Bayesian Classification of Vasculitis: A Simulation Study

Abstract

New classification criteria for vasculitic disorders have recently been proposed by the American College of Rheumatology. These classification criteria have limitations inherent to the method employed in their development. We propose a different approach to the quantitative analysis of the manifestations of vasculitis, which may improve the precision of classification criteria in this domain. Bayesian classifiers were developed for six vasculitides using literature-derived quantitative descriptions of these syndromes. These clinical data were also used in computer programs designed to generate simulations of vasculitis and control cases. The performance of Bayesian classifiers of vasculitis was then compared to that of the American College of Rheumatology criteria, using series of computer-simulated vasculitis cases. Bayesian classifiers identified simulated vasculitis cases with greater accuracy than those of the corresponding American College of Rheumatology 1990 vasculitis criteria in all six diseases studied. As predicted by theoretical considerations, Bayesian classifiers have the potential to identify vasculitis cases more accurately than the proposed American College of Rheumatology 1990 criteria.