Infectious disease modelling in the age of AI
Opening keynote
Slides Workshop page BVDOutbreakSize How I am LLM epiaware.org/approaches
Abstract
Infectious disease modelling has a long history, but the last ten years changed what it is asked to do during an outbreak and how fast. Firstly, the outbreaks we worked on. Ebola in West Africa in 2014, with CDC’s projections and real-time forecasts in Sierra Leone. COVID-19, where we produced reproduction number estimates daily for thousands of locations and weekly for the UK government’s advisers. Mpox in 2022, where the COVID-19 models could not be reused. The 2026 outbreak of Ebola disease caused by Bundibugyo virus in the Democratic Republic of the Congo, where again we could not reuse past work and built a joint model that is refit with each situation report. Then how an outbreak gets modelled and what keeps going wrong.
Secondly, the use of AI methods in outbreak modelling. Universal differential equations, physics-informed neural networks, a renewal process as a neural network layer, normalising flows, foundation models and reinforcement learning. They are under-explored and have seen little operational use that I am aware of.
Thirdly, agentic AI and outbreak modelling. Coding agents, working to a written modelling workflow, built the Bundibugyo model: they read the French situation reports, digitised the onset curve, wrote and checked the code, and refit it with every update. What that made possible, what went wrong, and agents that search over models.
The organisers asked me to open the workshop with a keynote of 20 to 30 minutes under the tentative title Infectious Disease Modelling in the Age of AI, with the content left to me. Their suggestion was some history of infectious disease modelling, a few practical examples from my own experience, and an open question on how AI fits into this space going forward. The audience is a mix of infectious disease and AI researchers, some with little or no modelling background. Steven Abrams and Pieter Libin speak straight after, followed by three case studies.
Later the same day
- Science communication with and under AI. The afternoon lecture. 13:40–14:10.
- Where do we go next?. The closing panel. 15:20–16:15.
- Research exchange with the VUB AI group and SIMID. The following day.
Resources
Outbreaks
- BVDOutbreakSize. A live joint model of the 2026 DRC outbreak of Ebola disease caused by Bundibugyo virus, built largely with coding agents. Source.
- The epiforecasts COVID-19 dashboard. Daily \(R_t\) estimates and forecasts, 2020 to 2022.
- EpiNow2. The real-time \(R_t\) tool from the pandemic.
Agents
- How I am LLM. What coding agents do and do not do in a research workflow.
- A workflow for infectious disease modelling. The workflow the agents are asked to follow.
Composable modelling
- What we want from an approach. The design considerations, and the two approaches we are taking.
- EpiAware. The Julia organisation building it.
- JuliaCon 2026. Three talks on composable modelling in Julia, August 2026.