What Slows Trial Enrolment, and What AI Recruitment Results Actually Measured

In February 2025 JAMA published a randomised comparison of AI-assisted and manual prescreening for a heart failure trial. At Mass General Brigham, 4,476 patients were randomised to record screening by study staff or by a large language model tool called RECTIFIER. Staff then checked the charts it flagged. The AI arm found 458 eligible patients against 284. It enrolled 35 against 19.

That is close to double the enrolment rate, and the strongest evidence I could find that AI patient recruitment can move the number that matters. It also means 2,207 of the 2,242 patients in the AI arm were never enrolled. That gap is the useful starting point for anyone trying to accelerate clinical trial enrolment. A screening tool works on one step of a long funnel. Most published evidence on enrolment delays points at two earlier steps: eligibility criteria that exclude patients before anyone looks, and sites that do not have the patients.

Where is enrolment actually lost?

The most-cited site benchmark has an awkward history. Kenneth Getz of Tufts CSDD wrote in Applied Clinical Trials in May 2012 that an older, widely repeated figure of 20% of sites enrolling nobody came from interviews. Tufts replaced it with sponsor data on nearly 16,000 sites in 151 phase 2 and 3 global trials run between 2008 and 2010.

The corrected picture is better but not good. 11% of sites ready to recruit did not "enroll a single patient". Counting those, 48% of all sites under-enrolled. Only 47% of studies finished enrolment on time, and one in six took more than twice as long as planned. Getz thought even that understated the problem, because timelines reset after amendments were hard to see. Registry studies count the same failure from the trial end.

Carlisle's team also looked for what failed trials had in common. Unsuccessful accrual was associated with more eligibility criteria, fewer sites, an active rather than placebo comparator, earlier phase, and public rather than industry funding. Speich's review found the same funding effect: industry trials were less often stopped for poor recruitment (adjusted odds ratio 0.32).

Read that list as an operations manager would. Criteria count and comparator are protocol decisions. Site count is a feasibility decision. Chart-review speed does not appear on it. Screening still matters, but it is the third place to look, not the first.

Which levers have a measured effect?

Little recruitment advice has been tested. A 2018 Cochrane review of 68 trials of recruitment strategies, with more than 74,000 participants, judged only three of 72 comparisons to be high-certainty evidence. Open trials recruited 10 percentage points more than blinded placebo trials. Telephone reminders to people who ignored a postal invitation added 6 points. A user-tested information leaflet made little or no difference. The authors summed up the field as "plenty of variety but little depth". The levers below each have a measured effect, with the thing measured named.

Eligibility criteria come first

Trial Pathfinder, built by Stanford and Genentech researchers and published in Nature on 7 April 2021, emulated completed advanced lung cancer trials on 61,094 patients from a nationwide EHR-derived database. Many common exclusions, including several laboratory-value cut-offs, had minimal effect on the trial hazard ratios. Broadening them with a data-driven approach more than doubled the eligible pool on average.

This is emulation on routine-care records, not a prospective trial. It shows which criteria cost patients without changing the answer, not how fast a broadened trial would recruit.

Timing matters as much as content. Tufts CSDD's 2022 data on 950 protocols, published in 2024, found 76% had at least one amendment. Getting from the need to amend to the last oversight approval took 260 days on average. A criterion removed before the protocol is final costs a meeting. Removed after the first site opens, it costs most of a year.

The other design lever cuts how many patients you need. Prognostic covariate adjustment shrank the total by 15% in CHMP's own worked example, and what CHMP qualified and refused in that opinion matters before anyone writes it into a protocol.

Site selection and feasibility

ICH E6(R3) section 2.2.1 already asks investigators to show recruitment potential "based on retrospective or currently available data". Many feasibility questionnaires still collect estimates. The gap between sites that never enrol and sites that beat target is the size of the prize. What is missing is a prospective comparison of data-driven against conventional site selection that counts randomised patients. The evidence is vendor case studies, weighed below.

EHR-based prescreening, with and without AI

The oldest controlled evidence is not AI at all. In a study published in 2005, Embi and colleagues switched on an EHR alert in a US academic health system. It prompted clinicians when a patient met selected trial criteria. Referrals per month rose more than tenfold. The enrolment rate doubled, from 2.9 to 6.0 a month. Tenfold more referrals produced twice the enrolments. Expect that pattern from any matching tool: finding candidates is not enrolling them.

What did the published AI recruitment results actually measure?

Most AI patient recruitment claims rest on four kinds of study. Each measured something real. Only one measured enrolment.

The most useful result in the Mendel study is a negative one. One of the three trials had failed to enrol. Applied retrospectively, the AI found no suitable patients for it either. A matching tool cannot find patients the criteria or the catchment have excluded.

The TrialGPT authors concede that "our pilot user study is of limited scope in sample size". Its 42.6% is time saved by two expert annotators on six semi-synthetic cases.

Read the JAMA trial precisely. Its primary outcome was time to an eligibility decision. The enrolment difference, from 54 enrolments in total, had a hazard ratio of 1.79 (95% CI 1.02 to 3.15). The authors name the limits: one centre, one trial, one disease. Even so, it counted enrolled patients, which is the right unit.

A July 2026 perspective in Patterns argues that matching has reached near-expert accuracy without raising enrolment, because the bottleneck is logistics, workflow and consent. That is an argument, not a measurement, but it fits the Embi and JAMA funnels. Read these figures like the time-to-first-draft numbers behind clinical AI cycle-time claims: real, and a step upstream of the outcome you pay for.

Four enrolment case studies, and what each one shows

These four, from pharma sponsors and a CRO, cover site selection, trial rescue and generative AI. Every number is a vendor claim.

