Keiji AI LogoKeiji AI

Clinical Operations guide

AI patient recruitment from EHR data: matching patients to trials, and Deep 6 AI alternatives

How AI finds trial patients in EHR data, the published accuracy of LLM matching, the kinds of vendors (including Deep 6 AI, now Tempus) and what to ask before buying.

AI recruitment tools read structured EHR fields and clinical notes, check each eligibility criterion, and return a ranked list of likely candidates for a coordinator to confirm. Vendors split into two models: networks that bring patients and sites, such as Tempus (which acquired Deep 6 AI in 2025), TriNetX and IQVIA; and software that runs on the hospital's or sponsor's own records, such as TrialMind. Published LLM matching approaches expert accuracy (TrialGPT: 87.3% per criterion, 42.6% less screening time), but every study keeps a clinician in the loop. Choose by whose patients you need and where the records may go.

How AI finds trial patients in EHR data

Most eligibility facts sit in two places. Structured fields hold diagnoses, labs, medications and procedures. Clinical notes hold the rest: biomarker results, prior lines of therapy, performance status and the reasons a drug was stopped. Keyword search over structured fields misses much of the second group.

AI recruitment tools combine both:

  1. Turn the protocol into criteria. Each inclusion and exclusion becomes a check that can be run.
  2. Filter on structured data. Diagnosis codes, age, labs and treatment history cut the pool quickly.
  3. Read the notes for what is left. A language model judges each remaining criterion and cites the sentence it relied on.
  4. Rank and hand off. Candidates go to a study coordinator with the evidence attached, and the coordinator confirms before anyone is contacted.

The result that matters for operations is the screening funnel: how many patients each criterion removes. It shows which criterion is the bottleneck and whether a protocol change would help.

How accurate is LLM patient matching?

Three studies give a fair picture.

  • TrialGPT (Jin, Wang et al., Nature Communications, 2024), a National Institutes of Health project that Keiji AI's founders co-authored, reached 87.3% criterion-level accuracy on 1,015 patient-criterion pairs, against 88.7% to 90.0% for experts. In a pilot user study it cut screening time by 42.6%.
  • RECTIFIER (Unlu et al.) used GPT-4 with retrieval on clinical notes for a heart failure trial. Overall eligibility sensitivity and specificity were 92.3% and 93.9%, against 90.1% and 83.6% for trained study staff (medRxiv preprint, 2024).
  • A JAMIA 2024 study showed smaller open models fine-tuned on synthetic data matched GPT-3.5 on eligibility judgments, which matters when notes cannot leave the hospital.

All three keep a person in the loop, and TrialGPT was not tested on full EHRs, lab values or imaging. Before using any tool at scale, measure it on a sample of your own charts against a coordinator's review. For the step-by-step method, see AI for clinical operations.

Two vendor models: bring patients, or work on your records

"AI patient recruitment" covers two different businesses. Decide which one you need before comparing features.

ModelExamplesWhat you getFits when
Network: the vendor brings patients or sitesTempus TIME (with Deep 6 AI), TriNetX, IQVIAAccess to provider sites and their patients, or outreach run for youYou need more patients or sites than you have
Software on your recordsTrialMindMatching and screening funnels on records you already hold, in your environmentYou are a hospital, academic medical center, or sponsor with data to search

Network vendors differ among themselves. Tempus TIME enrolls oncology trials through Tempus's provider network and uses Deep 6's EMR patient finding, which mines structured and unstructured records. TriNetX is a federated network of 14,200+ sites where patient data stays at each source. IQVIA is a CRO that runs AI-targeted outreach on its real-world data.

Alternatives to Deep 6 AI

Deep 6 AI was acquired by Tempus in March 2025 and its EMR patient finding now sits inside the Tempus TIME network (Tempus announcement). Teams looking for alternatives usually want one of three things:

  • Matching outside oncology, or outside a single network. TriNetX covers many indications across its federated sites.
  • Recruitment run for them. A CRO such as IQVIA takes on outreach and site work.
  • Matching on their own records, with nothing leaving the building. TrialMind runs in the customer's environment and turns each protocol into a screening funnel with the reason for every exclusion.

See TrialMind vs Tempus (Deep 6 AI) for a side-by-side on data, network and where patient data stays.

How TrialMind matches patients

TrialMind's patient matching converts a protocol's eligibility criteria into rules and runs them on the customer's patient data, with counts at each step and the reason for each exclusion.

Screening against the TROPION-Lung17 protocol narrowed 1,001 patients to 164 candidates (16.4%). Prior treatment was the largest single exclusion, which is the kind of finding that feeds back into criteria design.

  • Deployment: inside the hospital's or sponsor's environment. Raw patient records are not moved.
  • Review: results go to the study team; clinical decisions stay with people.
  • Not included: TrialMind is not a patient-outreach service and has no provider network.

Questions to ask a recruitment vendor

  1. Whose patients? Your records, the vendor's network, or outreach to the public.
  2. Where do records go? Get in writing whether any patient-level data leaves your environment.
  3. Notes, not only codes. Does it read clinical notes, and does each verdict cite the sentence it used?
  4. Unknown is an answer. Can it say a criterion cannot be judged from the record?
  5. Your accuracy, not theirs. Test on a sample of your charts against coordinator review before rollout.
  6. Screening funnel. Can you see how many patients each criterion removes?
  7. Audit trail. Is every verdict and reviewer override logged?
  8. Therapeutic area. Oncology-only networks may not help in other indications.

Frequently asked questions

Which companies offer AI for clinical trial patient recruitment from EHR data?
Network vendors include Tempus (which acquired Deep 6 AI), TriNetX and IQVIA. TrialMind, from Keiji AI, is software that runs patient matching on a hospital's or sponsor's own EHR data inside their environment.
What are the alternatives to Deep 6 AI?
Deep 6 AI is now part of Tempus. Alternatives include TriNetX for federated EHR matching across many indications, CROs such as IQVIA for outsourced recruitment, and TrialMind for matching on your own records without moving them.
How accurate is AI patient-to-trial matching?
Close to experts in published tests. TrialGPT reached 87.3% criterion-level accuracy against 88.7% to 90.0% for experts and cut screening time by 42.6% in a pilot. Keep a coordinator's confirmation step.
Can AI match patients without moving EHR data out of the hospital?
Yes. Federated networks keep data at each site, and tools deployed in the hospital's environment, such as TrialMind, run where the records already sit.

See TrialMind on your own question

A 30-minute walkthrough on a study like yours, with sources and code you can check.