P212: AI-Driven Pharmacovigilance: A Longitudinal Analysis of Improved Patient Safety and Operational Excellence
Poster Presenter
Rishi Verma
Medical Director
Roche & Stethy Australia
Objectives
To demonstrate the clinical and operational impact of an AI-powered pharmacovigilance platform in automating high-volume adverse event detection, triage, and reporting from unstructured medical communications.
Method
Stethy AI was deployed at scale across Roche Australia to monitor over 15,000 monthly communications; the system utilized multi-modal frontier models to process unstructured data and cross-reference the TGA database for automated triage and real-time reporting of safety information.
Results
The 2026 operational data reveals that Stethy AI has reached a level of precision and scale that defines a new industry standard for patient safety. Currently processing over 15,000 communications per month; a volume previously requiring massive manual oversight; the system identifies adverse events (AEs) with a verified accuracy of 99.9%. This high-precision filtering is supported by an industry-leading false-negative (FN) rate of <0.1%, ensuring that virtually no critical safety signals are missed during the automated scan of complex, multi-modal data, including handwritten attachments.
Operational efficiency has been transformed, with the average triage time per communication reduced to a mere 3 seconds. This speed allows for the immediate identification and routing of high-priority signals, ensuring regulatory timelines are met with 100% compliance. The automation of these high-volume, repetitive tasks has resulted in the reclamation of 18.5 full-time equivalents (FTEs) per month. These saved resources have been strategically reinvested into proactive signal detection, deep-dive clinical analysis, and addressing unmet medical needs. The data confirms that as the system processes higher volumes, the iterative learning cycles further sharpen the model's ability to distinguish between complex off-label use cases and standard adverse event reporting, creating a self-optimizing safety ecosystem.
Conclusion
The results conclude that AI-driven pharmacovigilance is no longer a pilot concept but a foundational necessity for modern medicine access programs. By achieving 98.7% accuracy at high scale, Stethy AI mitigates the inherent risks of human error in manual data entry while providing a transparent, auditable trail that exceeds global regulatory requirements. The primary implication of this study is that large-scale automation allows pharmacovigilance teams to shift from administrative processing to expert clinical oversight. This evolution enhances patient safety outcomes by accelerating the speed of signal detection and enabling a more responsive approach to pharmaceutical risk management.
