SIAM Chapter Seminar: Data assimilation and reinforcement learning in model-informed precision dosing

23.02.2022, 4:00 pm  –  Hybrid (Zoom and room 2.22)

Franziska Thoma and Katherine Briceno Guerrero

Abstract: Model-informed precision dosing (MIPD) employs a combination of mathematical modelling and individual patient measurements to calculate optimal drug dose and maximize treatment success in individualised therapy. Bayesian data assimilation and reinforcement learning can support informed decision making by providing uncertainty quantification and an understanding of underlying patient factors that influence dose decisions. In this talk, we will give you a glimpse into two approaches that aim to improve MIPD:
- We propose a neural network for smooth parameterization of the long-term return for neighboring states and actions that will efficiently reproduce the optimal policy on the reinforcement learning approach. Particularly when dealing with an increase of the complexity on the model by means of increasing the MIPD parameters (e.g. dosing time or rates of recovery).
- We develop and evaluate a fully Bayesian hierarchical approach on a toy model for continuous learning across patients when the models are employed in a real-world scenario aiming to improve the estimates of population parameters from a study population.

The log-in details are announced via mailing lists "mathe-phd" and "SIAM-chapter-list", or ask Undine or Florian. See also here.