01 / THE ARGUMENT
A stronger evidence base still leaves the enrollment question open
A 2026 review by Yin and colleagues included 121 studies and separated AI applications by recruitment task. Pooled sensitivity for patient screening was 0.91, with a 95% confidence interval of 0.84–0.95. Those figures describe a particular task across heterogeneous studies; they are not a promised enrollment rate for a new sponsor or site. The authors' literature search ended in June 2024, a material limitation when assessing today's language models.[1]
TrialGPT offers a useful concrete example. Its published evaluation used 183 synthetic patient cases and more than 75,000 trial annotations; the accompanying user study reported a 42.6% screening-time reduction.[2] That is a reason to test assisted review. It does not establish that real patients enroll sooner, that a trial closes earlier, or that underserved populations benefit. An investment committee should keep eligibility accuracy, reviewer time, enrollment, and participant retention in separate columns.
02 / THE ARGUMENT
Representativeness belongs inside the recruitment architecture
FDA's December 2025 final guidance recommends approaches to recruit a population that reflects the patients likely to use a product. Its scope includes demographic characteristics and factors such as organ dysfunction, comorbidity, disability, and residence.[3] For an operator, this means a recruitment strategy cannot be evaluated exclusively against the easiest-to-query patients in an academic electronic record.
The proposed design begins with an ascertainment map: which clinically relevant patients enter the searchable data, which are invisible, and which arrive too late to participate? Missing molecular tests, fragmented community records, inaccessible digital forms, language needs, and travel constraints should each produce an explicit work queue. Compare the disease population with the prescreened, approached, consented, enrolled, and retained cohorts. A balanced final cohort can conceal unequal attrition upstream; stratified conversion reveals where the program actually needs support.
03 / THE ARGUMENT
Treat site capacity as a live constraint on recommendations
The matching service should maintain a time-stamped operational registry alongside trial criteria: active protocol version, site and cohort status, screening slots, coordinator availability, investigator coverage, required procedures, and referral contacts. Registry recruitment status alone should not be treated as confirmation of a locally available slot. Before outreach, the receiving research team verifies that the opportunity remains open and clinically appropriate.
A hypothetical site that receives twenty additional plausible matches but can review only five creates a longer queue. To prevent that failure, route cases through documented states: candidate, missing evidence, clinician reviewed, site confirmed, patient invited, consented, screened, enrolled, and follow-up complete. Every state needs an owner and aging limit. Prioritization must follow approved clinical and research rules, while recording why a case was deferred. Marketing attribution should stop at the research boundary; a click is not consent or eligibility.
04 / THE ARGUMENT
Preserve uncertainty instead of manufacturing eligibility
A clinically useful interface distinguishes satisfied, contradicted, and unknown criteria. Each judgment should cite the source document, relevant text, observation date, and protocol version. Unknown washout timing or organ function should generate a request for clarification, not an inferred positive match. Screening decisions remain with authorized clinicians and research personnel. The audit trail should include human overrides and the reason for disagreement with the model.
The business case must include review and correction costs. False positives consume coordinator time; false negatives may deny a person consideration. Those errors have different consequences and should be evaluated separately by cohort. Before live routing, use a clinician-adjudicated sample to test extraction, temporal logic, contradictory records, and criterion interpretation. After deployment, review missed opportunities as well as rejected matches. This is a proposed governance design, not an effect size established by the cited studies.
05 / THE ARGUMENT
Underwrite participation rather than speculative launch acceleration
Start the economic ledger with verified paid research activity, avoided manual review, and the cost of additional patient support. Count an enrolled participant once and separate incremental enrollment from cases that would have entered through the existing pathway. Release of coordinator hours has value only when those hours are reassigned, replace purchased work, or increase appropriate throughput within staffed capacity. Travel assistance, interpreters, accessible scheduling, and follow-up are participation infrastructure, with their own budget and owner.
Do not capitalize an earlier product launch from screening-time improvement alone. The schedule may instead be constrained by site activation, competing trials, endpoint accrual, or regulatory requirements. A prospective comparison across sites or periods should estimate incremental enrollment, time to enrollment, retention, subgroup conversion, and cost per retained participant. If the model generates more recommendations while review backlogs grow or representation deteriorates, the correct response is to change the pathway before purchasing additional algorithmic reach.
FROM EVIDENCE TO ALLOCATION
The operating and investment case
Proposed design by Azis R. Dabas. These decisions and evaluation criteria are not outcomes established by the cited studies.
