Session Item

A Bayesian network (BN) is a probabilistic graphical model that represents a set of variables and their dependencies in a directed acyclic graph (DAG).  Within the graph, each node signifies a variable, and the direction of the arrow link between nodes represents the direction of causality, from the cause (parent node) to the effect (child node). Typically, BN structures can be specified by an expert in the domain of interest or consulted from the data via a learning algorithm. However, these methods have some limitations in the medical field. Generating a Bayesian network from a dataset can include casual relationships that are not possible or have no clinical meaning (e.g., causal links like gender to age). On the other hand, a network constructed by an expert might be biased based on prior experts'' knowledge and experience of the domain. In this study, we develop and validate a Bayesian network structure to predict local recurrence in patients diagnosed with locally advanced rectal cancer.

A retrospective cohort of 8566 diagnosed locally advanced rectal cancer patients from 2004 to 2014 from 14 international trial cohorts are analyzed for this study. A stratified  80-20 percent split per trial cohorts is used to train and validate the developed BN structure, respectively.  Continuous variables are categorized, and missing values are considered as a category (Unknown) for all variables but the response.  Multiple expert''s domain knowledge from different radiotherapy institutions is employed to develop and validate the Bayesian network structure.  Experts from Rome defined the casual relationship links between the variables which were independently reviewed by experts from Maastricht and Korea.  Our analysis was conducted in R version 3.6.1 using the bnlearn package and GeNIe (Graphical Network Interface) application.   The SMOTE function was used to adjust the rare event on the response of interest. The model performance is assessed by generating calibration plots and calculating the area under the receiver operating characteristics curve (AUC) on both training and validation datasets.

We excluded 1023 (15%) patients from the training cohort and 268 (16%) from the validation cohort due to information missing-at-random. The radiotherapy dose and adjuvant chemotherapy are excluded from the final structure because they did not influence the response of interest (no direct or indirect parental link). Figure 1 shows the developed Bayesian network structure and its calibration performance on the training data.

Table 1 shows the mean Accuracies, AUCs, and confidence intervals when the structure was used to predict two, three, and five-years local recurrence on the training and validation data.

We have developed and validated a Bayesian network structure from 14 trial cohort data for predicting local recurrence in locally advanced rectal cancer patients. The causal relationships between the variables in the structure are developed and validated by domain experts from different radiotherapy centers where treatment protocols may differ