Estimation of the accuracy of an ELISA test applied to bulk tank milk for predicting herd-level status for Salmonella Dublin in dairy herds using Bayesian Latent Class Models

Abstract

Enzyme-Linked Immunosorbent Assay (ELISA) test is commonly used for detection of antibodies to Salmonella Dublin in individual bovine milk samples. However, little is known about its accuracy when used on bulk tank milk for determining herd-level S. Dublin status and when evaluated without assuming a perfect reference test.The objectives of this study were: i) to estimate the herd prevalence of S. Dublin among dairy cattle herds in Québec, Canada; ii) to estimate the herd sensitivity and specificity of a commercially available ELISA test when used on bulk milk; iii) to examine how the diagnostic test accuracy varies with different bulk milk ELISA cut-offs; and (iv) to assess the added value of combining ELISA screening of bulk milk and individual serum of 10 animals for determining S. Dublin herd status.A cohort of 302 dairy herds selected in three regions (population 1) and 58 herds that have already tested positive to S. Dublin (population 2) were recruited. A total of 715 bulk milk samples and 7150 individual blood samples from cattle over 3 months old (10 animals per herd) sampled on two occasions were collected. Testing was conducted using PrioCHECK™ Salmonella Ab bovine Dublin ELISA test for milk (Bmilk ELISA: test under investigation) and for serum of 10 individual animals (Serum10 ELISA: imperfect reference test) to determine the herd-level S. Dublin status. A latent class model for two populations, two tests, allowing for conditional dependence between tests was fit within a Bayesian framework. At cut-off PP % ≥ 15 for a Bmilk ELISA, which is used by provincial authorities, the herd prevalence of S. Dublin estimated using informative prior was 6.8 % (4.3–9.9) in population 1. The herd sensitivity and specificity estimates (95 % Bayesian Credibility Intervals) for Bmilk ELISA were 40.6 % (15.6–88.8) and 91.9 % (88.3–95.8), respectively. Positive and negative predictive values of Bmilk ELISA applied in population 1 were 26.4 % (8.5–60.2) and 95.8 % (92.1–99.2), respectively. Increasing Bmilk ELISA cut-offs had little influence on predictive values. The combination of both ELISA tests did not improve the diagnostic accuracy of S. Dublin. Our study shows that a test-positive herd based on a single bulk milk sample would require complementary tests for status confirmation. However, a test-negative herd could be classified as true negative with a high certainty.