Bayesian Binary Mixture Models as a Flexible Alternative to Cut-Off Analysis of ELISA Results, a Case Study of Seoul Orthohantavirus

Serological assays, such as the enzyme-linked immunosorbent assay (ELISA), are popular tools for establishing the seroprevalence of various infectious diseases in humans and animals. In the ELISA, the optical density is measured and gives an indication of the antibody level. However, there is variability in optical density values for individuals that have been exposed to the pathogen of interest, as well as individuals that have not been exposed. In general, the distribution of values that can be expected for these two categories partly overlap. Often, a cut-off value is determined to decide which individuals should be considered seropositive or seronegative. However, the classical cut-off approach based on a putative threshold ignores heterogeneity in immune response in the population and is thus not the optimal solution for the analysis of serological data. A binary mixture model does include this heterogeneity, offers measures of uncertainty and the direct estimation of seroprevalence without the need for correction based on sensitivity and specificity. Furthermore, the probability of being seropositive can be estimated for individual samples, and both continuous and categorical covariates (risk-factors) can be included in the analysis. Using ELISA results from rats tested for the Seoul orthohantavirus, we compared the classical cut-off method with a binary mixture model set in a Bayesian framework. We show that it performs similarly or better than cut-off methods, by comparing with real-time quantitative polymerase chain reaction (RT-qPCR) results. We therefore recommend binary mixture models as an analysis tool over classical cut-off methods. An example code is included to facilitate the practical use of binary mixture models in everyday practice.

Automated summary

Serological assays like ELISA are important for detecting infectious diseases, but the classical cut-off approach may not be optimal for analyzing serological data. In contrast, binary mixture models can handle data heterogeneity and uncertainty more effectively, without the need for correction based on sensitivity and specificity.