NIHR), June 2025
AI answers consent questions
A language model limited to one trial's own documents answered questions about joining the trial with a mean accuracy rating of 4.8 from human raters and at about a grade 7.5 reading level. The human raters disagreed widely with one another (kappa 0.1, where 1 is full agreement), some finding answers accurate but lacking warmth, and the model has not yet been tested with real prospective participants.
Achieved whena group other than the developers tests it with prospective participants and measures whether they understood the trial better than with standard consent.
Claimed by
Moscatel and colleagues (Sinai Health and University of Toronto), in npj Digital Medicine, June 2026
Open
01
Eligibility rules as data
For every recruiting trial for one common condition, such as type 2 diabetes, turn the free-text eligibility criteria into structured rules (age, laboratory cut-offs, required and excluded diagnoses and drugs) with standard medical codes. Publish the rules, the code and a hand-made answer key for a random sample, and rerun monthly.
The briefCredit · unclaimed02
Enrolled versus who is ill
For the 20 conditions with the most US trial results, a public table that compares the share of Black, Asian, Hispanic and over-65 participants in trials, taken from ClinicalTrials.gov results, with their share among people who have the condition. Refresh it monthly.
The briefCredit · unclaimed03
Readable listings in Spanish
For every recruiting trial for one condition in one region, a short listing in English and in Spanish at about a grade 6 to 8 reading level, saying who can join, what happens, how many visits there are and where. Score each listing, fact by fact, against an answer key drawn from the registry record, and link each listing to that record.
The briefCredit · unclaimed
Open 01 · the brief
Eligibility rules as data
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For every recruiting trial for one common condition, such as type 2 diabetes, turn the free-text eligibility criteria into structured rules (age, laboratory cut-offs, required and excluded diagnoses and drugs) with standard medical codes. Publish the rules, the code and a hand-made answer key for a random sample, and rerun monthly.
Why it is still open
ClinicalTrials.gov stores eligibility as one block of free text, with only age, sex and healthy-volunteer status as separate fields. Research sets are one-off snapshots: Chia annotated the criteria of 1,000 trials once, in 2020. A 2026 test on 800 criterion-patient pairs from real hospital records found that simple AI prompting, which scored well on older test sets of short made-up patient summaries, nearly failed.
Built with what exists
the ClinicalTrials.gov API; Chia's annotation scheme as a template; Claude or Gemini for a first pass; two people to build the answer key; Python.
First step
Pull the recruiting trials for the condition through the API and annotate 10 by hand with Chia's scheme.
Done looks like
Every recruiting trial for the condition converted. On a 50-trial answer key, at least 90% of individual criteria captured correctly, field by field, with the error log published.
Open 02 · the brief
Enrolled versus who is ill
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For the 20 conditions with the most US trial results, a public table that compares the share of Black, Asian, Hispanic and over-65 participants in trials, taken from ClinicalTrials.gov results, with their share among people who have the condition. Refresh it monthly.
Why it is still open
The FDA's Drug Trials Snapshots report who enrolled in the trials behind each new drug, but not who has the disease. The largest recent study across conditions, a preprint of 17 November 2026, covers sex only, for 10 conditions; others take one condition at a time.
Built with what exists
the ClinicalTrials.gov API (the baseline characteristics in each results section); published US figures on who has each condition by race and age, such as CDC's NHANES survey data; Python; Claude or Gemini to map sponsors' varied race labels to standard categories, checked by hand.
First step
Pull every heart failure trial with posted results and list every distinct race label the sponsors used.
Done looks like
Ratios for four groups across 20 conditions. On a 30-trial hand check, every extracted count matches the registry. A methods page that names each prevalence source.
Open 03 · the brief
Readable listings in Spanish
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For every recruiting trial for one condition in one region, a short listing in English and in Spanish at about a grade 6 to 8 reading level, saying who can join, what happens, how many visits there are and where. Score each listing, fact by fact, against an answer key drawn from the registry record, and link each listing to that record.
Why it is still open
A Vanderbilt team measured 17,175 ClinicalTrials.gov summaries at an average reading grade of 18.7, graduate level. A 2025 study rewrote 20 recruiting trials for urologic cancer with AI and measured how readable they were, but did not check facts against a key or translate them.
Built with what exists
the ClinicalTrials.gov API; Claude or Gemini; the Python package textstat for reading level; two bilingual checkers.
First step
Choose the condition and region, and build the answer key for 10 trials.
Done looks like
Listings for every recruiting trial for the condition in the region. On a 40-trial answer key, no errors in age, sex, condition or location, at least 95% of key facts present, and a median reading grade of 8 or below in both languages.