trial-eligibility-agent

Clinical Trial Eligibility Agent

ID: biomedical.clinical.trial_eligibility Version: 1.1.0 Status: Production Category: Clinical AI / Trial Matching

---

Overview

The Clinical Trial Eligibility Agent automates patient-to-trial matching by parsing eligibility criteria and evaluating patient data against inclusion/exclusion rules. It reduces manual screening time while preserving traceability and highlighting data gaps.

---

Inputs

Field	Type	Notes
trial_id	str	ClinicalTrials.gov NCT number or sponsor protocol ID
patient_summary	str	Narrative summary of key facts
patient_structured	dict	Optional FHIR bundle or structured JSON
data_sources	list[str]	e.g., clinical_notes, labs, imaging, meds

---

Outputs

Eligibility Report

• Inclusion criteria status (MET / NOT MET / UNKNOWN)
• Exclusion criteria status (MET / NOT MET / UNKNOWN)
• Evidence snippets and confidence per criterion
• Overall recommendation: potentially_eligible, not_eligible, or needs_more_information
• Data gap checklist

JSON Schema (Recommended)

{
  "trial_id": "NCT00000000",
  "criteria": [
    {"id": "I-01", "text": "Age >= 18", "status": "MET", "evidence": "Age 58", "confidence": "high"}
  ],
  "eligibility_summary": "potentially_eligible",
  "data_gaps": ["Latest ECOG score"],
  "alerts": ["Confirm brain MRI within 30 days"]
}

---

Workflow

1. Acquire protocol - Pull eligibility section from ClinicalTrials.gov or sponsor PDF.
2. Parse criteria - Normalize to structured logic (AND/OR, thresholds, units).
3. Extract patient facts - Convert notes and FHIR data into a unified feature map.
4. Evaluate criteria - Assign MET/NOT MET/UNKNOWN with evidence.
5. Summarize gaps - List missing labs, imaging, or biomarker data.

---

Example Prompt

Patient summary:
- 58-year-old female
- Stage IIIA NSCLC
- EGFR exon 19 deletion
- Prior osimertinib, progressed after 14 months
- ECOG 1
- No brain metastases
- Creatinine clearance 72 mL/min

Check eligibility for NCT04487080 and list MET/NOT MET/UNKNOWN criteria.

---

Guardrails

• No final enrollment decisions: results are advisory only.
• Cite evidence for each criterion.
• Surface unknowns rather than inferring.
• PHI handling: use de-identified data and HIPAA-compliant environments.

---

Integration Patterns

FHIR-Based Matching

from fhir.resources.patient import Patient
from fhir.resources.condition import Condition

def extract_patient_features(fhir_bundle: dict) -> dict:
    features = {}
    for entry in fhir_bundle.get("entry", []):
        resource = entry.get("resource", {})
        if resource.get("resourceType") == "Patient":
            patient = Patient.parse_obj(resource)
            features["age"] = calculate_age(patient.birthDate)
        elif resource.get("resourceType") == "Condition":
            condition = Condition.parse_obj(resource)
            features.setdefault("conditions", []).append(
                condition.code.coding[0].display
            )
    return features

---

Dependencies

requests>=2.28
fhir.resources>=6.0
pandas>=1.5

---

References

• TrialGPT (NIH)
• Jin et al. "TrialGPT: Matching Patients to Clinical Trials with Large Language Models" (2023)

---

Author

MD BABU MIA Artificial Intelligence Group Icahn School of Medicine at Mount Sinai md.babu.mia@mssm.edu