Individualised dosing via latent-state estimation and reinforcement learning

Abstract

Personalising medical treatment is a crucial step towards improving therapeutic outcomes across a wide range of diseases. However, there remains a need for methods that can continuously adapt interventions to individual patients as new data become available. This is particularly challenging because the latent physiological factors that cause patients to respond differently to treatment are often only partially understood.

In this talk, we present a framework that combines data assimilation with reinforcement learning to support model-informed precision dosing and sequential treatment optimisation. Data assimilation integrates incoming patient data with a mechanistic model to estimate the patient’s evolving physiological state, while reinforcement learning uses these estimates to adapt treatment decisions over time.

We also investigate an experimental-design probem concerning how frequently and for how long a patient should be observed. The aim is to determine an observation strategy that enables data assimilation to estimate the patient’s state accurately while balancing the value and cost of additional measurements.