Agentic Workflow for Automated SUSAR (Serious Unexpected SUSAR) Identification in Trials

Manual comparison of serious adverse events (SAEs) against the Reference Safety Information (RSI) is slow and error-prone under tight regulatory clocks. An automated agent continuously evaluates incoming EDC data, applying expectedness logic to flag potential SUSARs within hours of database lock. This compresses the detection-to-triage window from days to minutes, virtually eliminating the risk of a late report and the associated regulatory findings.

Study teams waste hundreds of hours per trial manually reviewing SAE narratives and patient histories to assess expectedness. The workflow uses a causality assessment agent to pre-populate a structured analysis—retrieving relevant protocol sections, prior events, and lab data—presenting a confidence-scored recommendation. This allows medical monitors to focus on validation and complex edge cases, not clerical review, dramatically increasing their effective capacity.

Manual expectedness decisions are subjective, leading to inconsistent application of the RSI across reviewers and sites. The agentic workflow enforces a standardized reasoning framework (e.g., WHO-UMC), documents all evidence used, and generates a clear audit trail for every assessment. This creates a defensible, reproducible process that strengthens your position during regulatory inspections and simplifies responses to health authority queries.

By structurally coding every SAE's expectedness relationship at the point of intake, the workflow creates a clean, query-ready dataset. Safety scientists can then run continuous disproportionality analysis (e.g., PRR) specifically on unexpected serious events, dramatically increasing the signal-to-noise ratio. This enables earlier detection of emerging safety issues within the trial itself, allowing for proactive protocol amendments or investigator communications.

The fully burdened cost of a PV associate manually processing a complex SUSAR case is significant. Automating the identification, data aggregation, and initial assessment directly reduces the labor hours required per case. This creates a scalable operating model where increasing trial volume or case complexity does not require linear headcount growth, directly improving the cost structure of your clinical development program.

An isolated detection tool creates new manual handoffs. This workflow is architected as an orchestrated layer (e.g., using LangGraph) where a flagged SUSAR automatically triggers the expedited reporting pathway. It populates CIOMS/MedWatch draft templates, routes for approval, and initiates E2B generation—turning a detection event into a submission-ready action without re-entry or context switching. The value is in the closed-loop automation, not just the initial alert.

This agent establishes real-time connections to EDC systems (e.g., Medidata Rave, Oracle Clinical) and safety databases, ingesting new SAE case forms and lab alerts. It normalizes disparate data formats, extracts key fields (patient ID, drug, event term, severity, outcome), and enriches records with Reference Safety Information (RSI) from the trial's Investigator's Brochure. The agent handles API polling, manages data quality flags, and places validated records into a processing queue, eliminating manual data aggregation.

The core logic agent applies rule-based and ML-driven assessment to determine if a Serious Adverse Event (SAE) is a SUSAR. It retrieves the patient's trial drug timeline and medical history, compares the event against the RSI's listed expectedness, and evaluates temporal and biological plausibility using WHO-UMC or company-specific causality frameworks. The agent generates a preliminary assessment with a confidence score and a clear rationale, standardizing a subjective and time-consuming manual review step.

A LangGraph-based state machine orchestrates the entire workflow, managing the lifecycle of each case from ingestion to final disposition. It sequences agent tasks (ingest → assess → code → route), manages approval gates, and handles exceptions. Based on the assessment engine's output, it automatically routes confirmed SUSARs to the expedited reporting pathway, routes ambiguous cases to a medical reviewer queue, and closes out non-SUSAR SAEs with an audit trail. This layer ensures process integrity and auditability.

For confirmed SUSARs, this agent automates the generation of compliant regulatory documents. It populates CIOMS I or MedWatch 3500A templates with case data, generates E2B (R3) XML messages, and performs schema validation against ICH and regional requirements (FDA, EMA). The agent then manages electronic submission via the appropriate gateway (ESG, EVWEB) and tracks acknowledgments. It integrates with deadline management systems to ensure reporting within the critical 7 or 15-day clock, preventing compliance lapses.

A critical control plane providing pharmacovigilance physicians and associates with a structured review interface. It presents the agent's assessment, supporting evidence, and confidence scores side-by-side for efficient validation. Reviewers can approve, override, or request more information. All actions are logged in a defensible audit trail. The queue prioritizes cases by deadline and risk score, ensuring human expertise is focused on high-value exceptions and final sign-off, not on initial data sifting.

This operational component provides dashboards for throughput, accuracy rates, and deadline adherence. It monitors agent performance, tracks the volume of cases auto-routed versus escalated, and logs all decisions for aggregate analysis. A key feature is a continuous feedback loop where reviewer overrides are used to retrain and calibrate the assessment engine's models, improving accuracy over time. It also monitors data feed health and system latency, ensuring the workflow operates as a reliable production service.