Assessing colorectal cancer screening and surveillance strategies with colonoscopy: A Partially Observable Markov Decision Processes (POMDP) approach

Maoying Zhu, Purdue University


Background & Aims: Colorectal cancer (CRC) is the second leading cause of death from cancer in the United States, resulting in more than 49,960 deaths in 2008. Early detection of CRC through screening could decrease the mortality risk. Colonoscopy, a widely used CRC screening method in practice, can detect and remove polps at the precancerous stage so as to slow down or even prevent possible cancer development. Although it is shown in practice that screening and surveillance for colorectal cancer is cost-effective compared with no screening, it is not clear from previous studies whether the cost-effectiveness can be further improved by carefully designing screening age, surveillance frequency and other factors that specify screening and surveillance strategies. Methods: We modeled the CRC natural history and patient's decision with Partially Observable Markov Decision Processes (POMDP). The states of this Markov Decision model are probabilistic descriptions of a patient's health condition depicted by either the number and size of adenomas or the cancer stage. An action that a patient can take at the beginning of each year is whether or not to receive colonoscopy in the following year. The current screening and surveillance policy consists of a start age and an interval until next colonoscopy given the previous colonoscopy result. We used POMDP model to accommodate the uncertainty due to partial observability, because a decision maker doesn't know the patient's initial condition and could only obtain imperfect information on underlying states through colonoscopy which is not completely sensitive. The POMDP model is based on the combination of a biologically based neoplasia development natural history model and CRC development model from Vanderbilt-NC State Population Simulation Model (VNCS). The large number of precancerous stages enables us to evaluate colonoscopy screening strategies with much greater detail. The cost-effectiveness of a screening policy is evaluated by incremental cost effectiveness ratio (ICER). Results: For alternative groups of screening policies, we identified the optimal start age for patients at average risk. For patients who receive one-shot colonoscopy, the optimal start age is 65. We identified that the optimal start age is 61 for routine screening and screening with surveillance suggested by the current guidelines. The comparison between routine colonoscopy and colonoscopy based on previous screening results proved the necessity of surveillance. We also discussed deviation from the current policy, including adding an end age, varying frequencies of colonoscopy surveillance and changing the screening interval for some observations. The ICER could be further improved with compromise of QALYs. Sensitivity analysis on costs of screening and treatment found that the optimal start age is not very sensitive to cost fluctuation.




Kong, Purdue University.

Subject Area

Biomedical engineering|Public health

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server