Expert Judgment Supporting a Bayesian Network to Model the Survival of Pancreatic Cancer Patients

Purpose: Pancreatic cancer is known for its poor prognosis. The most effective treatment combines surgery with peri-operative chemotherapy. Current prognostic tools are designed to predict patient outcomes and inform treatment decisions based on collected data. Bayesian networks (BNs) can integrate objective data with subjective clinical insights, such as expert opinions, or they can be independently based on either element. This pilot study is one of the first efforts to incorporate expert opinions into a prognostic model using a Bayesian framework. Methods: A clinical hybrid BN was selected to model the long-term overall survival of pancreatic cancer patients. The SHELF expert judgment method was employed to enhance the BN's effectiveness. This approach involved a two-phase protocol: an initial single-center pilot phase followed by a definitive international phase. Results: Experts generally agreed on the distribution shape among the 12 clinically relevant predictive variables identified for the BN. However, discrepancies were noted in the tumor size, age, and ASA score nodes. With regard to expert concordance for each node, tumor size, and ASA score exhibited absolute concordance, indicating a strong consensus among experts. Ca19.9 values and resectability status showed high concordance, reflecting a solid agreement among the experts. The remaining nodes showed acceptable concordance. Conclusions: This project introduces a novel clinical hybrid Bayesian network (BN) that incorporates expert elicitation and clinical variables present at diagnosis to model the survival of pancreatic cancer patients. This model aims to provide research-based evidence for more reliable prognosis predictions and improved decision-making, addressing the limitations of existing survival prediction models. A validation process will be essential to evaluate the model's performance and clinical applicability.