Generative AI, especially Large Language Models (LLMs), could be used to extract and organise otherwise scattered and sometimes overlooked information on disease outbreak, according to our exploratory research. This information could help health authorities to better monitor and act upon disease outbreaks and other public health threats.
The findings, published in the JRC report Recent Advancements on Artificial Intelligence for Public Health Threats, also underline the importance of governance — covering validation, bias mitigation, and interoperable data standards.
From scattered health data to clear pictures
In cooperation with partners within the European Commission, as well as external organisations, the JRC tested AI systems that could identify early epidemiological signals more quickly than what traditional surveillance systems allow.
Researchers introduced AI‑based prototypes that can sift through thousands of outbreak reports, news reports and other unstructured information from the Epidemic Intelligence from Open Sources (EIOS) system.
These AI prototypes were specifically adapted to epidemiological information extraction and helped researchers to build an epidemiology knowledge graph which linked information from different sources. The researchers applied Retrieval Augmented Generation (RAG) over WHO Disease Outbreak News to improve the generation of detailed health threat narratives, thereby enhancing situational awareness and preparedness.
The results demonstrated the models’ ability to quickly synthesise large amounts of unstructured epidemiological information. This can reduce the time from outbreak detection to public health response, as scientists would need days to process the same data.
But these findings come with important considerations.
The prototypes have not yet been validated in operational public health setting, and it remains to be seen if they make outbreak detection and/or response times faster in real-world conditions.
Accuracy is another concern. Large language models can generate plausible-sounding information that is wrong, while their reasoning can be difficult to inspect. The report also highlights risks of bias, privacy concerns and the possibility that expanded digital surveillance could affect individual rights.
These concerns are particularly important in public health, where an incorrect interpretation of information could influence decisions about an emerging disease threat.
How to use AI in epidemiological surveillance?
The authors of the report explain that human expertise and validation remain essential and that AI should complement and not replace epidemiologists and other specialists.
The report calls for governance arrangements specifically adapted to AI in public health, including mechanisms for validation and bias mitigation. Common data formats and interoperability between EU Member States are identified as important conditions for wider adoption.
Rather than an immediate EU‑wide adoption of AI‑based epidemic‑intelligence tools, JRC scientists suggest a phased approach. AI systems could first be piloted by national or regional public health agencies, allowing authorities to assess how the technology performs in real surveillance environments. Successful pilots could then inform and scale-up to wider implementation supported by common standards, governance and training.
Background
This research is part of the EU’s wider effort to strengthen preparedness and response to cross-border health threats through digital technologies.
The findings are based on exploratory and prototype-stage work conducted in collaboration with public health stakeholders and institutions including the European Centre for Disease Prevention and Control (ECDC), the World Health Organization (WHO), as well as the European Commission’s Directorates-General for Health and Food Safety, and the Health Emergency Preparedness and Response Authority, among others.
Related content
Recent Advancements on Artificial Intelligence for Public Health Threats
Epidemic Intelligence from Open Sources (EIOS)
- Publication date
- 28 August 2026
- Author
- Joint Research Centre
- JRC portfolios 2025-27
