AI can help sites find potential trial participants by connecting details scattered across clinical notes, pathology reports, and treatment histories. To put that capability to work, start with the questions a coordinator needs answered: Which criteria does this patient appear to meet, what evidence supports that assessment, and what still needs confirming?
Those questions give teams a practical foundation for implementation. They guide how to translate the protocol, present potential matches, and test whether the system is ready for everyday screening.
Translate eligibility criteria with the people who use them
Choose an initial protocol with an established screening process and staff who know where interpretation gets tricky. Their experience will help distinguish requirements that can be checked directly from those that need more context.
An age range is relatively straightforward. Prior treatment requirements can involve several connected questions: what the patient received, when treatment ended, and whether it was stopped because of progression or intolerance. Finding a drug name in a note answers only part of that assessment.
This is where clinical and technical teams need to work together. Clear requirements can use explicit rules, while AI helps locate and interpret supporting information in narrative records. Coordinators and investigators should review how the system represents each criterion, especially where timing, exceptions, or combinations of conditions affect eligibility.
Keep the original protocol language accessible alongside that interpretation. When a criterion changes, teams should be able to see which matching logic needs updating and which previous assessments may need another look.
Give reviewers the evidence behind each match
A useful match should let a coordinator move directly into verification. Alongside the suggested trial, show the relevant criteria, the supporting record passages, their dates, and any unanswered questions.
Suppose a pathology report supports a biomarker requirement, but the available treatment history is incomplete. The reviewer should be able to see both findings immediately. A single overall score would obscure the distinction between a well-supported match and one that depends on information the site still needs to obtain.
Make room for “unknown” in the assessment. A missing result does not establish that a patient fails a criterion, and a reference to a planned procedure does not confirm that it happened. Preserving those distinctions helps staff decide whether to check another record, ask the care team, or move on.
It also makes feedback more useful. Instead of rejecting a match without explanation, a reviewer can identify the specific interpretation that went wrong. Over time, those corrections reveal where extraction, rules, or presentation need attention.
Use a pilot to learn where the system needs help
An initial pilot should include the kinds of records that make manual screening difficult. Clear-cut cases establish basic performance, but incomplete histories, conflicting notes, and complex treatment sequences show how the system handles uncertainty.
Have experienced reviewers assess a representative set of records and compare their findings with the system’s output. Look at the evidence behind disagreements. Did the tool miss a relevant passage, misunderstand its meaning, apply the wrong time window, or rely on outdated information? Each points toward a different correction.
Include records the system did not flag. Reviewing only suggested matches tells you little about potentially eligible patients it overlooks. Check whether performance differs across record formats, sites, or patient groups represented in the pilot, rather than relying entirely on an overall accuracy figure.
Once those initial checks are satisfactory, a period of use alongside the existing screening process can show how recommendations hold up in practice. Agree in advance what level of performance would justify broader use and which problems would require further testing.
Fit the recommendations into the site’s screening routine
Ask coordinators when a potential match would be most useful. For one site, that may be preparation for an upcoming clinic. For another, it may be a regular review of new diagnoses or updated test results.
Use that timing to shape delivery. A worklist should make it easy to distinguish new candidates from people already reviewed and explain why a previously assessed patient has reappeared. A newly available result may warrant another look; an unchanged record usually does not.
Decide who receives unresolved questions and how the care team becomes involved. AI can support prescreening, while authorized site staff confirm eligibility, determine appropriate next steps, and manage patient contact. Access to records and sharing of outputs should follow the site’s approved arrangements.
As use expands, track review time, useful matches, missed candidates, and reasons for rejection. Discuss the results with coordinators: which information helped them act, and what did they still have to find themselves? That conversation gives teams a concrete basis for improving the next version.
A successful implementation leaves sites with a repeatable way to identify candidates, check the evidence, and move appropriate patients toward screening. Build that process with the people doing the work, then expand as the results support it.
Continue the conversation at SCOPE Summit, March 1–4, 2027, in Orlando, Florida. Explore practical applications of AI, patient recruitment, and site operations with colleagues across clinical research. Explore the program and register.