Artificial intelligence is already being used in many areas of pharmacovigilance. It offers both opportunities and risks, as AI-generated information is not always accurate. Human review and assessment therefore remain an important part of its use.
Where can AI be used in pharmacovigilance?
In adverse event case processing, AI can read source documents, support case classification and identify potential duplicate cases. It can also be used to draft case narratives and support MedDRA coding.
In pharmacovigilance literature screening, AI can help review information and identify publications that require further assessment. In signal management, AI can be used to analyse data and identify potential safety signals.
AI can also support periodic safety reporting, for example by retrieving information, generating listings and drafting content for PSUR/PBRER reports. In addition, AI can help check consistency across case series and support quality and compliance activities.
AI can be convincingly wrong
AI-generated errors may not always be easy to detect because the content can appear completely credible.
For example, AI may add laboratory values or details about a patient’s medical history to a case narrative even though they are not included in the original source material. It may also suggest MedDRA coding that appears appropriate but is clinically incorrect.
In pharmacovigilance literature screening, AI may misinterpret the clinical context and classify a relevant article as irrelevant. In signal detection, gaps in training data may lead to safety signals being missed. This may affect, for example, rare events or certain patient populations.
Errors can also occur in data selection. An AI-generated listing may appear correct even if the wrong selection criteria have been applied. AI may also produce different case narratives or different clinical conclusions when processing the same case more than once.
For this reason, AI-generated information must be checked against the original source. AI itself should not be used as a source of information.
Human involvement remains essential
The use of AI in pharmacovigilance follows the human-in-the-loop principle. In practice, this means that a person reviews and assesses the work produced by AI or makes the final decision.
In adverse event case processing, a person confirms the classification and how the case should be handled. For case narratives and MedDRA coding, a medical reviewer checks that the content is correct. In signal management, an expert assesses the clinical significance of the findings and recommends any necessary follow-up actions. In safety reporting, AI-generated content is checked against the original data.
Clear processes are also needed at the organisational level. Organisations need to know which AI tools are being used, who approves their use, and how they are validated and monitored. It is also important to define where AI can be used in different pharmacovigilance activities and where its role should remain limited to supporting the expert.
AI literacy and regulation are increasingly important
AI regulation is developing rapidly. Article 4 of the EU AI Act requires providers and deployers of AI systems to take measures to ensure a sufficient level of AI literacy among their staff.
EMA and FDA have also addressed the use of AI throughout the medicinal product lifecycle. EMA’s guidance highlights areas such as explainability, traceability and inspection readiness. FDA and EMA have also published joint principles for good AI practice covering the entire medicinal product lifecycle.
In pharmacovigilance, AI users need to recognise the potential errors and limitations of these tools. AI can support the work of experts, but responsibility remains with people and, ultimately, with the marketing authorisation holder.
