Lung Cancer Survival Prediction Using Instance-Specific Bayesian Networks

Lung cancer is the most common cause of cancer-related death in men worldwide and the second most common cause in women. Accurate prediction of lung cancer outcomes can help guide patient care and decision making. The amount and variety of available data on lung cancer cases continues to increase, which provides an opportunity to apply machine learning methods to predict lung cancer outcomes. Traditional population-wide machine learning methods for predicting clinical outcomes involve constructing a single model from training data and applying it to predict the outcomes for each future patient case. In contrast, instance-specific methods construct a model that is optimized to predict well for a given patient case. In this paper, we first describe an instance-specific method for learning Bayesian networks that we developed. We then use the Markov blanket of the outcome variable to predict 1-year survival in a cohort of 261 lung cancer patient cases containing clinical and omics variables. We report the results using AUROC as the evaluation measure. In leave-one-out testing, the instance-specific Bayesian network method achieved higher AUROC on average, compared to the population-wide Bayesian network method.