Developing an interactive and responsive, large-language model (LLM) -based Public Health Knowledge System

Background:

Public Health information systems are often static knowledge systems, which suffer from a lack of interactivity and responsiveness (e.g. descriptive written reports, websites which merely present data, ...). An interactive and responsive query system that can be used to infer, combine, and extract information about public health, e.g. current information on diseases in Germany (influenza season, spontaneous outbreaks, etc.), disease interventions (vaccinations, etc.), and other public health information (pollen season, heat, policies, etc.) would be very valuable to public health and medical professionals alike. Such a system would make public health information significantly more actionable, e.g., by providing users who ask questions like “What is the current status of the influenza season this winter?” with a dedicated written report based on current disease monitoring, historical data, and up-to-date health recommendations.

Aim/s:

Aim of this project is to design and implement a prototype of such an interactive, responsive Public Health Knowledge System (HEALTHY) based on the currently available text-based knowledge/training data at RKI. HEALTHY will take the shape of a centralized knowledge hub that integrates RKI’s decade-spanning Public Health reports (Epidemiological Bulletins, etc.), emerging topics, and external data sources, which are already in use at ZKI-PH (climate data, image data, etc.), and can be further extended to also account for multimodality (e.g., images and quantitative data from public health research).

AI methods:

On the methodical level, this project requires the usage (training and/or fine-tuning) of a large language model (LLM) that is capable to produce text- and image-based replies to given questions that are based on RKI’s research and evidence gathering. Different open-source LLMs are readily available and can be implemented and trained for this task. Training data is also available in the form of epidemiological health records from the RKI’s Epidemiological Bulletin, which can be further combined with data products currently in use at ZKI-PH.

Keywords:

Artificial Intelligence, Natural Language Processing, Large Language Models, Public Health

to the top