An Empirical Validation of Recursive Noisy OR (RNOR) Rule for Asthma Prediction

Abstract

In 2004, an extension of the Noisy-OR formalism termed the Recursive Noisy-OR (RNOR) rule was published for estimating complex probabilistic interactions in a Bayesian Network (BN). The RNOR rule presents an algorithm to construct a complete conditional probability distribution (CPD) of a node while allowing domain causal relationships over and above causal independence to be tractably captured in a semantically meaningful way. However, to the best of our knowledge, the accuracy of this rule has not been tested empirically. In this paper, we report the results of a study that compares the performance of a data-trained expert BN (empiric BN) with the reformulated BN, using the RNOR rule. The original empiric BN was trained with a large dataset from the Regenstrief Medical Record System (RMRS). Furthermore, we evaluate conditions in our dataset which render the RNOR rule inapplicable and discuss our use of Noisy-OR calculations in such situations. We call this approach "Adaptive Recursive Noisy-OR".