AI in Patient Recruitment & Site Selection
How LLMs are reading EMRs to find eligible patients, how predictive models are ranking sites before startup, and how to deploy both without breaking HIPAA, GDPR, or ICH E6 (R3).
BY CAROLINA BOSCH · DIRECTOR, R&D CLINICAL DATA
What AI-Driven Recruitment Is
Patient recruitment is still the single biggest driver of trial delay. Roughly 80% of trials miss enrollment timelines, and about a third of Phase III sites never enroll a single patient. The bottleneck is not intent, it is matching complex eligibility criteria to messy, unstructured clinical records fast enough to compete with standard of care.
AI-driven recruitment uses NLP and LLMs to read EMR/EHR notes, pathology, and imaging summaries at population scale, surface candidates against protocol eligibility, and hand a ranked, pre-screened list to the study coordinator. Site selection uses predictive models over historical enrollment, therapeutic-area experience, and competing-trial density to rank sites before contracts are signed, not after enrollment stalls.
LLMs for EMR/EHR Screening
Criteria parsing
Protocol eligibility, prose written for humans, parsed into a structured, testable set of inclusion/exclusion predicates with source citations back to the protocol section.
Unstructured notes
LLMs read progress notes, pathology, radiology, and discharge summaries, the fields ICD/CPT codes miss, to surface patients a coded query would silently drop.
Longitudinal context
Prior lines of therapy, response, and washout windows reconstructed from years of encounters, not a single snapshot.
Ranked candidate list
Patients ranked by eligibility confidence, distance to site, and predicted retention, not a 40,000-row query dump.
Pre-screen summary
One-page structured summary per candidate with cited excerpts from the record, so the coordinator's review takes minutes, not an hour.
Human-in-the-loop
A qualified clinician reviews every AI-surfaced patient before outreach. The model shortens the funnel; it does not replace clinical judgment or the informed consent process.
Predictive Site Selection
Site selection used to be a spreadsheet of last-trial performance and a CRA's gut feel. Predictive models now combine sponsor-side history, public registries, claims-derived patient population density, competing trials, investigator publication activity, and startup timelines to rank prospective sites before the FPI clock starts.
The useful output is not a single score, it is a ranked list with the reasons each site is likely to enroll on time, plus a realistic enrollment curve. That is what a clinical operations lead uses to negotiate commitments that match reality, and to catch the low-enrolling site in month two instead of month eight.
"The wrong question is 'which sites should we pick?' The right one is 'given this protocol, this competitive landscape, and this timeline, which sites will actually deliver, and which are we adding for coverage we already have?'"
Privacy, HIPAA & GCP
PHI does not leave the covered entity's environment. LLM calls run under a BAA or on infrastructure the health system controls; de-identification, minimum-necessary access, and IRB oversight of pre-screen workflows are not optional. Sponsors receive counts and cohort characteristics, not identified patient rows.
Under ICH E6 (R3), the recruitment model is a computerized system: intended use documented, risk-based validation against a labeled cohort, versioned model and prompt, change control, and an audit trail that links every recommendation to source records and the human reviewer. If the sponsor cannot show which patients the model flagged and why, the workflow is not inspection-ready.
Metrics That Matter
The wrong metric is "patients screened by AI." The right ones are pre-screen precision and recall against a chart-reviewed gold standard, coordinator minutes per enrolled patient, screen-fail rate, time from referral to consent, and enrollment vs the model's own predicted curve per site.
Track them by therapeutic area, site, and protocol. If AI-assisted cohorts drift on precision or retention, you catch it before the DMC does.
Getting Started
- 1
Pick one indication
One therapeutic area, one protocol shape, at two or three sites you already work with. Bounded scope is the difference between a pilot that ships and a pilot that dies in IRB.
- 2
Build a labeled cohort
Chart review a stratified sample of eligible and ineligible patients end to end. That is your ground truth for precision and recall, not the vendor's benchmark.
- 3
Deploy assistive first
Every AI-flagged patient goes to a clinician before outreach. Measure precision, coordinator minutes saved, and retention at 90 days.
- 4
Validate as a computerized system
Intended use, risk assessment, validation against the labeled cohort, change control for model and prompt versions, audit trail per recommendation.
- 5
Layer on site prediction
Only once EMR screening is working, add predictive site selection on the next protocol. Rank sites, share the reasons, and revisit at the first interim enrollment review.
Rolling out AI-driven recruitment?
I am open to senior leadership roles in AI, Healthcare & Life Sciences, and Developer Ecosystems, plus select consulting engagements on AI-assisted recruitment, predictive site selection, and validated deployment inside health systems and sponsors.