Representation and incorporation of clinical information in outpatient oncology prognosis using Bayesian networks and Naïve Bayes

© 2016 IEEE. Many studies have focused on prognosis for oncology patients with the following characteristics: an inpatient setting (uniform sampling), binary outcomes, and predictor variables of patient demographics and tumor characteristics. This paper examines the problem of predicting prognosis in an outpatient setting (non-uniform sampling), discrete outcomes, and predictor variables of clinical observations. In particular, we consider how to represent the clinical observational data and reason over the prognosis using Bayesian networks (BN) and Naíve Bayes (NB). Different representations include trend behaviors using splines and differences over a time period and the clinical observations themselves. The best models were able to outperform the majority classifier with 8.5% (BN) and 10.2% (NB) higher accuracies in predicting a patient\u27s length of survival. The models with highest predictive performance both include a temporal behavioral representation