Learning dependencies among fetal heart rate features using Bayesian networks.

We present preliminary results on the use of Bayesian-network (BN) structure learning algorithms for deciphering dependencies from amongst different fetal heart rate (FHR) features collected from a real database. We used a greedy search-and-score procedure known as the K2 algorithm for the estimation of the BN structure. The database consists of a collection of discrete-valued features quantifying presence of morphological changes as prescribed by expert guidelines (updated by the National Institute of Child Health and Human Development (NICHD)) and implemented as rule-based programs. We compare the results of structure learning to the expert-guided structure and use decision functions for classification using posterior probabilities. It was found that the BN estimated from structure learning algorithms had comparable classification performance, but fewer edges, leading to more efficient characterization of conditional probability tables (CPD's). Moreover, structure learning algorithms offer a method of learning novel correlations between FHR features that may be exploited for automatic categorization.