Guideline generation from data by induction of decision tables using a Bayesian network framework

Decision tables can be used to represent practice guidelines effectively. In this study we adopt the powerful probabilistic framework of Bayesian Networks (BN) for the induction of decision tables. We discuss the simplest BN model, the Naive Bayes and extend it to the Two-Stage Naive Bayes. We show that reversal of edges in Naive Bayes and Two-stage Naive Bayes results in simple decision table and hierarchical decision table respectively. We induce these graphical models for dementia severity staging using the Clinical Dementia Rating Scale (CDRS) database from the University of California, Irvine, Alzheimer's Disease Research Center. These induced models capture the two-stage methodology clinicians use in computing the global CDR score by first computing the six category scores of memory, orientation, judgment and problem solving, community affairs, home and hobbies and personal care, and then the global CDRS. The induced Two-Stage models also attain a clinically acceptable performance when compared to domain experts and could serve as useful guidelines for dementia severity staging.