AI that actually finds more trial candidates: How MSK is scaling human-in-the-loop matching
The cancer center's first test of the technology matched all manually identified clinical trial candidates and surfaced additional appropriate patients. As one clinical IT leader put it: "The analysis was unambiguous: The approach works."
Memorial Sloan Kettering Cancer Center's mission is clear: end cancer. With as many as 1,800 open trials at any time, the team knows speed and precision in matching can change outcomes for real people.
The challenge
Clinical research runs on data, but much of what matters most lives in free-text notes, imaging narratives, outside records, and siloed systems. Even structured data like labs and vitals often fail to move cleanly from the EHR to sponsor systems.
"To find patient matches for our clinical trials, we have more than 400 research coordinators that spend countless hours combing through patient histories," said Joe Lengfellner, senior director of clinical research information technology at MSK. "Copy-and-paste has become an unfortunate but common bridge today."
The downstream effect: slower screening, more errors, and missed opportunities. AI scribes help clinicians, but they also generate more unstructured content, making the transformation step - structuring the unstructured - even more critical.
The proposal
MSK's Clinical Research Innovation Consortium, in partnership with the iHub, ran a head-to-head vendor evaluation focused on three pain points: patient matching, unstructured-to-structured extraction, and EHR-to-sponsor data movement.
"In mid-2024, we launched a head-to-head competition across multiple vendors and approaches," Lengfellner said. Triomics stood out with an oncology-tuned platform that reads unstructured clinical data at scale, extracts trial-relevant facts, and reconciles those against inclusion/exclusion criteria.
The bar was evidence. The team required a retrospective analysis across multiple completed trials: recover every patient coordinators found manually - and find any they missed.
Meeting the challenge
Triomics is connected to MSK's Epic EHR and research environments. Nightly pipelines ingest notes, labs, reports, and visit schedules. Models align patient-level features against trial criteria to surface high-confidence candidates, with alerts timed to upcoming visits.
"It's a human-in-the-loop approach where clinicians retain final eligibility decisions, and research coordinators remain central to quality control and longitudinal follow-up," Lengfellner said. Coordinators validate matches, correct extractions, and place near-matches on watchlists for ongoing monitoring.
With the retrospective analysis complete, MSK will run a controlled pilot across varied oncology subspecialties, then scale across more than 1,800 studies as performance gates are met.
Results
In the initial evaluation, the platform matched 100% of the patients MSK had already identified and found additional appropriate candidates. "The analysis was unambiguous: The approach works as intended and can widen access by reducing false negatives in screening," Lengfellner said.
Manual screening can take hours per patient. The platform processes hundreds of charts overnight. As adoption expands, coordinators can shift time to patient engagement, logistics, and data quality oversight. The expectation: faster matching, fewer missed opportunities, and better data hygiene across the portfolio.
What this means for healthcare leaders
- Treat unstructured data as a first-class citizen. Automate extraction for staging, biomarkers, prior therapies, and nuanced criteria hidden in notes.
- Insist on proof. Run retrospective studies against completed trials; measure precision, recall, and time saved before broad rollout.
- Keep humans in the loop. Let coordinators confirm, correct, and teach the system - with a feedback trail that improves models over time.
- Integrate with existing workflows. Push timely alerts to treating clinicians and coordinators; avoid adding clicks.
- Track operational impact. Monitor throughput, screening cycle time, near-miss recovery, and downstream data quality into sponsor systems.
- Plan for scale. Start with a controlled pilot, define performance gates, then expand by disease area and study phase.
How to get started
- Baseline your metrics: current screening time per patient, match rates, false negatives, and coordinator hours spent on manual review.
- Pick representative trials for a retrospective test - include variable complexity, data sources, and inclusion/exclusion nuance.
- Design your human-in-the-loop workflow on day one. Map who validates, how corrections feed learning, and how near-misses are tracked.
- Address interoperability early. Align data mapping with standards like HL7 FHIR to reduce friction as you scale.
- Keep patient access at the center. The goal is more eligible patients identified, sooner, with fewer administrative hurdles.
Why this approach works
Oncology trials have expanding criteria and complex histories. A system that continuously reads notes, reconciles eligibility, and flags near-misses gives teams a wider, more accurate view of who qualifies - without adding overhead.
The key is pairing automation with expert oversight. Technology does the heavy lifting; clinicians and coordinators make the call.
Advice from the field
"Start with a pragmatic approach and be evidence-driven," Lengfellner said. "Run a retrospective analysis before any broad deployment. Set rollout metrics, build human-in-the-loop workflows, and use the technology to reduce administrative burden and elevate expert work."
Resources
- Background on trial design and eligibility from the National Cancer Institute: NCI Clinical Trials Overview
Upskilling your team
If you're standing up AI capabilities across clinical operations or research, a focused learning plan helps. See curated AI programs by job role here: Complete AI Training.
Your membership also unlocks:
