药物安全创新的下一个前沿：人工智能支持的信号管理

The Next Frontier of DrugSafety Innovation: AI-Supported STEPHANIE SENN,Product Manager, Vigilance Signal, IQVIA Table of contents IntroductionOpportunities for innovation in the signal management spaceWhere could AI implementation benefit safety signal management?How is IQVIA approaching AI in safety signal management?What’s on the horizon?ReferencesAbout the authors Introduction Safety signal management is the process of detecting and assessing adverseevent data and relevant supporting pharmacologic, clinical, and epidemiologicevidence to determine if there is a new risk associated with a medicinal product The rapid growth of data volume and source complexity,coupled with advances in technology has led pharmacompanies, regulators, technology providers, and Agentic AI offers a promising approach for assistingsignal management. These systems can autonomouslyadapt to new data, prioritize tasks, and proactively been the standard data sources, regulators and industryhave demonstrated successful use of Real-World Data(RWD) in specific scenarios, including in retrospectiveobservational studies utilizing electronic health recordsas a supplementary source. With advances in AI, it is Opportunities for innovation in Technological advances and the desire to reducethe operational burden of signal management have 1. A growing regulatory emphasis on improving signal 2. Increased interest in utilizing supplementary data Across the industry and all stages of the drugdevelopment and manufacturing lifecycle, AI tools arebeing implemented to increase efficiency and reduce resource overload. In June 2025, the FDA launchedElsa, a GenAI tool to help FDA employees work moreefficiently; thus far, Elsa has been used to speed upclinical protocol reviews and evaluations.1From a drugcompany perspective, GSK has implemented PVLens, 3. Increased industry adoption of AI 4. A desire to move toward widespread use of Data mining algorithms and manual data review havelong been the accepted approach of regulators todetect and validate signals. Regulators and MarketingAuthorization Holders (MAHs) have researched andrefined disproportionality algorithms over the years, but In an IQVIA-commissioned survey and study conductedby IDC, real-time signal detection is forecasted to become the primary focus area for automation initiativesin the next two years.3We have already started toimplement AI solutions to support signal management,but within the near future we can see the potential to Over the past ten years, regulatory-led and -backedinitiatives including IMI PROTECT, DARWIN EU, and FDASentinel have paved the way for continued research and improvements in this area. In 2025, a newinitiativefrom the Innovative Health Initiative (IHI) consortium The availability of data has increased exponentiallyand will continue to grow alongside the expectationto utilize all relevant evidence in signal management. Where could AI implementation benefit safety signal management? While much of the data review for signal management has historically been manual, with an agentic AI approach,these processes could be automated to increase efficiency and reduce manual work. When and where it makes Scenario 1: Detection algorithm improvement AI is being applied to develop better algorithms that can uncover potential signals and consolidateinformation from various sources. Bothk-means and random forest algorithms have been explored in signal detection, and the IHI consortium is now exploring algorithms to enable faster, more accuratesignal detection, marking a drastic shift from the disproportionality algorithms — Empirical BayesGeometric Mean (EBGM), Information Component (IC), Proportional Reporting Ratio (PRR), and Reporting Odds Ratio While signal detection has traditionally been performed on a monthly or quarterly basis, AI workflows are driving a shifttoward real time detection that refreshes as new data comes in. Predictive analysis is another possible advantage, thanksto AI’s ability to detect trends across compounds and adverse events that can help forecast a product’s safety profile. Scenario 2: Data contextualization When data volumes are high and coming in a variety of different formats — including company data,literature, labelling documents, the FDA Adverse Event Reporting System (FAERS), the EudraVigilance qualitative signal detection may be less effective. This is due to differences in the way information is reported across AI can be used in signal management to uncover potential trends in specific populations or through related terms whileincluding supporting evidence in its response. It can also be used to reduce false positives and identify connections inmulti-source data. AI will augment the previously manual signal analysis step and ensure that traceability is maintained, Scenario 3: Data summarization AI can be leveraged to conduct data extraction from health authority documentation, competitor labels, and literatur