Multi-Agent Based Automation of Consistent Causality Assessment (e.g., WHO-UMC)

Manual causality assessment is a high-variability task where physicians spend hours retrieving patient history, drug timelines, and lab data from fragmented sources. An agentic workflow automates evidence gathering and applies the WHO-UMC framework rules uniformly, reducing assessment time from 2-3 days to under 2 hours per complex case and ensuring every case is evaluated against the same objective criteria.

The bottleneck of physician review for causality limits overall case processing throughput. By deploying retrieval agents to auto-populate assessment worksheets and a reasoning agent to propose a score with justification, the workflow creates a 'review-ready' package. This allows PV physicians to focus on validation and sign-off, potentially doubling the number of cases a single reviewer can finalize per week.

Inconsistent or poorly documented causality assessments are a common audit finding. This automated workflow enforces a step-by-step, evidence-traced process. Every agent action—data retrieval, rule application, score proposal—is logged in an immutable audit trail. This creates a defensible record that demonstrates systematic, reproducible decision-making, significantly strengthening compliance posture.

PV physicians are a constrained, high-cost resource. Automating the evidence compilation and preliminary analysis allows them to operate at the top of their license. Instead of data gathering, they perform high-value oversight: validating agent logic, handling complex edge cases, and making final determinations. This improves job satisfaction and allocates medical talent to where it has the most impact.

As product portfolios grow or during safety crises, manual processes collapse under volume spikes. An orchestrated agent system (built with LangGraph or similar) scales horizontally. New agents can be added to handle increased data source load or parallel assessment queues without proportional increases in headcount. This creates a resilient operating model capable of handling fluctuating case volumes without compromising timelines.

Faster, standardized causality assessments mean aggregated safety data is richer and more timely. Consistent 'Certain', 'Probable', and 'Possible' labels across the entire case database improve the quality of signal detection analyses (e.g., disproportionality measures). This leads to earlier, more confident identification of true safety signals, directly impacting patient safety and risk management.

This agent is triggered upon case intake and autonomously queries connected clinical systems (EHRs via FHIR, EDC, patient registries) to retrieve longitudinal patient data. It extracts and timelines key evidence: drug administration dates, medical history, concomitant medications, and dechallenge/rechallenge outcomes. It normalizes this data into a structured timeline, eliminating hours of manual chart review and ensuring the assessment is based on a complete patient narrative.

The core logic agent applies the WHO-UMC framework deterministically. It ingests the structured timeline and evaluates against six criteria: temporal relationship, biological plausibility, alternative explanations, dechallenge/rechallenge response, dose-response, and previous knowledge. Using a rule-based engine supplemented by LLM reasoning for narrative interpretation, it proposes a causality score (Certain, Probable/Likely, Possible, Unlikely, Unassessable) with a confidence percentage and a clear, point-by-point rationale.

This agent transforms the scoring logic into a compliant, audit-ready case narrative. Using Retrieval-Augmented Generation (RAG) over internal SOPs and past approved assessments, it drafts a summary that explicitly links evidence to each WHO-UMC criterion. It highlights gaps in information (e.g., 'rechallenge not performed') and flags cases requiring senior review. This ensures medical writers receive a high-quality draft, cutting narrative writing time by over 50% while improving regulatory defensibility.

The central nervous system built on LangGraph manages the multi-agent workflow state. It sequences agent execution, passes context, and enforces business rules (e.g., 'if score is Certain, route directly to QP review'). It handles exceptions by routing cases with low-confidence scores or missing data to a human-in-the-loop queue for clarification. This orchestration layer is critical for maintaining a reliable, observable, and reproducible process that integrates with legacy PV systems like ARISg or Veeva Safety.

The mandatory human control point. The workflow does not auto-finalize assessments. All agent-proposed scores and narratives are presented to a qualified PV physician or QP via a review interface that shows the source evidence, the agent's reasoning, and a side-by-side comparison to historical similar cases. The physician can approve, modify, or reject the assessment. All changes are logged in an immutable audit trail, preserving final human accountability as required by EMA/FDA regulations.

Deployment involves containerized agent services communicating via a message broker (e.g., RabbitMQ). The primary integration is with the Safety Database (Oracle Argus, Veeva Safety) for case data and final storage. Secondary integrations pull from EHR/EDC APIs and internal document management systems. A rollout typically starts with a 4-6 week pilot on a single product line, measuring consistency gains and time savings before scaling. Critical success factors include high-quality training data for the reasoning agent and clear SOPs for physician review of AI-assisted outputs.