Clinical trial patient matching, explained.

Clinical trial patient matching explained: one protocol's rules against every chart a site holds

From a protocol's rules to a referral: the five steps at a cancer center

1. The protocol arrivesVLT-411 has 24 rules: 12 inclusion and 12 exclusion, each with its own thresholds and time limits.
2. The study team confirms what the rules meanDoes maintenance therapy count as part of the same line of treatment? A physician answers before anyone applies the rule.
3. Someone checks charts against the rulesEvery patient against every rule: met, not met or unknown. This is where a blank turns into a lost patient, or a question.
4. The investigator decides who is approachedA physician reads the evidence and refers the patient to screening, holds, or rules the patient out.
5. Consent, then screening, then enrollmentThe patient says yes or no. Protocol tests confirm eligibility. Many who are eligible on paper never enroll.

Step 3 is the subject of this guide. The study, site and patients in this example are fictional.

Why a blank in the chart can lose a patient, and three more surprises

Inclusion, exclusion, look-back window and unknown: four terms to know before reading a chart

Twelve terms are defined in the summary and glossary.

Prescreening vs screening in clinical trials: one consent form separates them

In shortPre-screening is a look at existing records before consent. Screening is the protocol's own tests after consent, and only screening can confirm eligibility. Software, a coordinator or an investigator reading charts can produce a shortlist with evidence. None of them can produce an enrolled patient.

The human research office of the National Institutes of Health (NIH) draws the line in one sentence: pre-screening "happens before a person signs a consent form", its purpose is to see whether someone might qualify and is interested, and it "must not include research procedures". Screening to determine further eligibility happens after consent, unless an institutional review board (IRB, the committee that approves human research at a site) has waived it.6

An information sheet from the US Food and Drug Administration (FDA), on the books since January 1998, says something similar from the other side. Procedures done for a patient's clinical care may be performed and the results later used to judge eligibility without first obtaining consent. Any clinical screening procedure done only to determine research eligibility needs consent first.7 So a records review that orders nothing sits on the first side. Ordering a test just for the trial sits on the second, and it needs consent.

Pre-screening and screening side by side: timing, evidence, output and who decides

Inclusion and exclusion criteria: what 24 rules ask, and which document answers each

In shortA criterion is a fact, a threshold, a time limit and a document type. Only a minority of them sit in coded database fields. On the sample protocol, six of 24 do; the rest live in sentences. A published heart failure study found the same shape: coded data settled 5 of 6 inclusion criteria and 5 of 17 exclusions, so people read notes for the other 13.

Take the 24 rules of VLT-411, this guide's fictional Phase 2 study in advanced non-small-cell lung cancer. Some look easy: age 18 or older. Some are a number with a unit and a date: platelets at least 100 × 10⁹/L within 14 days. Some are a judgment a clinician wrote in a sentence: a tumor growing on scans despite a third-generation EGFR inhibitor. One depends on a block of tissue in a pathology department's storage.

The chart sorts them by the record that proves them, the first thing a coordinator should write down for any protocol.

The rules above are from the guide's fictional protocol, and 14 of the 24 are shown. I marks an inclusion rule and E an exclusion. NSCLC is non-small-cell lung cancer. ECOG is a 0-to-5 scale of how well a person can carry out daily activities. QTcF is a heart-rhythm interval. A strong CYP3A4 inducer is a medicine that sharply lowers the blood levels of drugs cleared by that enzyme.

How much a coded query can reach, measured on a heart failure trial

Before building their screening tool, a Mass General Brigham team counted what structured record data could settle for one heart failure trial. It could reliably settle 5 of 6 inclusion criteria and 5 of 17 exclusions. The remaining 13 of 23 went to trained, non-licensed staff doing manual chart review.9 Exclusions lean hardest on prose, because a past illness or a reason a medicine stopped is usually a sentence.

One study in one disease. Measure the split on your own protocols. It takes an hour with a highlighter.

Met, not met or unknown: how look-back windows, units and threshold direction decide a criterion

In shortEvery patient against every rule has three possible answers, and each sends work somewhere different. Three small mistakes produce confident wrong answers: reading old evidence as current, comparing numbers in different units, and forgetting which way an exclusion runs. The worked examples below use the guide's fictional case.

Why a May ECOG score cannot answer an October question

A protocol says how recently a fact must have been shown. VLT-411 asks for ECOG 0 or 1 within 28 days before screening. Another fictional patient, P-0412, had a score recorded on May 19. On October 13 that score is 147 days old. It may still be true, and it does not count. His status on this rule is unknown, and the fix is to record a new score at Thursday's visit.