A phase 3 sponsor using claims data to re-pick sites. In Definitive Healthcare's case study, an unnamed company had enrolled only 20% of its patients against a 24-month goal. Claims analysis found that 18 of the top 20 sites with eligible patients were not in the trial. The remaining 80% then enrolled within 12 months, with an "estimated" $10 million saved and approval 12 months earlier than anticipated. It shows a site problem, diagnosed with the kind of data available at feasibility. It gives no baseline, no method for the saving, and no reason to credit approval timing to enrolment alone.

A CRO testing two explanations before rescuing a trial. Citeline ran the analysis for Biorasi, a CRO, on a phase 3 oncology trial with high site decline rates and too few patients passing screening. It tested whether the sites were wrong or the protocol was the barrier. The sites were mostly sound. The analysis found 199 more US sites and eight countries for expansion, and supported a protocol amendment "projected" to lift enrolment by 15–20%. The order of diagnosis is the lesson. The uplift is a projection, and on Tufts' numbers the amendment would take months to approve.

A mid-sized sponsor choosing rare-disease sites with AI. Medidata reports a 50% increase in site count for a rare-disease trial using its Intelligent Trials product. Site count is an input, not an outcome: more sites can mean faster enrolment, or more sites each enrolling one patient at full start-up cost. The public page gives no enrolment figure.

Sanofi and generative AI for recruitment content. Muse, launched by Formation Bio, Sanofi and OpenAI on 12 November 2024, aims to cut recruitment strategy and content creation from months to minutes. Sanofi planned the first deployment in phase 3 multiple sclerosis studies. The launch page gives no measured outcome. Its agent applies IRB and regulatory guidelines with humans in the loop, but ethics approval of every piece of material still sets the pace.

Which rules bind prescreening? Position current at 12 October 2026

Prescreening means reading identifiable health records before the patient has agreed to anything. The rules decide who may look, where data may go, and what ethics committees approve first.

United States. Under 45 CFR 164.512(i)(1)(ii), a covered entity may let a researcher review records to prepare research if no protected health information is removed. The University of Wisconsin–Madison IRB reads this as letting the institution's own workforce contact potential participants, while outside researchers need a partial waiver of authorisation. For a sponsor the consequence is architectural: the model runs inside the hospital, and the sponsor sees counts, not charts. For federally funded research, Common Rule 46.116(g) lets an IRB approve screening by "accessing records" without consent. FDA's Recruiting Study Subjects information sheet, guidance rather than law, treats direct advertising as "the start of the informed consent and subject selection process".

ICH E6(R3). The Principles and Annex 1 were adopted on 6 January 2025 and, per the EMA's E6 page, took effect in the EU on 23 July 2025. Article 47 of the EU Clinical Trials Regulation obliges sponsors and investigators to take appropriate account of ICH GCP. Section 1.2.2(e) puts recruitment advertisements before the IRB or IEC, and 2.4.2 requires their approval before the trial starts. Appendix B.5.3 asks the protocol to describe pre-screening. Appendix C.3.1 counts records of recruitment and pre-trial screening as essential. It also expects trial-specific systems to be validated, and clinical-practice systems to be assessed as fit for their use in the trial. A matching tool that feeds your screening log falls under one of those, and which E6(R3) clauses govern a clinical AI model covers the validation detail.

European Union. Article 7(1)(c) of Regulation 536/2014 makes recruitment arrangements a Part II matter, assessed by each Member State for its own territory. Annex I section K requires the application to give "a clear indication of what the first act of recruitment is". Advertising material must be submitted, including the communications used to invite participants. A generative tool that drafts fifty ad variants in minutes has drafted fifty items for ethics review. Submit a small set.

The EDPB's Opinion 3/2019 of 23 January 2019 says trial informed consent must not be confused with consent as a GDPR legal basis. For research processing it points to public interest or legitimate interests, with Article 9(2)(i) or (j). Prescreening happens before any trial consent, on data collected for care. Article 9(2)(j) needs a basis in Union or Member State law, and Article 9(4) lets Member States add conditions for health data. So the answer differs by country, and the hospital's data protection officer gives it, not the vendor. A coordinator on every flag also keeps you clear of Article 22's limits on decisions based solely on automated processing.

What does not bind: the FDA information sheet, the EDPB opinion and any vendor's description of its own compliance. None replaces the hospital's legal basis or the ethics committee's approval.

What this means in practice

The steps run in lead-time order, starting with the one that informs the rest.

1. Measure the funnel on your last studies. From screening logs, count candidates identified, approached, consented, screened and randomised, with screen failures by criterion and site. Without this baseline, no later result can be compared with anything.
2. Test eligibility criteria before the protocol is final. Count how many real patients each criterion removes. Loosen the ones that cost patients without protecting safety or the endpoint, and ask whether a prognostic covariate design could cut the number needed. After launch, each change is an amendment averaging 260 days.
3. Choose sites on data, not questionnaires. Ask investigators for the retrospective counts E6(R3) section 2.2.1 expects. Compare where eligible patients are, from claims or EHR counts, with where your shortlisted sites are.
4. Settle the prescreening route site by site. In the US, choose workforce prescreening or a partial waiver at each hospital. In the EU, get the legal basis from each hospital's data protection officer. Run the model inside the hospital.
5. Put recruitment into the ethics submission. Describe pre-screening in the protocol and submit a fixed set of recruitment materials, including any drafted with generative AI.
6. Pilot matching on randomised patients per site-month. Keep a coordinator confirming every flag and compare against the baseline. Screening hours saved are a cost figure, not an enrolment figure.
7. Watch recruitment as a risk indicator during conduct. E6(R3) section 3.11.4.5.2(h) puts recruitment rates in the monitor's remit. Act on silent sites early, or a recruitment dashboard repeats the gap between adopting risk-based monitoring and changing what monitors do.