Design of an Offline Agentic Decision Support System for Pharmacovigilance Case Processing Using Rule-Augmented Local Language Models

Pharmacovigilance (PV) involves continuous monitoring of adverse drug reactions through structured workflows such as case intake, coding, causality assessment, and regulatory reporting. These processes are highly manual, repetitive, and prone to variability. This study proposes an offline agentic decision support system that combines deterministic rule-based logic with a locally deployed language model to automate template selection in PV case processing. The system processes structured and semi-structured inputs such as adverse events and suspect drugs, applies predefined rules, and retrieves appropriate templates while providing explainable outputs. The proposed hybrid system demonstrates improved accuracy (93.2%), increased consistency (100%), and more than 90% reduction in processing time compared to manual workflows. The offline architecture ensures data privacy, regulatory compliance, and suitability for real-world pharmacovigilance environments.