How is AI developing in pharmacovigilance?

Despite all the hype, AI and ML are not new technologies: they are not even new in pharmacovigilance (PV)!^1,2 There have also been several recent special issues dedicated to the use of AI in PV and medicine safety^3,4 and several literature reviews.^5–8 Why then the hype now and another collection on AI in PV? The combination of the rapid advances in the technology itself, the wave of increased interest in awareness of AI, and the sudden ease of access for many to populist AI tools has led to huge advances in interest, capability, and scientific research into AI in PV: some of which is therefore presented in this special issue. The dramatic increase in focus can be traced to the release of ChatGPTv3.5 ⁹ and the public access to the tool toward the end of 2022 and in particular the GenAI advances that seem to have accelerated since then. There is little doubt that AI is leading to great advances in our everyday lives, but with it, we are seeing both enormous publicity and therefore the associated hype/concern about use of these technologies. Other ML research of course continues to be used see, for example, Chakraborty and Tiwari ¹⁰ and Cherkas et al., ¹¹ but over the last few years, we have seen a sudden influx in scientific publications discussing experimentation with generative AI and large language models (LLMs) in pharmacovigilance for a range of problems.^12–15 We must now begin as a field as an absolute minimum to consider implications of this “brave new world” ⁹ but also not be constrained to historical requirements or approaches.

There is much opportunity to advance PV. ¹⁶ At its core, technological advances should allow us to do things we couldn’t before, or earlier/more effectively/efficiently. It is already shown, we would argue, that ML and AI is adding value across the PV lifecycle, to address new problems or address them in entirely new ways, for example, duplicate detection^2,17; to do well established tasks better or more effectively ¹⁸ and even to reinvent aspects of PV process entirely to enable maximized and appropriate use of AI.^19,20

Given, then it felt important and arguably pressing to reconsider the current state of the use of AI in pharmacovigilance today and where the field is heading: we were, in particular, keen to explore the

• 1. Impact and challenges and specific opportunities the newer branches of AI and ML, in particular GenAI, would afford PV. There being very limited literature at the time we embarked on this special collection.^12,13,15

• 2. Extent to which experimentation with AI was leading to production usage, which has been limited to specific examples previously. ²

• 3. End to end impact of AI across the PV lifecycle, much of the literature previously while promising on specific tasks has focused on specific use cases, for example, Bergman et al. ¹⁸ and Cherkas et al., ¹¹ without attempting to assess the wider PV lifecycle impact, as is beginning to advance in other fields. ²¹

• This would include not just multiple AI agents connecting and interacting, an area of much current research but also hybrid systems with connectivity to rules-based tools as part of the overall system, where rules-based solutions hold particular usefulness (e.g., transparency, explainability, ability for predictable error prevention).

• 4. Use of other data streams in addition to individual case safety reports (ICSRs) in AI ⁸ ; particularly given the need to enrich a data stream that, while a keystone of PV, is often poor in quality and can be enriched substantially with existing technology.

Niazi and Mariam ²² review the wider AI landscape, describe exploration of AI across the entire R&D landscape including PV and argue it is already having clear impact across the R&D landscape with even great future potential. They do also emphasize the importance of the combination—and therefore interaction—between human and computer.

In PV, one area where there has been much focus is on automating case intake and processing. Automation ²³ is sorely needed given the complexity of processes to handle the large volumes of data and heterogeneity not just in data but also in expectations around data management. While this key area of interest has led to much rules-based automation, it is clear that GenAI holds the promise of further advances as Roemming et al. clearly articulate in their perspective review ²⁴ while discussing the many decisions around considering GenAI implementation for case intake and processing and make clear it is far from facile. Similarly, to what extent can LLMs be used to detect drug–drug interactions (DDIs)? Sicard et al. ²⁵ study this with a range of LLMs and conclude at this stage that LLMs can detect DDIs leading to pharmacovigilance cases but cannot reliably exclude DDIs in cases without interactions emphasizing the crucial role pharmacologists play in assessing whether a DDI is implicated in an ADR.

Nagar et al. ²⁶ review the current use of AI and consider the regulatory use as well as indications of routine usage, emphasizing for trusted and broad use that a holistic approach involving regulatory alignment, that a transparency, and ongoing validation is crucial to ensure AI systems in PV that not only perform as expected but also can adapt as needed as the landscape of drug safety evolves.