Windows also close on patients who are eligible today. P-0883's evidence all counts on Thursday, October 15. Her CT, read September 22, stops counting on October 20. If her visit slipped a week, she would have an unknown again.

Platelets, creatinine clearance and the units that fool a query

Why one accuracy number says little about a 24-rule patient

Eligibility requires every rule to hold. If a reader is right 95% of the time on each of 24 independent rules, it is right on all 24 for a given patient only about 29% of the time (0.95 raised to the 24th power). Real errors are not independent, so treat this as a warning about arithmetic. Always ask whether a published or quoted accuracy is per criterion or per patient.

The rules that govern trial pre-screening, and what changed between 2025 and 2026

In shortNo single rule covers matching. Good clinical practice, the international standard for running trials, sets the investigator's duties and now names pre-trial screening records. FDA's human subject rules and the Common Rule, the US federal rule protecting research participants, govern consent. HIPAA governs the health information. FDA's 2025 and 2026 guidance pushes sponsors to widen eligibility, which can make criteria harder to apply by eye. Guidance on Diversity Action Plans, a sponsor's plan for enrolling underrepresented groups, is still a draft, and the law's plan requirement waits on its final version.

The names below: ICH is the International Council for Harmonisation, and E6(R3) is its good clinical practice guideline. CFR is the Code of Federal Regulations, the US rulebook: 45 CFR 46 is the Common Rule, 45 CFR 164 holds HIPAA's privacy rule, and 21 CFR 312 is FDA's rule on investigational drugs. EMA is the European Medicines Agency.

HIPAA and consent when pre-screening: preparatory to research, limited data sets and the date trap

In shortTwo bodies of rule apply at once. The IRB governs the research activity; HIPAA governs the health information. A review preparatory to research lets a researcher look without authorization, on a promise that nothing leaves. A file that can test a look-back window has kept dates, so it is a limited data set or identifiable data read in place, unless a statistical expert has documented that the risk is very small. Without that finding, "de-identified" is the wrong word for it.

Under 45 CFR 164.512(i), a covered entity (a hospital or practice, in this setting) may let a researcher use protected health information, the identifiable health information it holds, without authorization for a review preparatory to research. It must first obtain three representations.23

The second promise shapes how a tool can be built. A tool that copies charts to someone else's environment to read them is a harder conversation with your privacy office than one that reads them where they live. A vendor that handles protected health information on your behalf, for a function HIPAA regulates, is a business associate and needs a written agreement.27

Whether a research use fits that definition is a question for your privacy office and IRB. Contacting a patient is a separate question again, so write down who may do it and on what basis.

Why a 28-day window and safe-harbor de-identification cannot share a file

Under the safe harbor, all elements of dates except the year must be removed for any date directly related to an individual, along with ages over 89. The rule's other route to de-identification is a documented finding by a statistical expert that the risk of identifying someone is very small.4

A limited data set may keep dates. It excludes direct identifiers such as names, street addresses and medical record numbers, stays protected health information, and is used for research, public health or operations under a data use agreement.4 Eligibility is arithmetic on dates, so without an expert's finding, "identifiers removed, dates kept" is the honest description of a pre-screening view.

What a sponsor may see from a site's pre-screening

A sponsor needs to know whether a site has patients. It rarely needs to know who they are. A clean split is counts by site and week: records read, eligible, one question away, referred. A spreadsheet of candidate patients emailed to a recruitment vendor is the kind of disclosure that needs its own legal basis first. FDA's December 2025 sentence on real-world data carries the same condition: privacy maintained, permissions for access to identifiable data obtained and kept.16

Eight decisions for a site's pre-screening procedure

1. Which basis does each study use: 46.116(g), an IRB waiver, or consent?
2. Which HIPAA route: preparatory to research, a waiver of authorization, or a limited data set under agreement?
3. Who may read records, and is that person part of the covered entity's workforce?
4. Which records and clinics are in scope, written down, including satellite clinics?
5. Who approaches a patient, and who tells the treating clinician first?
6. What goes in the screening log, including a record of a patient who declined?
7. What the sponsor sees, and what it never sees?
8. Who answers when a sponsor or vendor asks for more than that?

Pre-screening errors: blanks read as no, stale evidence, misread amendments and thin screening logs

In shortThe damage comes in two shapes. Some errors show up at the screening visit, in front of a patient. Others never show up, because a missed patient leaves no trace. FDA's inspection findings on clinical investigators cover records, consent and protocol compliance, and the pre-screening record now sits inside that scrutiny.

