Bayesian network structure for predicting two-year survival in patients diagnosed with non small cell lung cancer

Abstract

Purpose The aim of this study was to develop and internally validate a clinically plausible Bayesian network structure to predict two-year survival in patients diagnosed with non-small cell lung cancer (NSCLC) and primarily treated with (chemo) radiation therapy by combining expert knowledge and a learning algorithm. Summary of background The incidence of lung cancer has been increasing. Healthcare providers are trying to acquire more knowledge of the disease biology to treat their patients better. However, the information available is more than humans can efficiently process. Predictive models such as Bayesian networks, which can intricately represent causal relations between variables, are suitable structures to model this information. However, commonly known methods for developing Bayesian network structures are limited in healthcare. Patients and Methods 545 NSCLC patients treated primarily with (chemo) radiation therapy from Maastro clinic in The Netherlands between 2010 to 2013 were considered to develop this Bayesian network structure. All continuous variables were discretized before analysis. Patients with missing survival status and variables with more than 25% missing information were excluded. The causal relationships (arcs) between variables in the data were determined using the hill-climbing algorithm with domain experts restrictions. The learning algorithm was run on a number of bootstrapped samples (B=400) and for the final structure, we kept the arcs present in at least 70% of the learned structures. Performance was assessed by computing the area under the curve (AUC) values and producing calibration plots based on a 5-fold cross-validation. In addition, an adapted pre-specified expert structure was compared with a structure developed from the method in this study. Results Tumor load was included in the main structure due to its high percentage (37%) of missingness and lack of added value. The final cohort used to develop the structure was reduced to 499, excluding 46(8.4%) patients with missing survival status. The resulting structure mean AUC and confidence interval to predict two-year survival was 0.614 (0.499 - 0.730 ). The AUC of the compared structures was only slightly above the chance level, but the structure based on the method in this study was clinically more plausible. Conclusion The results of this study show that Bayesian network structures which combine expert knowledge with a rigorous structure learning algorithm produce a clinically plausible structure with optimal performance.