Trial eligibility screening is a translation problem. Protocol inclusion and exclusion criteria are written by regulatory affairs teams and medical writers in a precise, structured language: "patients aged 18 to 75 with a confirmed diagnosis of heart failure with reduced ejection fraction (HFrEF), LVEF below 40 percent, NYHA class II-IV, receiving stable guideline-directed medical therapy for at least 90 days." Patient records are written by clinicians for other clinicians: "56F, NYHF class III HF, known low EF (echo Jan '25: 34%), on max-dose carvedilol and sacubitril/valsartan."
The protocol says LVEF below 40 percent. The note says low EF. The protocol requires 90 days of stable therapy. The note says "max-dose" without specifying duration. Neither of these mismatches stops a skilled screener from correctly classifying the patient, but they do stop a keyword search engine, and they slow down any screening process that does not have a systematic way to resolve clinical shorthand to structured criteria.
The two-stage structure of most eligibility screening failures
When eligibility screening is done manually at scale, two distinct failure modes appear consistently in audits. The first is false exclusion: a potentially eligible patient is screened out because a clinical term in their record does not match the lexical form the screener was checking against. "Chronic kidney disease stage 3b" in the notes and "eGFR 30-44 mL/min/1.73m2" in the exclusion criteria describe the same patient population using different coding systems. A screener who checks for the phrase "CKD stage 3b" and does not also check the associated lab value will miss the connection.
The second failure mode is false inclusion, where an exclusion criterion is overlooked because the relevant clinical history is buried in a section of the chart that was not reviewed for that specific criterion. A contraindicated prior medication documented in a hospital discharge summary from three years ago, not carried forward to the active problem list, will not be found in a screening process that only reviews current medications. That patient may pass initial eligibility screening and fail rescreening at the point of enrollment, which has cost implications for both the site and the patient.
Why structured extraction differs from keyword search
Most clinical informatics systems that support eligibility screening still rely heavily on keyword matching against structured fields: ICD codes for diagnoses, NDC codes for medications, LOINC codes for labs. This approach works well for criteria that map cleanly to structured fields and for institutions with high coding quality. Its limitations appear when criteria require clinical context that is only available in notes.
Structured extraction operates differently. Rather than asking "does this record contain ICD-10 I50.20?" it asks "does this record contain evidence of a heart failure diagnosis, what kind, and what is the ejection fraction evidence associated with it?" The extraction unit is a clinical concept with associated attributes, not a code match. This distinction matters for three common eligibility scenarios: criteria that require modifier information the code does not carry (severity, laterality, acuity), criteria that require documentation from a specific time window (history of X within the past 12 months), and criteria that require the absence of a condition rather than its presence.
Mapping protocol language to extractable entities
The first step in using note extraction for eligibility screening is translating each criterion into an extractable query. Not all criteria are best handled by note extraction. Demographic criteria (age range, sex), lab value thresholds (eGFR, hemoglobin A1c), and vital sign criteria are usually more reliably obtained from structured fields than from notes. Note extraction adds value for criteria that require clinical narrative: diagnosis history and severity, medication history including prior medications no longer active, procedures performed with specific techniques or outcomes, and clinical findings documented in examination notes.
For each note-extractable criterion, we define: the target entity type (diagnosis, medication, procedure, finding), the required attributes (NYHA class for a heart failure criterion; ejection fraction value for an LVEF criterion), the required time window (current, within 90 days, ever), and the required relationship to treatment (naive to therapy, previously treated, currently on stable dose). This translation step is done collaboratively with the clinical team before extraction configuration begins. Ambiguous criteria, the ones that could be interpreted two ways by a skilled screener, are flagged for explicit decision by the principal investigator before the extraction logic is finalized.
A concrete screening example: heart failure trial, 280 patients
A site running a phase II trial in HFrEF needed to pre-screen 280 patients from their cardiology patient population before contacting potential candidates. The protocol had eight inclusion criteria and twelve exclusion criteria. Of the twenty criteria total, six required note review beyond structured fields: NYHA classification, current GDMT for at least 90 days, specific exclusion for prior cardiac resynchronization therapy (not always coded), exclusion for severe hepatic impairment documented in notes, exclusion for active malignancy, and exclusion for a contraindicated co-medication with narrow therapeutic index that the pharmacy system did not flag automatically.
We configured extraction to target those six note-based criteria across the 280-patient record set. The extraction run took approximately four hours for the full corpus. The output was a per-patient candidate report showing, for each criterion, whether extraction found supporting or contradicting evidence, the source note and sentence, and the confidence score. Patients where all note-based criteria resolved cleanly above the confidence threshold were pre-classified. Patients with at least one uncertain criterion were flagged for analyst review.
Of the 280 patients, 141 resolved cleanly and did not require note-level manual review for the note-based criteria. The remaining 139 had at least one flagged criterion requiring review, but the analyst review was focused: they reviewed only the flagged criterion and the cited source note rather than conducting a full chart review. Pre-screening for all 280 patients was completed in 3 days rather than the 12 days the site had estimated for full manual pre-screening. Fourteen patients who would likely have been missed in a keyword-only pre-screen were identified as eligible through note-level diagnosis evidence, three of whom enrolled.
The exclusion criteria challenge: absence of evidence
The conceptually tricky half of eligibility screening is exclusion criteria, specifically those requiring the confirmed absence of a condition. "No history of atrial fibrillation" and "no prior stroke" are common exclusion statements. Clinical note extraction can find positive evidence for these conditions. It cannot certify their absence from the medical record, because absence in the extracted record does not equal absence in reality.
We are clear about this boundary with the teams we work with: note extraction for exclusion criteria identifies patients who have positive evidence of an exclusion condition, which supports automatic exclusion, and patients who have no extracted evidence, which does not support automatic inclusion but reduces the manual review burden by eliminating confirmed excludes from the manual review queue. The final decision on a patient with no extracted evidence of an exclusion condition still requires a trained screener to confirm, but they are reviewing a shorter list that does not include the definitively-excluded patients.
Screening documentation and audit trail
Clinical trial eligibility screening requires documentation under GCP. Every inclusion and exclusion criterion assessment must be documented, traceable, and reproducible. Note extraction outputs an audit trail by design: every extracted entity has a source note, a note date, a specific source sentence, a model version, and a confidence score. When a screening decision is based on extracted evidence, the documentation package includes the extraction output with citations, not just the screener's judgment.
For sites that have experienced screening documentation deficiencies during monitoring visits, the explicit citation trail from structured extraction resolves the most common finding: "basis for eligibility determination not documented." The extraction output is the documentation. It does not replace the screener's final judgment, but it provides the evidentiary basis for that judgment in a form that satisfies monitoring requirements without requiring the screener to reconstruct their thought process from memory after the fact.
