How it works
From protocol design to enrolled patient.
Three products on one federated engine — author the protocol, de-risk its eligibility criteria, then find and enroll the patients — and the hardest part, matching patients, happens without their data ever leaving the hospital.
Protocol authoring
Write or absorb a full, editable ICH-M11 protocol — versioned, amendable, exportable.
Protocol de-risking
Fix enrollment-strangling criteria before the protocol is locked, on real federated counts.
Enrollment intelligence
Find and enroll eligible patients across hospitals, privately, end-to-end.
01
Author or absorb the protocol
Start from the ICH-M11 / SPIRIT template, or upload existing documents — a draft protocol, synopsis, or investigator brochure. GhostTrials extracts the text, maps it into structured sections (tracking which source each block came from), and gives you a fully editable protocol you can version, amend, and export to PDF or DOCX.
02
De-risk the design
Before the protocol is locked, a federated dry-run shows — grounded in real de-identified patient counts across the hospital network — which eligibility criteria strangle enrollment, how many patients each relaxation would unlock, and a per-criterion data-completeness score. An explicit what-if editor tests the trade-offs; the output is a shareable Criterion Sensitivity & De-risking report.
03
Parse & approve criteria
An AI Protocol Parser extracts every inclusion and exclusion criterion into structured form — ontology codes (SNOMED, LOINC, RxNorm), biomarker gene/variant, value thresholds, temporal windows, negation. A regex/NLP floor guarantees a baseline; a human approves before anything matches.
04
Match — inside the hospital
For each site, a federated query runs against the local FHIR server. Patients are evaluated where they live. Only Bloom-filter pseudonyms and Laplace-noised counts (ε = 0.5) return — never names, dates of birth, or MRNs.
05
Resolve the ambiguous
Candidates are scored per-criterion. Confident matches advance; anything ambiguous — a stale ECOG, a conflicting biomarker report, missing staging — is routed to a human-in-the-loop queue with the exact blocking criterion and source evidence. A Conflict Resolution agent arbitrates against the protocol text and defers to a human below 0.75 confidence.
06
Orchestrate enrollment
Identified → outreach → pre-screen → eConsent → enrolled, tracked live. An eConsent agent drafts ICH-compliant documents; a Regulatory agent prepares IRB/HREC packages; a Site Re-router redistributes candidates when a site nears capacity. When a protocol amendment changes eligibility, the affected sites and re-consent needs are flagged automatically.
The agents
Eight specialized agents, one audit trail.
Protocol Author
Absorbed document text → clean ICH-M11 protocol prose.
Protocol Parser
Protocol → structured eligibility criteria.
Match Orchestrator
Ranks candidates with per-criterion explanations.
Conflict Resolution
Arbitrates disagreements; defers to humans when unsure.
Outreach Composer
Drafts coordinator-to-coordinator messages.
eConsent Generator
Produces ICH E6(R2)-compliant consent documents.
Regulatory Affairs
Assembles IRB/HREC submission packages.
Site Re-router
Redistributes enrollment as sites reach capacity.
Every agent run is recorded with its model, token usage, and decisions — so the system’s reasoning is always inspectable. Agents enrich a deterministic floor; the core flows work even without a model in the loop.