What FDA inspectors cite at investigator sites

FDA's summary of common investigator observations includes inadequate or inaccurate case history records and study records, failure to follow the investigational plan, and informed consent issues.28 The list carries no counts, and we found no published count that names pre-screening. It is safest to treat the worksheet and log as records an inspector may read.

Missing data in health records: why ECOG goes blank, and three kinds of "not eligible"

In shortA chart's blank may only mean nobody wrote the fact down. Performance status, the ECOG score, shows it best: about 60% of records in one large oncology dataset had a coded score, and text extraction raised that to 73%. Patients and doctors gave the same score about two times in three, so a score is a judgment, and an old one says less. Separating the kinds of "no" in your own log is the cheapest measurement you can make.

The authors call performance status "a subjective measure with no gold standard of measurement".31 Scoring any reader against one person's rating measures agreement with that person. For a judgment like this, finding the recorded score and its date is more defensible than inferring one from "doing well".

Three kinds of "not eligible" hiding in one column of your screening log

No public figure says what share of "not eligible" entries belong to the third kind. The published funnel below counts ineligibility as sites recorded it, and those logs usually merge all three. To find your own share, take 50 consecutive "not eligible" entries, find the rule each failed on, sort them into these three, and note what question would have settled each third-kind entry. It takes an afternoon.

Underrepresented patients and the pre-screening list: which clinics get read decides who gets asked

In shortMost patients are lost before anyone asks them. In a pooled analysis of 8,883 cancer patients, a trial was unavailable for 55.6%, and 21.5% were ineligible for an available trial. Community programs enroll a fraction of what comprehensive centers do. Inside one institution, the list a coordinator has time to read works the same way: choosing the main-campus list decides who gets considered.

A list is a sample. Nobody can read several thousand charts against 24 rules by hand, so a coordinator reads the tumor-board list and new-patient notes. Those patients tend to be at the main campus, and that choice is rarely written down as a decision. Reading more clinics is one of the few levers a research office controls without changing the protocol or adding a site.

Two rules keep this work defensible. Count each group only as the record states it, and never infer race or ethnicity from a name, an address or a language preference. Report counts by site and week, and never send a sponsor a list of people.

FDA's December 2025 final guidance still recommends approaches to enroll "a representative population".16 Sponsors plan differently from one another, so ask yours what it expects. A count of who was considered, from where, is useful under any version.

What slow enrollment costs a sponsor and a site: daily trial cost, longer recruitment and trials that never finish

In shortA day of delay costs about $40,000 in direct trial cost on average, and more in Phase 3. Lost sales add $500,000 or $800,000 a day, depending on which of two figures from Tufts University's drug development center (Tufts CSDD) you read. Recruitment itself is a small budget line. Median recruitment in industry Phase 3 trials went from 13 months for trials completed in 2008 to 2011 to 18 months for 2016 to 2019, and about one adult cancer trial in five failed to complete.

Trials run in phases. FDA's rule describes Phase 1 as studying how a drug acts in the body, generally in 20 to 80 people; Phase 2 as testing whether it works, usually in no more than several hundred; and Phase 3 as gathering more evidence, in several hundred to several thousand. Phase 4 studies come with marketing approval or after it.35

Both Tufts sales figures come from one 645-drug dataset. Getz wrote about $500,000 a day in Applied Clinical Trials in June 2024; the August 2024 white paper says about $800,000.37,36 Both give $40,000 a day in direct Phase 2 and 3 cost.

Where a Phase 3 budget goes, and how small the recruitment line is

Two things follow. Spending more on finding patients is cheap compared with what it can shorten. And cost per patient is an output: a study that enrolls slowly spends more over more months, whatever the recruitment line says. Market reports on patient recruitment software give widely different figures with no visible dataset, so this guide gives no market size.

LLM clinical trial matching accuracy: what five studies measured, and what regulators expect

In shortLarge language models (LLMs, AI that reads and writes text) read single criteria about as well as trained staff, in published tests. One randomized trial shows more patients reached in the same staff hours. One deployment shows a tenfold fall in chart review with no eligible patient dropped on its validation trial. None proves a figure for a whole oncology protocol at your site. A person still has to decide, and a replay on your own closed trial is the fairest test.

Two words run through these studies. Sensitivity is the share of truly eligible patients kept. Specificity is the share of ineligible patients correctly set aside.

A miss costs a patient, silently; a false alarm costs a few minutes of review. For pre-screening, ask about sensitivity first.

