Treatment Optimisation

What if mathematical models could predict the best treatment for each patient?

In Simple Terms

Treating a long-running disease such as multiple myeloma is a sequence of decisions made over years: when to test, when to change therapy, when to wait. Each decision is taken with incomplete information about what the disease is doing between visits. Mathematics offers a way to reason about this carefully, weighing the benefit of acting now against the cost and uncertainty of acting later. We build models that learn from a patient's own trajectory to suggest when and how to intervene, aiming to personalise follow-up rather than apply one fixed schedule to everyone.

The Science

This programme develops stochastic control and reinforcement-learning methods for medical follow-up optimisation, modelling disease progression as partially observed piecewise-deterministic Markov processes and controlled hidden semi-Markov models. We estimate relapse-time distributions from longitudinal biomarker trajectories, then use Monte-Carlo planning and Bayesian adaptive deep reinforcement learning to derive intervention policies under model uncertainty. Multiple myeloma is our primary clinical application, connecting the molecular RNA and epitranscriptome signatures characterised elsewhere in the group to quantitative decision support for personalised treatment.