Nonparametric calibration of two common susceptibility tests using interval-censored data with measurement error.

摘要

Drug dilution (MIC) and disk diffusion (DIA) are the two common tests to determine pathogen susceptibility to antibiotics. In the drug dilution method, the classification breakpoints are concentrations and based directly on the pharmacokinetics and pharmacodynamics of the drug. For the disk diffusion method, these concentration breakpoints need to be converted into zone diameters and this is not a straightforward calculation. Historically, the error-rate bounded method has been effective in producing reasonable DIA breakpoints. In the last few decades, however, pathogens have become more drug resistant and as a consequence, the error-rate bounded method has become less effective in identifying appropriate breakpoints.;Recently, a statistical model was proposed to determine the DIA breakpoints. While this approach produced very consistent results, there are concerns regarding its robustness due to several parametric assumptions. In this research, we've proposed a more robust method of DIA breakpoint determination by relaxing these parametric assumptions. Based on M-spline theory, we employ nonparametric density estimation and nonparametric monotone regression techniques to estimate the MIC density, the true MIC/DIA relationship, and the DIA breakpoints. The bootstrap is used to assess the uncertainty in the DIA breakpoint estimates.;Our simulation studies have shown that this nonparametric method performs very well. Compared with the error-rate bounded method, the resulting DIA breakpoints are considerably more precise. Our results also show that when the underlying model is that used in the parametric approach, our approach performs comparably to the proposed statistical model. However, when the underlying relationship falls short of the specified parametric form, our approach again performs well while the parametric approach gives biased estimates. While our focus is specifically on DIA breakpoint estimation, we feel our nonparametric density and monotone regression methods could be used in more general errors-in-variables model situations with and without interval censoring.