- Decision
- Fund one protocol-specific assisted-recruitment pathway with a defined eligible catchment and prospective comparator before expanding across indications.
- Accountable owner
- Sponsor clinical operations and the site's principal investigator jointly sponsor the program; a research operations lead owns routing and capacity, with privacy and patient-access review.
The delivery sequence
- Define the disease catchment and missing-data pathways
- Extract criteria with dated source evidence
- Verify clinician appropriateness and local site capacity
- Offer accessible, consent-based participation
- Track screening, enrollment, retention, and reasons for attrition
The economics
Incremental completed and paid study activity plus demonstrably redeployed review capacity, less software, integration, adjudication, coordinator expansion, and participation support. Keep enrollment economics separate from speculative sponsor launch value.
The measures that govern expansion
- Additional enrolled and retained participants versus comparator
- Time from clinical identification to site-confirmed invitation
- Review minutes per adjudicated case and unresolved queue age
- Approach-to-enrollment conversion by relevant subgroup
- Cost per incremental retained participant and protocol deviations
Stop or redesign when
Pause automated routing when unsupported eligibility judgments, protocol-version errors, unmanageable coordinator backlog, or worsening subgroup attrition exceed prospectively agreed limits. Continue clinician-led recruitment while correcting the cause.
THE EVIDENCE LEDGER
What supports the argument
Study findings, policy requirements and market signals support different claims. Their boundaries remain visible.
[1] peer-reviewed · April 23, 2026
Artificial intelligence in clinical trial participant recruitment and retention: A scoping review and meta-analysis
- Design or status
- Scoping review of 121 studies with task-specific random-effects meta-analysis; literature January 2018–June 2024.
- Verified finding
- Patient-screening tools had pooled sensitivity 0.91 (95% CI 0.84–0.95), while performance varied across recruitment tasks.
- Boundary
- Heterogeneous task definitions, datasets, study designs, and implementation settings; model metrics do not establish incremental representative enrollment, retention, or sponsor ROI.
[2] peer-reviewed · November 18, 2024
Matching patients to clinical trials with large language models
- Design or status
- Patient-to-trial matching benchmark using 183 synthetic patients across three cohorts, manual criterion assessment, and a screening-time user study.
- Verified finding
- The authors' NIH project record reports more than 75,000 annotations, 87.3% criterion-level matching accuracy, and a 42.6% screening-time reduction in the user study.
- Boundary
- Synthetic patient evaluation and a bounded user study do not demonstrate real-world incremental enrollment, trial representativeness, or long-term retention.
[3] policy · December 2025
Enhancing Participation in Clinical Trials — Eligibility Criteria, Enrollment Practices, and Trial Designs: Guidance for Industry
- Design or status
- Final regulatory guidance; recommendations rather than a clinical intervention study.
- Verified finding
- Recommends recruitment approaches to increase representation across demographic and non-demographic baseline characteristics relevant to expected users of a drug or biologic.
- Boundary
- Guidance is not empirical proof that any specific AI tool improves participation; individual protocol and regulatory requirements remain applicable.
FOLLOW THE SOURCE
Sources and editorial method
Selected evidence was reviewed through October 7, 2026. Numbered references connect claims to their underlying records. Economic mechanisms and business cases are the author’s analysis unless a source is cited. The review is selective; publication dates retain the precision available in the source.
Artificial intelligence in clinical trial participant recruitment and retention: A scoping review and meta-analysis
Ziran Yin, Yun-Chung Liu, Jonathan Chong Kai Liew, Rui Yang, Stephanie Hendren, Elisa Ma, Zhaomei Geng, Jiahan Wang, Henry Foote, Christopher Lindsell, Chuan Hong. Journal of Clinical and Translational Science. . peer-reviewed.
DOI: 10.1017/cts.2026.10743 · Primary verification record · Consensus paper record
Matching patients to clinical trials with large language models
Qiao Jin, Zifeng Wang, Charalampos S. Floudas, Fangyuan Chen, Changlin Gong, Dara Bracken-Clarke, Elisabetta Xue, Yifan Yang, Jimeng Sun, Zhiyong Lu. Nature Communications. . peer-reviewed.
DOI: 10.1038/s41467-024-53081-z · Primary verification record
Enhancing Participation in Clinical Trials — Eligibility Criteria, Enrollment Practices, and Trial Designs: Guidance for Industry
U.S. Food and Drug Administration. FDA final guidance. . policy.
Study authors retain credit for their work. Researcher affiliations and publisher names do not imply affiliation with or endorsement of this analysis.