Allowing generative artificial intelligence to analyse large quantities of data from open sources could allow health authorities to detect public health threats more quickly, according to a report published on Tuesday 14 July by the European Commission’s Joint Research Centre.
The research, launched in February 2024, was carried out in collaboration with the European Union’s disease surveillance agency, the European Centre for Disease Prevention and Control (ECDC); the World Health Organization (WHO); the European Commission’s Directorate-General for Health and Food Safety (DG SANTE); and the Directorate-General for Health Emergency Preparedness and Response (DG HERA).
It involved studying the use of artificial intelligence text analysis models to automatically extract epidemiological information from the Epidemic Intelligence from Open Sources (EIOS) system, which compiles information disseminated by ProMED – a platform launched in 1994 that has published over 66,000 messages on infectious diseases – and WHO bulletins on epidemic outbreaks.
The researchers present several use cases, including the creation of an epidemiological knowledge graph, the use of several artificial intelligence models simultaneously to improve the reliability of the results, and using retrieval-augmented generation (RAG) to produce health threat scenarios from epidemiological reports.
Tests performed on a corpus of 171 reports annotated by experts showed that the commercial GPT-4 and GPT-3.5 models, developed by US company OpenAI, provided the best results for information extraction. It should be noted, though, that comparable performance can be achieved by combining three open-access artificial intelligence models.
Furthermore, the study revisits a number of limitations in these uses, such as the lack of transparency of certain models, the risk of “hallucinations”, data-related bias, confidentiality issues, and the spread of medical disinformation. For that reason, they stress that these technologies must remain under human control.
Read the report: https://aeur.eu/f/myu (Original in French by Nithya Paquiry)