A partially observable markov decision process approach for clinical decision support in cancer treatment: implementation for colon cancer

The cancer treatment process involves uncertainty by its nature. Since the disease evolves continuously and diagnostic tests used to detect the level of the disease are not totally accurate, the actual state of the disease remains unknown. Therefore, physicians should make treatment decisions in a stochastic environment. This study aims to develop a mathematical model of the history of the colorectal cancer treatment process by using partially observable Markov decision process. To understand the impact of the partially observable environment on modeling the history of the disease, a comparative analysis of the outputs of the partially observable Markov decision process model, in which the patient's actual health status is estimated from the blood carcinoembryonic antigen level change and computed tomography results as observational states, with a basic Markov decision process model that assumes the patient's actual health status is fully known. has been made. The output of the proposed model has been compared to 5-year survival outcomes that come from Surveillance, Epidemiology, and End Results database. A series of hypothetical scenarios have been presented to understand the effectiveness of the model and some limitations encountered in the modeling process have been mentioned along with suggestions for future studies will be made.