Data Science Seminar: Could Uncertainty Quantification be Useful During the Next Pandemic?06 Mar 2023, by Sponsored events in
Speaker: Daniel Williamson (Exeter)
29 March 2023 at 12:30
Lecture Theatre 2.41, Fry Building, School of Mathematics, University of Bristol, UK[Lunch will be served in room 2.01 at 12:30 before the seminar]
Though decision makers use the output from computer models to support many of the decisions that shape our everyday lives, never has the general public been so acutely aware of their influence as during the COVID-19 pandemic, when nightly broadcasts from Downing Street presented predicted hospitalisations and deaths to help them to understand current and potential limits on their freedoms. SPI-M, a modelling group, drawn from the UK’s leading researchers in epidemiological modelling, was formed to provide the modelling and uncertainty statements that would support government ministers in taking decisions on behalf of all of us.
Uncertainty Quantification (UQ), the accepted term for the collection of methodologies focussed on quantifying sources of uncertainty induced by studying complex systems with computer simulators, was not used throughout (the clarifier is necessary as of course uncertainties were quantified, just not with UQ). But could UQ have been used?
Our project, “UQ4Covid”, in partnership with one of the SPI-M modellers, aimed to demonstrate that it could, by providing real-time UQ to SPI-M through the local measures, vaccination and variants phase of the pandemic. In this talk I’ll explain how we came to realise that the UQ methods we sold so confidently would not be able to feed into policy support. I will then present a new approach to calibrating infectious disease models in real time that embeds model discrepancy on the state vector within the simulations through data assimilation. I will show how this approach solves the seeding/outbreaks problem and show promising early results.
Co-authors: Trevelyan McKinley, Xiaoyu Xiong, James Salter, Leon Danon, Rob Challen, Ben Youngman, Doug McNeall
Daniel Williamson is Associate Professor of Bayesian Statistics in the department of Mathematics and Statistics at the University of Exeter and a Turing Fellow. He obtained his PhD in statistics with Michael Goldstein at Durham University where he also worked as a postdoc on Uncertainty Quantification for ocean models. He won an EPSRC fellowship and took a faculty position at Exeter in 2013. His work develops Uncertainty Quantification for the challenge of supporting policy makers when numerical models can be used to understand how a complex system might evolve or respond to policy. Usually that work has focussed on climate models, but during the pandemic, he led an EPSRC project, UQ4Covid, that used UQ to support SPI-M-O.
In cooperation with the Jean Golding Institute, University of Bristol
Registration: More information and registration on the Bristol Data Science Seminar Series