Clinical study deviation evidence assembly

The business situation

Illustrative reference architecture. Validate requirements for your organization.

Illustrative architecture: A clinical study operations team receives visit exceptions across sites, laboratory feeds, and electronic data capture. Reviewers need to distinguish missing data from a potential protocol deviation and document their assessment. The agents assist records management and operational review, not participant care or trial eligibility decisions.

Problem

Protocol amendments differ by site activation date, and late lab results can change the evidence available at review. Flattening these timelines into a summary risks applying the wrong requirement or implying a clinical conclusion unsupported by source data.

Why a multi-agent system?

Specialists independently align protocol versions, reconcile visit events, gather approved source references, and assemble the review packet. Their separation supports explicit evidence ownership and prevents a drafting agent from amending captured clinical observations.

Intended outcome

The intended result is a reviewable operational decision with source evidence and explicit ownership. Evaluate preparation effort together with source traceability and reviewer correction frequency. The proposed architecture does not establish regulatory compliance; study owners must map their actual procedures, systems, and oversight requirements before deployment.

Specialist agent team

Protocol version agent

Determine the applicable approved protocol and site activation interval.

Tools

• Protocol repository
• amendment registry

Authority boundary

Cannot interpret clinical exceptions or change approved requirements.

Visit reconciliation agent

Compare scheduled and recorded visit events with source timestamps.

Tools

• Clinical data capture read API
• visit calendar

Authority boundary

Cannot modify case report forms or impute missing observations.

Laboratory evidence agent

Match incoming results to permitted study subjects and visits.

Tools

• Laboratory feed reader
• specimen tracking API

Authority boundary

Cannot assess participant safety or recommend treatment.

Deviation packet agent

Draft the chronology, conflicting evidence, and reviewer questions.

Tools

• Quality case draft API
• provenance manifest

Authority boundary

Cannot classify final deviation severity or submit external reports.

Enterprise systems

Clinical data capture platform

Authoritative captured observations, queries, and correction history.

Study and protocol management systems

Supply site assignments, approved amendments, and effective activation dates.

Laboratory and quality systems

Provide specimen events and controlled review records without granting model write access to source evidence.

Architecture diagrams

flowchart LR
    N0["Clinical study deviation evidence assembly"]
    N1["Protocol version agent"]
    N2["Visit reconciliation agent"]
    N3["Laboratory evidence agent"]
    N4["Deviation packet agent"]
    N5["Evidence ledger"]
    N6["Human decision"]
    N7["Authorized adapter"]
N0 --> N1
    N0 --> N2
    N2 --> N3
    N1 --> N4
    N3 --> N4
    N4 --> N5
    N5 --> N6
    N6 --> N7

sequenceDiagram
    participant O as "Case orchestrator"
    participant S as "Clinical data capture platform"
    participant H as "Accountable reviewer"
O->>S: Read visit evidence revision
    S-->>O: Current study observations
    O->>H: Present deviation chronology
    H-->>O: Approve documented assessment
    S-->>O: Late corrected observation
    O->>H: Open linked amendment review

Runtime flow

1. 1

   Register exception

   Create a study-scoped event using subject pseudonym, site, visit, and source revision; preserve the original signal.

2. 2

   Resolve protocol interval

   Match the site's activation history to the event date and expose unresolved amendment applicability for reviewer attention.

3. 3

   Reconcile chronology

   Compare scheduled visits, collection timestamps, receipt timestamps, and late corrections without conflating their meanings.

4. 4

   Collect evidence

   Retrieve the minimum permitted study records and build a provenance manifest that points back to controlled originals.

5. 5

   Draft assessment packet

   Describe the possible deviation, missing evidence, and contradictory facts; reserve clinical significance and final classification for qualified reviewers.

6. 6

   Record disposition

   Save the approved assessment with reviewer identity and source versions; subsequent source corrections reopen dependent review sections.

Failure recovery

Security & decision boundaries

• Use study pseudonyms in prompts and isolate reidentification mappings from agent access.
• Enforce site and study entitlements on all source queries and exports.
• Preserve source records and corrections; agent drafts never overwrite captured observations.
• Require qualified review of deviation classification and any participant-impacting decision.
• Record amendment version, site activation interval, and evidence cutoff in every packet.

Operational metrics

Measurement definitions, not promised improvements.

Evidence traceability

Reviewed packet assertions with resolvable controlled source references divided by audited assertions.

Version resolution rate

Exceptions with unambiguous applicable protocol versions divided by all registered exceptions.

Review preparation time

Elapsed time from exception registration to reviewer-ready chronology, segmented by site and event class.

Reopened assessment rate

Assessments reopened for late or corrected evidence divided by completed assessments.

Architecture tradeoffs

Rollout plan

Archived study exercise

Reviewers validate amendment selection and chronology on representative exceptions with known dispositions.

Read-only site pilot

One study permits draft packets after pseudonymization, source access, and correction-history tests pass.

Controlled quality integration

Approved dispositions enter the quality system only after late-evidence reopening and audit reconstruction are demonstrated.

Connected concepts

• 12 — State Persistence
• 21 — Authorization-Aware RAG
• 28 — Deterministic Data Grounding
• 30 — Non-Repudiation Logs
• 38 — Data Lineage Tracking
• 44 — Human-in-the-Loop (HITL)
• 68 — Time-Weighted Vector Retrieval

Primary references

• NIST AI Risk Management Framework
• HL7 FHIR R5: Security and privacy considerations
• Temporal: Workflow execution