Reinforcement learning for precision medicine: Integrating medical knowledge into decision-making algorithms
Sophia Yazzourh (McGill University)
Wednesday 7th January 12:00-13:00
Maths 311B
Abstract
Precision medicine aims to tailor treatments to individual patient characteristics. This paradigm relies on the framework of Dynamic Treatment Regimes (DTRs), which seek to determine an optimal decision rule at each stage of intervention. The construction of such rules draws on a variety of methodological approaches, including causal inference, Bayesian modeling, and machine learning.
This presentation focuses on reinforcement learning (RL) in the context of DTRs. In particular, we explain why, among the available algorithms, Q-learning is especially well suited to the challenges posed by medical data. However, the practical deployment of these approaches in clinical settings remains limited, as both clinicians and patients often perceive learning-based methods as “black boxes” whose decision-making processes are difficult to interpret.
We therefore explore a key direction for improvement: the integration of medical knowledge into RL models. To this end, we introduce a probabilistic approach to reward construction based on expert medical preferences. Applied to case studies in diabetes and cancer, this framework enables the generation of data-driven rewards while avoiding biases induced by manual reward specification, thereby ensuring closer alignment with clinical objectives.
Add to your calendar
Download event information as iCalendar file (only this event)