Reply to “A Comment about AI Methods that May Help Advance Pharmacovigilance”

We appreciate the perspective provided by Anderson et al. on our article and agree that the integration of artificial intelligence (AI) methods in pharmacovigilance workflows represent promising avenues for advancing safety assessment of medicines.1 As highlighted by the authors, AI is being actively explored across multiple steps of the pharmacovigilance workflow, and international initiatives such as those from CIOMS reflect this growing interest.2 Yet, several of the applications proposed by authors also illustrate both the promise and the current challenges of AI. For example, automated causality assessment at the individual case level is an important goal, but recent studies suggest that even advanced models face substantial difficulties when confronted with sparse and inconsistently structured data.3 More broadly, while AI has been explored in pharmacovigilance for decades, including in combination with disproportionality approaches, its incremental benefit in real-life settings has, to date, remained modest. We do anticipate this will change as AI advances, coupled with trustworthy frameworks and best practice implemented.2, 4 Moreover, we would like to clarify that the primary aim of our article was to address the current and increasing misuse of adverse event reporting data through disproportionality analysis.1 In this context, we believe it is important to emphasize that methodological advances alone, whether based on traditional statistics/workflows or AI, do not ensure that all the limitations are solved and it is hard to anticipate that human oversight will not remain vital for the foreseeable future. Issues such as nonrandom reporting, missing or inconsistently recorded clinical information, delays in reporting, overinterpretation of disproportionality measures, automation bias, and the difficulty in establishing a causal link remain major obstacles to AI large-scale implementation, especially in complex evidence-to-decision scenarios requiring advanced clinical reasoning.5 Overall, we fully support continued research in this area and agree that such approaches may, in time, enhance signal detection and evidence integration. Yet importantly, the same risks of misuse and overinterpretation that we described for conventional approaches may also apply to AI-generated outputs if they are not critically appraised. As such, AI may both enhance the quality of pharmacovigilance research and, if misused and unsupervised, contribute to distrust its value by generating unreliable results. We therefore view AI as a valuable complement and sometime system component, but not a substitute, for rigorous rules-based methods, domain expertise, and critical interpretation. Ensuring that these tools are implemented responsibly will be essential to realizing their potential in strengthening evidence-based actionable pharmacovigilance. No funding was received for this work. All opinions expressed in this commentary represent the views of the authors and do not represent the views of any listed employer or institution. We have not been paid to write this article and have no conflict of interest related to this study. One of us (AB) is an employee of a pharmaceutical company; his role is, like the rest of us, an expert in the analysis of spontaneous reports, and he is widely cited, having, for example, been first author of one of the earlier publications (Bate et al, Eur J Clin Pharmacol 1998). He reports no conflict of interest with this specific study.