A ClinicalTrials.gov study exposes its eligibility as a single free-text block
(protocolSection.eligibilityModule.eligibilityCriteria) plus a few typed fields
(sex, minimumAge, maximumAge, healthyVolunteers). This skill turns that
prose into structured inclusion / exclusion criteria and matches each rule
against patient facts that OpenMed extracted — producing an explainable
eligible | ineligible | unknown verdict per criterion.
This is decision support, not enrollment. The output is a candidate list and a rationale for a clinician to review, never an automated eligibility decision.
searching-clinicaltrials and need its eligibility as
machine-readable rules.The typed gates are deterministic — apply them first. The free-text criteria need parsing into bullet-level inclusion/exclusion items.
# Study from ClinicalTrials.gov v2 (see searching-clinicaltrials)
elig = study["protocolSection"]["eligibilityModule"]
raw = elig["eligibilityCriteria"] # free text, often markdown bullets
sex = elig.get("sex", "ALL") # ALL | FEMALE | MALE
min_age = elig.get("minimumAge") # e.g. "18 Years"
max_age = elig.get("maximumAge") # e.g. "75 Years"
healthy_ok = elig.get("healthyVolunteers") # bool
def split_criteria(text: str) -> dict[str, list[str]]:
"""Split the prose into inclusion / exclusion bullet lists."""
sections, current = {"inclusion": [], "exclusion": []}, None
for line in text.splitlines():
low = line.strip().lower()
if "inclusion criteria" in low:
current = "inclusion"; continue
if "exclusion criteria" in low:
current = "exclusion"; continue
bullet = line.strip(" -*•\t")
if bullet and current:
sections[current].append(bullet)
return sections
criteria = split_criteria(raw)
Each bullet is a candidate rule. Structure it into a comparable predicate: condition present/absent, lab threshold, age/sex, prior-therapy, performance status (e.g. ECOG ≤ 2), pregnancy status, etc.
from dataclasses import dataclass
@dataclass
class Criterion:
kind: str # "condition" | "lab" | "age" | "sex" | "medication" | "other"
polarity: str # "include" | "exclude"
text: str # original bullet
target: str | None # e.g. "ECOG", "diabetes", "metformin"
op: str | None = None # "<=", ">=", "==", "present", "absent"
value: float | str | None = None
Build the patient profile from openmed.analyze_text outputs plus structured
demographics, then evaluate each criterion to a three-valued result.
patient = {
"age": 61, "sex": "FEMALE",
"conditions": {"type 2 diabetes", "hypertension"}, # OpenMed Disease spans
"medications": {"metformin", "lisinopril"}, # OpenMed Pharmaceutical
"labs": {"hba1c": 8.1, "ecog": 1}, # from a labs extractor
}
def evaluate(c: Criterion, p: dict) -> str:
if c.kind == "sex" and c.target:
return "pass" if p["sex"] == c.target or c.target == "ALL" else "fail"
if c.kind == "condition" and c.target:
has = c.target.lower() in {x.lower() for x in p["conditions"]}
ok = has if c.polarity == "include" else not has
return "pass" if ok else "fail"
if c.kind == "lab" and c.target and c.target.lower() in p["labs"]:
v = p["labs"][c.target.lower()]
cmp = {"<=": v <= c.value, ">=": v >= c.value, "==": v == c.value}
return "pass" if cmp.get(c.op, False) else "fail"
return "unknown" # fact not present → needs human review, never assume pass
Aggregate: a patient is a candidate only if every inclusion criterion is
pass (or unknown, flagged) and every exclusion criterion is not fail.
Surface the unknown items prominently — missing data is the most common reason a
real screen needs a human.
sex, minimumAge, maximumAge) — cheap, exact.Criterion (kind, polarity, target, op,
value). NER on the bullet via openmed.analyze_text finds the condition / drug
/ lab targets; numeric thresholds come from a regex/units pass.pass | fail | unknown.unknown facts that block a confident decision.openmed.analyze_text over the patient note
to populate conditions (Disease), medications (Pharmaceutical), and oncology
context; normalize via coding-icd10 / normalizing-rxnorm so comparisons are
code-based, not string-based.openmed.analyze_text over each eligibility
bullet to identify the condition / drug / lab the rule references, improving
target extraction beyond keyword spotting.eligibilityModule
already populated — this skill is the next stage.unknown as pass enrolls
ineligible patients; treating it as fail drops eligible ones. Surface it.openmed.clinical (see resolving-clinical-context) so negated/historical
mentions are not counted as present.mapping-loinc) helps.resolving-clinical-context