Drug dilution (MIC) and disk diffusion (DIA) are the two antimicrobial susceptibility tests used by clinicians to determine an unknown pathogen's susceptibility to various antibiotics. This determination is made in part by comparing the test result to antibiotic-specific breakpoints. Because the MIC test deals with concentrations of a drug, its breakpoints are based primarily on the pharmacokinetics and pharmacodynamics of the drug. However, finding comparable DIA breakpoints, which deals with the diameter of a clear zone, is not straightforward. The traditional approach selects a set of 300-500 pathogens and obtains test values for both tests. The DIA breakpoints are then estimated based on minimizing the observed classification discrepancies.
In 2000, I proposed a Bayesian hierarchical modeling approach to determine DIA breakpoints and showed that this approach increases precision and has better accuracy compared to the traditional approach. The underlying errors-in-variables model takes into account the measurement error of each test as well as the differing test characteristics. This model, however, makes a strong parametric assumption about the relationship between the two true latent test values that may not reasonable in all drug/bug combinations. In this talk I will describe a generalization of this method by considering a more flexible non-parametric relationship between the true test values. This is done using I-Splines (Ramsey 1988), which preserve the monotonically decreasing relationship between the two true tests values. This approach performs comparably to the original model when the assumed parametric relationship suffices, and exceeds it when the parametric assumption is not reasonable. This is joint work with Glen DePalma, Purdue University.