Evaluation of 4 predictive algorithms for intramammary infection status in late-lactation cows

Abstract

The objective of this observational study was to compare 4 cow-level algorithms to predict cow-level intramammary infection (IMI) status (culture and MALDI-TOF) in late-lactation US dairy cows using standard measures of test performance. Secondary objectives were to estimate the likely effect of each algorithm, if used to guide selective dry cow therapy (SDCT), on dry cow antibiotic use in US dairy herds, and to investigate the importance of including clinical mastitis criteria in algorithm-guided SDCT. Cows (n = 1,594) from 56 US dairy herds were recruited as part of a previously published cross-sectional study of bedding management and IMI in late-lactation cows. Each herd was visited twice for sampling. At each farm visit, aseptic quarter-milk samples were collected from 20 cows approaching dry-off (>180 d pregnant), which were cultured using standard bacteriological methods and MALDI-TOF for identification of isolates. Quarter-level culture results were used to establish cow-level IMI status, which was considered the reference test in this study. Clinical mastitis records and Dairy Herd Improvement Association test-day somatic cell count data were extracted from herd records and used to perform cow-level risk assessments (low vs. high risk) using 4 algorithms that have been proposed for SDCT in New Zealand, the Netherlands, United Kingdom, and the United States. Agreement between aerobic culture (reference test; IMI vs. no-IMI) and algorithm status (high vs. low risk) was described using Cohen's kappa, test sensitivity, specificity, negative predictive value, and positive predictive value. The proportion of cows classified as high risk among the 4 algorithms ranged from 0.31 to 0.63, indicating that these approaches to SDCT could reduce antibiotic use at dry-off by 37 to 69% in the average US herd. All algorithms had poor agreement with IMI status, with kappa values ranging from 0.05 to 0.13. Sensitivity varied by pathogen, with higher values observed when detecting IMI caused by Streptococcus uberis, Streptococcus dysgalactiae, Staphylococcus aureus, and Lactococcus lactis. Negative predictive values were high for major pathogens among all algorithms (≥0.87), which may explain why algorithm-guided SDCT programs have been successfully implemented in field trials, despite poor agreement with overall IMI status. Removal of clinical mastitis criteria for each algorithm had little effect on the algorithm classification of cows, indicating that algorithms based on SCC alone may have similar performance to those based on SCC and clinical mastitis criteria. We recommend that producers implementing algorithm-guided SDCT use algorithm criteria that matches their relative aspirations for reducing antibiotic use (high specificity, positive predictive value) or minimizing untreated IMI at dry-off (high sensitivity, negative predictive value).