As Ramcharran et al. ²⁷ argue having guardrails in place is important for a field like pharmacovigilance for wider use of AI where one wants to be sure that very rare unanticipated outcomes are important to effectively identify and appropriately addressed, and there is robust and appropriate governance to ensure the AI system works as needed, including quality assurance and controls see, for example, Cain et al. ²⁸

Consideration is needed to manage the planned use and expected outcomes of systems to identify and handle any anomalies. ²⁹ In a field like PV, one must pay attention to the field’s particular needs, for example, ensuring a system that works well for both more common well established safety outcomes, and also rare, potentially unknown outcomes.^30,31 A perspective on how to ensure appropriate governance is in place for PV is articulated by Glaser and Littlebury. ²⁹

Richer data through linkage to strengthen insights of relevance to PV is important. Crown et al. ³² show this potential capability. They point out that social determinants of health (SDOH) are not widely captured in electronic medical records (EMRs), insurance claims databases, and other forms of real-world data (RWD). The article explores how SDOH data might be linked by using ML to generate SDOH clusters, which can then potentially be linked to RWD—this richer data could enable enhanced causal inference in drug safety analyses.

Another key data source with a long-established role in safety is healthcare databases. ³³ In order to analyze multiple healthcare databases at once and rapidly, often each database is converted for analysis to a standardized format a “common data model” for standardized analytics usage. This conversion can sometimes be a slow and difficult process ³⁴ and can potentially provide a barrier to access and therefore surveillance. Not to mention converting and maintaining such a standardized data model can be expensive, for example, the IHI project EHDEN that ran from 2018 to 2024 looked to do build a federated data network of allowing access to the data of 100 million EU citizens standardized to a common data model and cost more than 30 million Euros. ³⁵ If AI has the potential to remove the need for such common data models and enable more rapid and richer access to data for pharmacovigilance then as Painter et al. ³⁶ argue, it should be explored.

Following the theme of linking or incorporating data that one rarely have access to, Crown et al. ³⁷ present interesting simulation work on behavioral data. Specifically, an agent-based simulation model to estimate the impact of a peer navigation program in Tanzania on antiretroviral therapy participation and adherence. Clearly, an effective ability to incorporate behavioral aspects of drug treatment patterns through agent-based simulation which we wouldn’t be able to incorporate otherwise could help in better understanding of vaccine and drug treatment benefit-risk.

To what extent do we deliver on our aims? Well, this special issue is full of examples of new approaches to tackling challenges in PV, including exploring the use of GenAI and the particular challenges these newer types of AI bring. Although some potential important applications, for example, safety outcome prediction and personalized pharmacovigilance do not feature herein, suggesting there is much more we can expect in the future. We clearly also see plenty of applications of PV using data outside of ICSRs, adding support to the notion that PV is becoming increasingly multimodal—this being accelerated by emerging AI capabilities and needs for more/better richer data. We see much discussion of how to implement systems around the PV as part of overall PV systems, but limited evidence of routine usage of PV. Conversely, there is opportunity for PV data to be used outside of our regulated domains to enhance patient safety and the information available to healthcare providers. There appears to be opportunity for convergence of data sources to offer a more personalized approach to benefit-risk than current generic population-based approaches.

The field seems still to focus at least in research terms on using AI for specific tasks, rather than evidence of impact across large parts of PV systems, perhaps the advent of agentic AI and the ability of AI algorithms to interact with one another and propose actions will lead to a change on this moving forward. The wider use of AI will sharpen the focus on certain challenges with the use of AI, for example, will automation of first draft text with GenAI for PV applications actually increase burden on human reviewers either requiring more time from humans (as checking is obfuscated) or as expectations on human for rapid human review increase? Or indeed, will it lead to overreliance on automated content “automation complacency,” at the potential expense of quality? Perhaps synergistic working with GenAI models to test human thinking and identify concordant or conflicting data points to their conclusions might prove to be a more effective use than a drafting replacement approach? Clearly, evaluation and monitoring post deployment will need to focus on time taken to generate final products and the quality of this final product rather considering the time and quality of a first draft as a useful metric for analysis. ³⁸

Like any scientific discipline, there is much in pharmacovigilance that could be improved we need to continue to critically question and study the approaches and implications of their implementation used across the field of medicine safety^16,19,20,39 as we strive to maximize the opportunities that advances in AI can afford. For us to see widespread value of AI and help patient safety: to ensure effective and trusted use, strong international collaboration, for example, through CIOMS is essential to enabling such a future vision. ⁴⁰ The future is exciting, but we must together ensure that AI unlocks the advances we want, and patients expect, in understanding and preventing harm from medicines and vaccines.