Optimising outcomes for resectable pancreatic cancer by learning lessons from military strategy and the stockmarket: creation of a prognostic Bayesian belief network that makes personalised pre and post-operative predictions of outcome across competing treatment strategies

Aims: Bayesian Belief Networks (BBN) have been successfully used to perform risk assessment and predictions under uncertainty within complex adaptive systems to optimise outcomes in high-risk industries including: planning military attacks, assessing corporate bankruptcy risk and Stockmarket predictions. The aim of this study was to create a prognostic BBN that can make personalised pre and post-operative predictions of 3year or more survival time post resection of PDAC. Methods: BBN consists of variables, known as nodes, with arcs depicting causal relationships from parent to child nodes. Each node has a defined and exclusive set of states and the dependencies between nodes are quantified through a set of conditional probability tables with that of a child node defined by the state of its parent nodes. Parent nodes were identified following comprehensive search of survival analysis studies contained within the PubMed database (n=48691). Each identified variable underwent a two-stage weighting process. The normalised weight for each variable became the weighted mean of the TNormal distribution for the corresponding parent node. Parent nodes were weighted based on their overall ranking of normalised weights to calculate the child nodes. The model was validated against a prospectively maintained patient database. Results: Area Under the Curve (AUC) was 0.94 (P-value 0.002; 95% CI 0.859-1.000) for accuracy of pre-operative predictions. Performance for prognostic updating based on post-operatively available information was 0.97 (P-value 0.000; 95% CI 0.908-1.000). Conclusions: The future application of this model will be in encompassing emerging genomic, pathology and clinical data to make increasingly precise, personalised predictions.