What FDA, EMA and EU law expect of AI in this work

The limits on patient matching that remain after any model improves

How research coordinators pre-screen today: the people, the morning's lists and where the days go

In shortA coordinator reads short lists against rules held in memory, chases documents across departments, then asks an investigator who has clinic three days a week. Sites do this by hand, buy query tools that reach only coded fields, or buy software that reads notes. Sponsors sometimes pay. The questions for each approach are who reads, what it reads, and who keeps the evidence.

Coordinator to sponsor: who handles a pre-screen

A small survey of 85 current and former coordinators at academic centers linked staying to adequate pay and to a principal investigator who is engaged, respectful and collaborative.49 A coordinator who leaves takes the working knowledge of several protocols along.

Hand review, query tools, chart-reading software and paid outside help

P-0883 at Lakemont: from 1,212 records read overnight to a signed referral at 9:44 a.m.

In shortThis is the example from the top of the page, inside the product. Overnight, six agents read 1,212 records against VLT-411's 24 criteria. In the morning, Nadia Okarven, the senior research coordinator, works from a grid instead of a pile. One newly filed report turns P-0883's last unknown green. Dr. Samuel Ferravane, the principal investigator, reads the worksheet and signs.

1

The overnight read: 29,088 checks, each with its line

At 2:00 a.m. the criteria parser loads 24 checks from protocol v2.0, confirmed by Dr. Ferravane on October 6. The record reader goes through 48,216 documents from the main campus and three satellite clinics. The criterion checker runs 29,088 checks, and the question writer turns 244 unknowns, for the 142 patients not ruled out, into questions.

The ranker lifts the closest patients to the top. The run ends at 2:41 a.m.

2

The grid replaces the morning's lists

Tuesday, October 13. Nadia runs pre-screening for four lung trials. VLT-411's grid shows 9 patients eligible on all 24, 47 one question away, 86 two or three away and 1,070 with a criterion not met. Each dot is a criterion: green met, red not met, hollow amber unknown.

For exclusions, met means the exclusion is absent, and the screen says so.

3

An old score stays unknown, and becomes a question

P-0412, 58, main campus, meets every criterion but one. I6 asks for ECOG 0 or 1 within 28 days, and his last score is from May 19. The checker will not carry it forward. It marks I6 unknown and writes one question: "ECOG not recorded since May; ask at Thursday's visit".

Nadia can mark it asked, or record the answer when she has it, and the cell fills in.

4

A scanned outside report settles the last rule

P-0883, 64, Eastgate clinic, meets 23 of 24. I3 is unknown because the note says only "EGFR+ per outside records". Eastgate scanned in a three-page report at 4:40 p.m. on October 12, and the grid flags it as new. Nadia asks the record reader to read it.

It finds "EGFR exon 19 deletion (E746_A750del)", from a report dated August 14, 2026. Protocol section 5.1.3 accepts it by tissue test. Eligible on all 24 rises to 10, and one question away falls to 46.

5

Every criterion beside its line in the chart

Nadia opens the record. The 24 criteria run down the left, and the lines that decide them are highlighted in place: the September 30 clinic note (ECOG 1, no brain metastases on the September 24 MRI), the outside report, and the October 7 labs: ANC (a white blood cell count) 3.8, platelets 241, creatinine clearance 72 mL/min.

6

Dr. Ferravane reads the worksheet and signs at 9:44 a.m.

The screening-log writer drafts the pre-screening worksheet: each criterion as the protocol wrote it, the verdict, the line and the source with its date. Nadia sends it at 9:42 a.m. Dr. Ferravane, in clinic, chooses refer, hold or not eligible, and signs with his PIN and a note: "Approach at Thursday visit; consent then screening labs".

The decision goes into the screening log under his name.

The solution's design targets are 3.4 times the eligible patients found per site from the same records, and about 3 minutes to pre-screen a chart with every criterion and its line. In the sample data, 3.4× compares a projection of 17 eligible on this one day (10 now plus 7 expected from this week's answers, at the 70% rate similar questions closed over the last 90 days) with 5 found by hand last quarter. Treat both as targets to measure on your own trials.

Fourteen more screens, from the coordinator's question list to a sponsor's counts by site

In shortOne patient is the story. These screens show the rest of the work: the list of questions by visit, plain-words search over the records, the protocol beside its checks, an amendment waiting for the investigator, the funnel by clinic and group, and the site's own settings.

