- A Phase III trial excludes patients with "cardiac catheterization."
A chart says "central venous catheter placement" — an ICU line for IV access.
Generic AI sees "catheter," flags ineligible. An eligible patient, gone.
That one class of error is why trials miss enrollment. 🧵 - The two procedures share a word, not a meaning.
Cardiac catheterization (SNOMED CT 41976001) sits under "Procedure on heart."
Central venous catheterization (392230005) sits under "Catheterization of vein."
Different branches. One patient. Two answers. - This isn't hypothetical. Published evaluations confirm AI models make this exact "cardiac catheterization = central venous puncture" error (Fierce Biotech, 2025).
LLMs fixed keyword search's easy cases — not the ones where terms share words but differ medically. - Second failure: exception clauses.
"Exclude hypertension unless well-controlled on stable meds for 3+ months."
LLMs see "hypertension" and either exclude — losing an eligible patient — or skip the temporal check. Protocols now average 27+ criteria, many nested. (IQVIA) - Third failure: non-determinism.
Run the same patient through an LLM matcher twice, with slightly different context, and you can get two answers.
Regulators need to know exactly why each patient was in or out. "The model decided" is not an audit trail. - What ontology-driven matching does instead:
Map every term to a SNOMED CT concept ID. Reason over the hierarchy, not token proximity. Parse "unless/except" with deontic and temporal logic.
Same input, same output, every time — with a trace you can hand a regulator. - The economics behind this:
80% of trials miss enrollment timelines. Every day of delay costs ~$800K in lost sales (Tufts CSDD). Cardiovascular delays run $1.4M/day.
Screen failures cost ~$1,200 each — and 20–80% of screened patients fail.
The vendor map shares one blind spot. - Tempus (Deep 6) ships a 94%-accurate LLM matcher. http://IQVIA.ai, ConcertAI, Medidata — all probabilistic too.
94% sounds great until it's 6 in 100 patients misjudged, non-reproducibly. None ground eligibility in SNOMED CT graphs.
So we build that layer. - Custom matching that reasons over SNOMED CT graphs and deterministic logic, runs inside your firewall, and integrates to Rave / Veeva / Clinical One — for sponsors, CROs and AMCs where screen failure is measured in millions.
Honest question for trial ops teams: - when your matcher flags a patient ineligible, can you produce the exact medical reasoning — reproducibly — or only a similarity score?
That's where eligible patients get screened out — and you never know which. #ClinicalTrials #ClinicalAI - We wrote up the full failure taxonomy and the ontology approach here:
https://veriprajna.com/solutions/clinical-trial-recruitment-ai