Telfer Research Seminar Series - Brian Denton

Risk models are frequently used to inform physicians about making patient recommendations. However, the lack of reproducibility of results from multiple medical studies leads to model ambiguity. When physicians select a model based on a single study, they risk making a poor choice that could lead to suboptimal recommendations and regret. In this talk, we propose a multi-model partially observable Markov decision process (MPOMDP) to address model ambiguity under partial observability caused by latent disease factors. The proposed MPOMDP model aims to learn the distribution of the true model from system outputs over time and to find an optimal policy that maximizes the expected future rewards across models. In this talk, we discuss properties of the MPOMDP, providing some insight into the model and motivating solution methods. Methods suitable for real-world disease surveillance models are described and applied to a case study of active surveillance for prostate cancer using four of the most well-known clinical studies from Canada, Europe, and the United States. We present results illustrating the ”gap” between individual model recommendations and further demonstrate the positive impact of considering model ambiguity in the prostate cancer setting.