A Partially Observable Markov Decision Process for Optimal Design of Surveillance Policies for Bladder Cancer

Bladder Cancer is the fourth most common cancer in men and eighth in women in the United States. For patients with a history of bladder cancer, the probability of recurrence at one year ranges from 15% to 70%; and the probability of progression to high risk muscle invasive bladder cancer at 5 years ranges from 7% to 40%, depending on the patient's particular risk factors. Cystoscopy is regarded as the gold standard for surveillance of bladder cancer recurrence and progression. However, no consensus exists about the best frequency of follow-up cystoscopy for patients with a history of low grade Ta disease. In this thesis we use stochastic models to investigate policies for bladder cancer surveillance. First, we formulate a partially observable Markov model. The model includes states defining stages of bladder cancer, the effects of treatment, death from bladder cancer, and all other cause mortality. Simulation is used to compare published recommendations for bladder cancer surveillance policies based on expected quality adjusted life years (QALYs) over the patient's lifetime. We compare the American Urological Association (AUA) guideline, the European Association of Urology (EAU) guideline, and several other policies. Next, we extend our model to a partially observable Markov decision process (POMDP) to determine the optimal surveillance policy. We present a series of computational experiments. Results show that age and comorbidity significantly affect the optimal surveillance policy. We find that younger patients should have more intensive surveillance than older patients and patients having comorbidity should have less intensive surveillance. We perform sensitivity analysis to evaluate the influence of model input parameters. Among them we find that disutility of cystoscopy has a significant influence on the optimal surveillance policy. In general, the lower the disutility of cystoscopy, the more intensively surveillance should be performed. Finally, we extend our initial POMDP model to incorporate a new urine based biomarker test into the surveillance process. We study the incremental benefit of the optimal policy that includes a biomarker test and cystoscopy over the optimal policy based on cystoscopy alone. We also compare the optimal policy to easy-to-implement heuristic policies using a biomarker to direct the frequency of cystoscopies. We find that introduction of a biomarker does not significantly improve the optimal policy based on cystoscopy alone; however, biomarkers may significantly improve the heuristic policies we investigated.