Six agents read and draft; the investigator confirms the rules and signs every referral

In shortEach agent has one job and limits it does not cross. None decides eligibility, contacts a patient or sends a patient to a sponsor. Four-eyes control, where a second person checks the work, runs on both ends: the coordinator sends what the agents drafted, and the investigator confirms the protocol's rules before any match and signs each referral after.

Criteria parser to screening-log writer: what each of the six does

Every verdict keeps its line and date. Agent steps and people's decisions are recorded apart on an audit trail you can export. Controls like these, role-based access, maker-checker, e-signature and an audit trail, support your compliance with 21 CFR Part 11, FDA's rule on electronic records and signatures, and good clinical practice; a product cannot be certified to them. Botminds is SOC 2 Type II audited, with information security managed to ISO 27001.

Nadia, Dr. Ferravane, Marcus Hale and Lena Ortiz in the screens

What the same records showed by clinic and group

Protocols and site records in; an eligibility grid, questions by visit, signed referrals and counts for sponsors out

In shortIt reads the protocol and the records a site already keeps, where they live, and leaves the coordinator a grid and a list of questions, the investigator a worksheet to sign, and the sponsor counts. Your record systems stay as they are.

Systems a pre-screening rollout connects to

For each, check whether it has an API, a file feed or only screens, and who owns it. Those owners set your timeline.

Your record systems, procedure and committees stay in charge

What the solution holds to on evidence, blanks and signatures

What slow pre-screening costs you in coordinator hours, empty slots and patients nobody asked

In shortStart from your own numbers. Nobody publishes how many eligible patients a site misses, because a missed patient leaves no record. Published figures frame the cost: what a day of trial delay costs a sponsor, how little of a trial's budget goes on recruiting, and what a chart takes by hand. Your screening log can tell you the rest in about a week.

Re-counting a screening log to find your own baseline

Patients nobody asked, sites that stay empty, and coordinators who leave

What pre-screening software returns when you count coordinator hours and extra enrollments against its yearly cost

In shortPut in your own numbers. The calculator asks how many charts you read, how long they take, what the questions add back, and what one more enrolled patient is worth to you. Every input marked illustrative is an assumption to replace. The sponsor's daily cost is shown apart, because it belongs to a different budget.

Reach, chart time and the log: where automated pre-screening saves

Charts a month, minutes a chart and the value of one more enrollment

A site reading 10 charts per trial per month across six trials would not cover a $100,000 yearly cost from time alone. Its case rests on reach: reading the charts it cannot get to today. The randomized trial measured reach: 35 enrolled against 19, with staff time split equally.5

How ready your research office is: trials, records, your log, your IRB and who owns accrual

In shortPick the answer closest to your site today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

A low score means re-counting your screening log first

A total can hide a gap. A zero on any question is the first thing to fix, whatever your band: a site scoring 13 with no owner of accrual still has no one to run the work.

What to ask clinical trial matching software about scanned reports, blanks, evidence and your records

In shortAsk every vendor the same questions, us included. The answers separate a filter over coded fields from a system that decides each criterion from any evidence and shows its line. Hold out for a replay on your own closed protocol.

We found little independent testing of commercial matching products, as of November 2026. The studies in section 11 are mostly academic builds. A replay on your own trial is the fairest comparison, and it measures your current process too.

Replay a closed trial, then run one open protocol at one clinic, and decide on sensitivity and patients found

In shortMeasure first, replay on a trial whose answers already exist, then widen one clinic at a time. Sensitivity, false alarms and eligible patients found decide it. A replay is no guarantee, and its number holds only until the next protocol or clinic, so keep measuring.

1. Re-count your log, 50 entries, into the three kinds of no, and time 20 charts. That is your baseline.
2. Settle four decisions: the source records (E6(R3) asks the investigator to define them before the trial), the privacy route, who may approach, and what a sponsor sees.13
3. Have an investigator confirm the checks for one closed protocol, criterion by criterion.
4. Replay it blind with records frozen as they stood at each decision, and compare with your log.
5. Go live on one open protocol at one clinic, with one coordinator. Test whether the questions can be asked on a real clinic day. Then add a clinic.

Sensitivity, false alarms and patients found on your closed trial

Keep the scoring honest with the note in section 8: for a judgment like performance status, your reference reader's answer is itself uncertain.31 Sample size matters: zero misses out of 20 eligible patients is weak evidence, so say how many there were.

Clinical trial pre-screening questions answered: consent, HIPAA, AI accuracy and eligibility

In shortThe questions research teams and buyers search for most, each answered on its own.