AI Agentic Workflow for Automated WHO Drug Dictionary (WHO-DD) Coding

Manual coders spend 5-15 minutes per drug entry cross-referencing trade names, salts, and dose forms against the WHO-DD. An agentic workflow uses a fine-tuned NER model to extract drug mentions and a retrieval-augmented reasoning agent to disambiguate against the dictionary, handling synonyms and investigational compounds. This cuts per-case coding labor by over 80%, directly reducing PV operational costs.

WHO-DD coding is a sequential bottleneck before MedDRA coding and narrative finalization. By automating this step inline with case intake, you compress the end-to-end ICSR processing timeline. The workflow integrates directly with safety databases (e.g., ARISg, Veeva Safety) via API, applying codes in real-time or in batch for backlogs, enabling cases to move to the next review stage without delay.

Manual coding leads to variance between coders and over time. The AI agent applies the same logic consistently, referencing the latest WHO-DD version and internal coding guidelines. It generates an audit trail with confidence scores and source rationale for each mapping. For 'not codeable' scenarios, it routes exceptions with context to a human reviewer, ensuring 100% quality control without 100% manual review.

Legacy cases with uncoded or poorly coded drug data are unusable for aggregate analysis. This workflow can be deployed as a batch remediation pipeline, processing thousands of historical cases to standardize drug data. This unlocks trapped data for signal detection, PSURs, and regulatory queries, turning a compliance liability into an analytical asset without taxing current staff.

The workflow is built on an orchestration layer (e.g., LangGraph) that manages the state of each coding request, coordinates the NER and mapping agents, and handles integration calls. It's designed to scale with case volume and can be deployed in a cloud-native environment adjacent to your PV system. This creates a durable automation asset that grows with your portfolio without linear cost increases.

The business case is clear: reduce FTEs dedicated to repetitive coding, reallocate PV expertise to higher-value signal assessment and medical review, and mitigate regulatory risk from coding inconsistencies or delays. Implementation typically involves a 4-6 week pilot on a specific product stream, validating accuracy and throughput gains before full rollout, delivering payback in under 12 months.

This initial agent parses incoming case narratives, intake forms, and EMR data to identify drug mentions, including investigational codes, trade names, and misspelled variants. It uses a fine-tuned NER model combined with a custom dictionary to extract candidate strings with high precision, normalizing dose forms (e.g., 'tab', 'capsule') and strengths. It passes structured entities—drug name, form, strength—to the mapping agent, eliminating the manual copy-paste step that consumes 15-30 minutes per complex case.

The core reasoning agent receives extracted drug entities and queries the WHO-DD API or local dictionary snapshot. It handles ambiguity (e.g., 'Prozac' maps to Fluoxetine) and selects the correct ATC code based on form and strength context. For combination products, it decomposes the name and maps individual components. It assigns a confidence score and, for high-confidence matches, auto-applies the code to the case in the safety database (e.g., Veeva Safety, Oracle Argus). Low-confidence matches are queued for review.

Not all drug names can be automatically coded. This component manages the workflow for ambiguous or novel compounds. Cases with low mapping confidence, new molecular entities not yet in the dictionary, or conflicting data are routed to a dedicated queue in the PV team's task management system (e.g., Jira, ServiceNow). The queue is prioritized by case seriousness and reporting deadline. The agent provides context—source text, extraction results, and mapping attempts—to accelerate the human coder's review, turning a 20-minute investigation into a 2-minute validation.

The central nervous system of the workflow. Built on LangGraph, this layer defines the state machine that orchestrates data flow between the extraction agent, mapping agent, safety database, and review queue. It manages retries, handles API failures gracefully, and maintains a complete audit trail of all decisions—which agent acted, what data was used, and why a code was assigned or escalated. This orchestration is deployed as a microservice, integrating via REST APIs with legacy PV systems, ensuring the automation layer is decoupled and maintainable.

A specialized agent for one-time cleanup projects or periodic backlog processing. It ingests thousands of historical cases with uncoded or poorly coded drug terms, runs them through the recognition and mapping pipeline in batch mode, and outputs a spreadsheet of proposed codes with confidence scores for bulk reviewer approval. This agent operates on a separate compute cluster to avoid impacting real-time case processing, enabling teams to remediate data quality issues without adding to operational load.

Continuous operational control. This component tracks key metrics: automated coding rate, accuracy vs. human gold standard, reviewer override reasons, and processing latency. Dashboards (e.g., Grafana) provide real-time visibility to PV ops. Drift detection triggers alerts if model performance degrades. A feedback loop automatically collects reviewer corrections into a training set, scheduling periodic retraining of the NER and mapping models to adapt to new drug launches and terminology shifts, ensuring the workflow's ROI compounds over time.