Matching Patients to Clinical Trials with AI

Introduction

Finding the right patients for a clinical trial is often slower than the trial itself. A research coordinator may spend 30 to 45 minutes reviewing a single patient's chart against a page of eligibility criteria, and that work multiplies across every trial a site runs. Slow recruitment delays every downstream stage of research, including the therapies that eventually reach patients.

AI clinical trial matching tools aim to compress that screening step by reading electronic health records against trial criteria automatically. This article looks at what independent evaluations show, distinct from the drug discovery process covered in Blog 11, and how organizations can judge whether a matching tool is ready for real use.

What the Tools Actually Do

1. Extract structured and unstructured data from a patient's EHR, including diagnoses, labs, medications and clinical notes.
2. Compare that data against trial eligibility criteria, which are often written in dense, ambiguous free text.
3. Score or rank eligibility at the criterion level and the trial level, flagging likely candidates for human review.
4. Present ranked matches to research coordinators or referring physicians.

Retrieval-augmented approaches, which ground each eligibility decision in the specific trial protocol text rather than relying purely on a model's memory, have become common because trial criteria change often and must be applied exactly as written.

What the Evidence Shows

A randomized evaluation published in Nature Communications in early 2026 compared human review alone, AI alone and human-AI collaboration for oncology trial prescreening using real-world patient records. The researchers found AI support was particularly helpful for complex eligibility criteria such as tumor biomarkers and cancer staging, areas where manual review is especially error-prone, and concluded that AI can augment existing screening processes without replacing the people doing them.

The NIH's open-source TrialGPT system, evaluated in peer-reviewed research, retrieved over 90% of relevant trials for a patient while requiring review of less than 6% of a site's full trial list, and cut screening time by roughly 40% in pilot testing. A separately validated commercial system (TRIAGE) reported strong sensitivity, specificity and criterion-level agreement with expert clinical research coordinator review in a retrospective oncology study, and was also able to surface enrollment opportunities that had been missed under standard workflows.

These are genuinely encouraging results. They also share a common thread: every study evaluated AI as a screening aid that narrows the pool for human decision-making, not as an autonomous enrollment system.

Where Matching Still Struggles

• Ambiguous or poorly structured criteria. Trial protocols are written for humans and often contain implicit assumptions that are hard to encode precisely.
• Data not captured in structured fields. Biomarker results, prior treatment history and performance status often live in free text or scanned documents.
• Protocol version control. Criteria change during a trial's life, and a matching system must stay synchronized with the current version.
• Generalization across sites. A tool tuned and validated at one health system may need local recalibration before results transfer directly.

Practical Use Cases Beyond Enrollment Speed

• Widening access. Automated pre-screening across a full patient population can surface eligible patients who would never have been proactively referred, which several studies link to more diverse trial participation.
• Site and feasibility assessment. Predictive analytics applied to historical trial performance and real-world patient data can help sponsors judge which sites are likely to recruit successfully before a trial even opens.
• Reducing coordinator burnout. Narrowing hundreds of charts to a short, ranked list changes a manual search into a review task.

Implementation and Evaluation Checklist

• Validate locally against your own patient population and past trials before relying on rankings.
• Track sensitivity and precision separately, since missing eligible patients and wasting time on ineligible ones carry different costs.
• Keep a documented human sign-off step for every enrollment decision.
• Monitor for equity effects. Confirm the tool doesn't systematically under-flag patients from groups underrepresented in its training or validation data.
• Plan for protocol updates as a routine maintenance task, not a one-time setup step.

Data and Privacy Considerations

Trial matching tools process identifiable health information to compare against protocol criteria, so the same governance questions that apply to any clinical AI system apply here: where data is processed, whether external vendors retain it, and what agreements govern its use. Research use of PHI carries its own regulatory considerations distinct from routine clinical care, and institutional review boards and privacy officers should be involved early.

Bottom Line

AI clinical trial matching has moved from a promising idea to a validated screening aid in several published studies, with meaningful time savings and improved handling of complex criteria. The strongest evidence supports human-AI collaboration rather than automation, and local validation remains essential before a tool is trusted to shape who gets offered a trial.

This article is educational and does not provide medical advice.

Suggested Internal Links

• RAG applications
• healthcare software solutions

Sources / References

• "Study shows AI can help identify more patients for cancer clinical trials," Winship Cancer Institute, Emory (summarizing Nature Communications study), 2026:
• "Expert Validation of an Artificial Intelligence-Enabled Trial Matching Solution Using Real-World Data From Patients With Cancer," PubMed:
• TrialGPT performance summary (secondary source, citing peer-reviewed study):

CTA

Whether you're building research analytics tools or integrating a trial-matching system with your EHR, Techverse Solutions can help design the data pipeline and validation approach. Book a consultation to discuss your needs.