Optimizing active surveillance for prostate cancer using partially observable Markov decision processes

We describe a finite-horizon partially observable Markov decision process (POMDP) approach to optimize decisions about whether and when to perform biopsies for patients on active surveillance for prostate cancer. The objective is to minimize a weighted combination of two criteria, the number of biopsies to conduct over a patient's lifetime and the delay in detecting high-risk cancer that warrants more aggres-sive treatment. Our study also considers the impact of parameter ambiguity caused by variation across models fitted to different clinical studies and variation in the weights attributed to the reward crite-ria according to patients' preferences. We introduce two fast approximation algorithms for the proposed model and describe some properties of the optimal policy, including the existence of a control-limit type policy. The numerical results show that our approximations perform well, and we use them to compare the model-based biopsy policies to published guidelines. Although our focus is on prostate cancer active surveillance, there are lessons to be learned for applications to other chronic diseases.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This article presents a finite-horizon partially observable Markov decision process (POMDP) approach to optimize biopsy decisions for patients under active surveillance for prostate cancer. The objective is to minimize the number of biopsies and the delay in detecting high-risk cancer. The study considers parameter ambiguity and patients' preference variability, and proposes two fast approximation algorithms for the model.