EU Joint Research Centre · Health

AI Can Enhance Disease Surveillance in Europe, But Human Oversight is Crucial

Generative AI, particularly Large Language Models (LLMs), shows potential for improving disease surveillance in Europe by extracting and organizing scattered public health information, according to exploratory research. The findings, published in a JRC report, emphasize the need for robust governance, including validation, bias mitigation, and interoperable data standards.

The JRC, in collaboration with European Commission partners and external organizations, tested AI systems designed to identify early epidemiological signals faster than traditional methods. Prototypes were developed to analyze thousands of outbreak reports and news articles from the Epidemic Intelligence from Open Sources (EIOS) system, helping to construct an epidemiology knowledge graph. By applying Retrieval Augmented Generation (RAG) to WHO Disease Outbreak News, the AI improved the generation of detailed health threat narratives, enhancing situational awareness and preparedness.

These AI models demonstrated a rapid ability to synthesize large volumes of unstructured epidemiological data, potentially reducing the time between outbreak detection and public health response from days to much shorter periods. However, the research highlights significant considerations. The AI prototypes have not yet been validated in operational public health settings, and their real-world effectiveness in speeding up outbreak detection and response remains to be seen.

Accuracy is a key concern, as LLMs can produce plausible but incorrect information, and their reasoning processes can be difficult to scrutinize. The report also points to risks of bias, privacy issues, and potential impacts on individual rights due to expanded digital surveillance. These concerns are particularly critical in public health, where inaccurate information could lead to flawed decisions regarding emerging disease threats.

The report's authors stress that human expertise and validation are indispensable, with AI intended to supplement, not replace, epidemiologists and other specialists. The report advocates for governance frameworks tailored to AI in public health, including validation and bias mitigation mechanisms. Establishing common data formats and interoperability among EU Member States is identified as crucial for broader adoption.

Instead of immediate EU-wide implementation of AI-based epidemic intelligence tools, JRC scientists propose a phased approach. National or regional public health agencies could pilot AI systems to evaluate their performance in real surveillance environments. Successful pilots could then inform wider scaling-up, supported by common standards, governance, and training.

AI-samenvatting op basis van de bron.