# What AI Could Change in Pharmacoepidemiology Research - Medscape

*Источник: Medscape*
*Дата: 2026-10-01*
*Язык: en*

**Кратко:** MILAN — At the 42nd annual meeting of the International Society for Pharmacoepidemiology (ISPE), held in Milan from August 29 to September 2, 2026, AI was among the major topics discussed. John Diaz-Decaro, who leads ISPE’s Digital Technology and Artificial Intelligence Special Interest Group, spoke with Univadis Italy, part of the Medscape Professional Network, about some aspects of the increasingly frequent and highly practical interaction between pharmacoepidemiology and AI.

MILAN — At the 42nd annual meeting of the International Society for Pharmacoepidemiology (ISPE), held in Milan from August 29 to September 2, 2026, AI was among the major topics discussed.
John Diaz-Decaro, who leads ISPE’s Digital Technology and Artificial Intelligence Special Interest Group, spoke with Univadis Italy, part of the Medscape Professional Network, about some aspects of the increasingly frequent and highly practical interaction between pharmacoepidemiology and AI.
AI is taking a central role in pharmacoepidemiology. In which areas is it already making a difference today, and in which areas is its potential overrated?
AI represents a huge opportunity for the fields of pharmacoepidemiology and real-world evidence. Our field, like many others, is going through a period of transformation. What once seemed like science fiction is now reality. We can complete well-structured literature reviews in a few hours instead of days or weeks or support clinicians in identifying potential patient populations before the end of a meeting. However, we are still in the early stages. I think it is premature to ask where AI is already making a significant difference. The question is valid, but it is still too early.
AI is becoming a focal point of our research; nevertheless, our field is only now beginning to explore potential use cases of these tools across pharmacoepidemiology activities such as safety surveillance and protocol development. We recognize the enormous potential of these technologies, but reproducibility is often a challenge. I would say that the key shift we are seeing today is this: We can no longer be only pharmacoepidemiologists; we must be pharmacoepidemiologists trained in the use of AI.
It is inaccurate, even irresponsible, to think research teams can simply press a button and generate a protocol, a report, or a finished result. These systems require an iterative process. Future adoption and implementation must include an “epi-in-the-loop.” Any AI-based workflow that is developed must include expert review before the workflow proceeds. AI itself needs additional contextual knowledge that only a qualified epidemiologist can provide.
Which parts of the pharmacoepidemiologic research process do you think are most likely to be transformed by AI in the coming years?
In short, all of them. I do not think there is any aspect of our research that AI cannot influence or improve in some way. At ISPE’s 2026 annual meeting, my colleagues presented innovative research showing how AI can be used across all aspects of the research process, and there are many other projects underway that have not yet been presented. I have had the privilege of leading ISPE’s Digital Technology and Artificial Intelligence Special Interest Group for the past 12 months, and, to better understand all of this work, we are conducting a living scoping review to assess how AI is being used in our field.
Rather than replacing pharmacoepidemiologists, could AI radically change the nature of their work? What skills will a pharmacoepidemiologist need in 5 or 10 years that might not be essential today?
I am convinced AI will radically change our profession and that we must proceed with extreme caution. It is easy to delegate phases of critical thinking to these large language models (LLMs), such as drafting manuscripts or funding proposals or developing study designs. These have always been tasks requiring careful reflection, but what we are seeing today is that these phases of critical thought have not been eliminated entirely — they have simply shifted downstream.
For example, suppose an LLM drafts an entire protocol for us. We entrust all that reflective work to the model. The LLM identifies objectives, inclusion and exclusion criteria, and all other components of the protocol up to the statistical analysis plan. We receive that result, but before sharing it with colleagues or regulators, we still must review it critically. That is the change. Now we will reflect critically on what the model has done for us instead of doing the critical thinking ourselves. The crucial point here is the architecture we use to build the AI system.
Soon, pharmacoepidemiologists will need to become “builders.” We will need to be able to create the architecture and ecosystem required for AI to perform specific tasks. It may be a slow process, but we must take on this builder role. We are the subject matter experts, and no one is better qualified to develop an agent capable of performing a pharmacoepidemiology task than someone who has done that work for 5, 10, 15, or more years.
Some studies presented at the meeting highlighted current weaknesses of AI, from inaccurate citations to errors in synthesizing research findings. In a field where scientific evidence can ultimately affect regulatory decisions and patient safety, what level of validation and human supervision should we require before relying on AI-generated results?
To be clear, much of the research presented at the conference relied on commercial models, such as ChatGPT or one of the Claude models. These are off-the-shelf systems. The authors added extra context or even used excellent prompts, but none of that is deterministic or fully reproducible. Given the regulatory nature of our work and the possibility that the evidence we generate could affect policy or patient care, we must certainly think carefully about validation and human oversight.
Looking further ahead, we’re hearing not only about generative AI but also about “agentic” AI systems capable of performing increasingly complex research tasks. Can we imagine a future in which AI designs and conducts substantial parts of a pharmacoepidemiologic study autonomously? If so, where should we draw the line between what can be delegated to a machine and what must remain a human scientific responsibility?
First, let us clarify that agentic AI represents an evolution of generative AI systems as we know them. Instead of just inputting a prompt and receiving an output, agentic systems can perform an action on behalf of the user. That capacity for action allows us to use such systems for complex research activities. Agentic systems are already being used to perform pharmacoepidemiology tasks, and their use will undoubtedly increase. But ultimately it should be the pharmacoepidemiologist who determines how these systems are used to design and conduct our studies.
What strikes me is that the question we were asking last year was “should we use AI in our work?” and now it has become “how can we use AI responsibly in our work?” The boundary to draw concerns validation and benchmarking. We are talking about how to validate individual agents to ensure the right tools are called during execution, in the correct order, and that they provide a similar output each time they are used. I deliberately said, “a similar output,” not “the same output.”
It would be a mistake to view agents as deterministic because that is not how they behave. The issue of validation is broad and complex because we are not just validating individual agents but entire workflows and systems. The most important priority right now is the development of validation frameworks for our field in relation to the output of these agents, asking what is fit for use in the specific context of pharmacoepidemiology.
This article was translated from Univadis Italy.

[Оригинал](https://www.medscape.com/viewarticle/what-ai-could-change-pharmacoepidemiology-research-2026a1001085)