Application of behavior data to predictive exploratory models of metritis self-cure and treatment failure in dairy cows

The list of standard abbreviations for JDS is available at adsa.org/jds-abbreviations-24. Nonstandard abbreviations are available in the Notes. The objective was to evaluate the performance of exploratory models containing routinely available on-farm data, behavior data, and the combination of both to predict metritis self-cure (SC) and treatment failure (TF). Holstein cows (n = 1,061) were fitted with a collar-mounted automated-health monitoring device (AHMD) from −21 ± 3 to 60 ± 3 d relative to calving to monitor rumination time and activity. Cows were examined for diagnosis of metritis at 4 ± 1, 7 ± 1, and 9 ± 1 d in milk (DIM). Cows diagnosed with metritis (n = 132), characterized by watery, fetid, reddish/brownish vaginal discharge (VD), were randomly allocated to 1 of 2 treatments: control (CON; n = 62), no treatment at the time of metritis diagnosis (d 0); or ceftiofur (CEF; n = 70), subcutaneous injection of 6.6 mg/kg of ceftiofur crystalline-free acid on d 0 and 3 relative to diagnosis. Cure was determined 12 d after diagnosis and was considered when VD became mucoid and not fetid. Cows in CON were used to determine SC, and cows in CEF were used to determine TF. Univariable analyses were performed using farm-collected data (parity, calving season, calving-related disorders, body condition score, rectal temperature, and DIM at metritis diagnosis) and behavior data (i.e., daily averages of rumination time, activity generated by AHMD, and derived variables) to assess their association with metritis SC or TF. Variables with P-values ≤0.20 were included in the multivariable logistic regression exploratory models. To predict SC, the area under the curve (AUC) for the exploratory model containing only data routinely available on-farm was 0.75. The final exploratory model to predict SC combining routinely available on-farm data and behavior data increased the AUC to 0.87, with sensitivity (Se) of 89% and specificity (Sp) of 77%. To predict TF, the AUC for the exploratory model containing only data routinely available on-farm was 0.90. The final exploratory model combining routinely available on-farm data and behavior data increased the AUC to 0.93, with Se of 93% and Sp of 87%. Cross-validation analysis revealed that generalizability of the exploratory models was poor, which indicates that the findings are applicable to the conditions of the present exploratory study. In summary, the addition of behavior data contributed to increasing the prediction of SC