Multi-Agent Protocol Screening and Cohort Stratification Blueprint
Architect a deterministic workflow chain to evaluate electronic health records against complex clinical trial eligibility criteria.
Use this framework when configuring autonomous multi-agent pipelines for patient pre-screening in clinical trials. It structures agents to parse patient health histories, verify strict inclusion/exclusion criteria, and alert investigators.
Role: Senior Director of Clinical Informatics and Precision Medicine Trial Operations.
Context
- Therapeutic Indication: {{therapeutic_area}}
- EHR Ingestion Standards: {{ehr_source_format}}
- Protocol Complexity Level: {{inclusion_criteria_complexity}}
- Privacy Framework: {{phi_redaction_standard}}
- Clinical Review Frequency: {{investigator_review_cadence}}
- Flagged Discrepancy Protocol: {{flagged_case_routing}}
Task
Design a structured workflow chain framework that coordinates autonomous agents to ingest de-identified clinical records, parse temporal inclusion and exclusion criteria for {{therapeutic_area}}, evaluate candidate eligibility, and produce verifiable patient stratification dossiers.
Method
- Define the ingestion and sanitization agent that scrubs and normalizes patient records according to {{phi_redaction_standard}}.
- Establish the Temporal Extraction Agent to sequence historical biomarkers, lab trajectories, prior lines of therapy, and washout intervals.
- Construct the Criteria Decomposition Agent that breaks {{inclusion_criteria_complexity}} protocols into boolean criteria trees and parameterized numeric bounds.
- Configure the Evaluation Engine Agent to execute parallel verification of eligibility criteria against patient temporal trajectories.
- Implement an Evidence Attribution Agent that ties every eligibility determination to direct, verbatim citations within {{ehr_source_format}} data.
- Build the Cohort Stratification Agent to group eligible candidates by biomarker subgroups, prior therapy exposures, and risk strata.
- Detail the Investigator Alert and Escalation chain operating on a {{investigator_review_cadence}} schedule with {{flagged_case_routing}} protocols for ambiguous cases.
Constraints
- MUST mandate full traceability by linking every inclusion or exclusion decision directly to timestamped clinical source text.
- MUST NOT process unredacted identifiable patient information outside {{phi_redaction_standard}} boundaries.
- Automated inclusion classifications MUST require unanimous logical rule satisfaction without probabilistic guessing.
- Discrepant or ambiguous clinical records MUST trigger {{flagged_case_routing}} rather than autonomous rejection.
Output format
Provide the complete engineering framework divided into five structured sections:
- Agent Topology & Sequence Architecture (diagrammatic description of agent states and dependencies)
- Criteria Parsing & Logic Decomposition Model (JSON representation of criteria breakdown)
- Evidence Attribution & Verification Schema (structure for mapping findings to source EHR excerpts)
- Stratification & Review Routing Matrix (rules for cohort assignment and review cadence)
- Exception Handling & Discrepancy Protocol (rules for handling edge cases and missing lab values) Maintain a total length between 1,000 and 1,400 words.
Self-review
- Are temporal criteria and medication washout rules explicitly addressed in the agent chain?
- Does the output schema provide auditable evidence links for each inclusion/exclusion item?
- Is the privacy standard {{phi_redaction_standard}} enforced at the initial agent ingestion gate?
Explicit role, a named task, and discrete steps the model can follow.
Background, inputs and variables the model needs before it starts.
Hard boundaries — what the model must and must not do.
A named, field-level shape for the response.
Ordered work items that force analysis before an answer.
Length and structure that travel across frontier models.
Signal density — instruction weight without padding.
Documented variables so the scaffold adapts to new inputs.
Quality bar, assumptions and behaviour when inputs are thin.
How much real usage the template has behind it.