Over the next few weeks, I will do a deep dive into AI in pharmacovigilance. I will read the key papers in this field and share what I learn with my network.
My goals are simple.
Become more knowledgeable about how AI is being applied to drug safety.
Share these insights so more people in the PV community can follow important developments.
Learn together with my network and discuss ideas that can help move the field forward.
I started with a review article written by Robert Ball and Gerald Dal Pan from the FDA, “Artificial Intelligence for Pharmacovigilance: Ready for Prime Time?”. The link is in the comment.
The review focuses on how AI has been used so far for processing and evaluating Individual Case Safety Reports. It highlights several promising areas where NLP and machine learning have already shown real value, such as extracting key clinical features, identifying seriousness, detecting duplicates, filtering low value reports, and supporting components of causality assessment. The paper also provides helpful examples from both industry and FDA efforts over the past decade, which gives readers a clear picture of where the field currently stands.
The authors conclude that while AI has made meaningful progress, current methods are not yet ready for full automation of ICSR evaluation. They emphasize that human expertise remains central, and that full automation would require very high performance levels. Their perspective reflects the regulatory need to avoid missing any important safety information, which naturally calls for a high standard. My own view is that AI does not need to reach near perfect F1 scores before becoming useful. If AI can match or exceed typical human performance, or strengthen AI plus human workflows, that may already bring significant value. I see their high bar as a thoughtful regulatory safeguard, while also believing that real world adoption and iterative improvement will be important for long term progress.
The review also points out several broader challenges, including the need for larger annotated datasets, a computable cognitive framework for causality assessment, and improved integration of AI models into production PV workflows. These insights highlight both the complexity of PV and the opportunities for future innovation.
I will continue reading the papers referenced in the review as well as the newer papers that cite it. After that, I will conduct a systematic literature search to broaden my understanding of the space.
If you work in PV, safety science, or AI, I would love to hear your recommendations on important papers or researchers to follow.
Let’s learn and advance this field together.