Clinical cohorts & trial matching
On this page
"Which patients meet these criteria?" is a rule-matching question, and it sits at the center of clinical research, quality measurement, and decision support. Eligibility criteria are conditions over labs, diagnoses, demographics, and notes, exactly the mix a fast filter engine is built for.
The stakes🔗
Patient recruitment is the chronic bottleneck of clinical research: roughly 80% of trials fail to meet their enrollment timeline, and recruitment and retention are the cause of about 80% of trial delays (Power). Those delays are expensive. Estimates run from $600,000 to $8 million for every day a trial is held up (Applied Clinical Trials). Finding eligible patients faster, against criteria that are easy to adjust, moves that number directly.
How it works🔗
Eligibility criteria become a rule evaluated against the patient population. For example, "patients aged 40–75 with HbA1c above 7.0% (stored ×10), carrying diagnosis code 250 or 401, and not on the excluded medication 88." Lab values and age are number checks; diagnosis and medication codes are category checks; free text in clinical notes is matched by text search.
🤖 With AI in the loop
Clinical AI ships wrapped in rules. UC San Diego Health's deep-learning sepsis model scores every emergency-department patient hourly, but a nurse is alerted only when the score crosses a set threshold and eligibility rules pass (patient still in the ED, treatment not already given), a combination that cut sepsis deaths 17% (relative) across two hospitals. The TREWS early-warning system adds a provider-confirmation rule before any alert drives treatment. For trial matching, NIH's TrialGPT puts fast deterministic retrieval in front of the language model (pruning ~26,000 trials to ~500 before the model reads anything) and cut screening time by 42.6%. In Winnow, model scores and AI-extracted fields are columns; the thresholds, eligibility gates, and suppression logic around them are rules clinicians can read. More patterns: Deep Dive: AI + Rules Engines.
Why Winnow fits🔗
- The clinical data model maps cleanly: lab values as number checks, ICD/ATC
codes as category checks, notes as searchable text,
with exclusions expressed by
NOT. - Cohort building in one pass: evaluate the criteria across millions of patient records at once to assemble a cohort.
- One patient, many trials: run the question in reverse, matching a single patient against hundreds of trial-eligibility rules in microseconds, surfacing every study they qualify for at the point of care.
- Live data, stable answers: criteria and patient records can change while a search runs; every search sees one complete, consistent view of the data.
🧪 In simulation — 50,000 trial criteria, 500,000 patient screenings
To put numbers on the above, we built an eligibility book at roughly the scale of every actively recruiting trial in the world — 50,000 criteria rules: per-indication inclusion criteria, lab bands, age-and-sex arms, exclusion-heavy rules full of nots, and site feasibility screens — and screened 500,000 patient records against all of it, lab values held as exact decimals so no floating-point rounding can move an HbA1c cutoff. Three quarters of patients matched at least one recruiting trial, about five on average. Then we ran the same book in the other direction: one pass that placed a 100,000-patient population into every trial's candidate list at once.
What this is: a test run on generated criteria and generated patient records, not a clinical deployment — the trials and patients are invented, the speeds and costs are measured. It ran on one ordinary 16-core desktop computer, priced at a standard cloud rate of $0.043 per processor-hour running around the clock, the same basis as the Deep Dive: Cost Savings. Every match was double-checked against a slow, exhaustive pass over all 50,000 criteria before any stopwatch started — that slow way takes about 19 times longer — and the whole-base sweep was re-checked patient by patient against the one-at-a-time path. Treat the speeds as approximate, moving by about a fifth with machine load, and the counts as exact; the test ships with Winnow for your team to re-run on your own criteria.
Winnow evaluates the criteria; it is not a medical device and makes no clinical determination. A clinician reviews the candidates it surfaces.