Evaluating the performance of metabolic profiling for dairy cattle lameness classification: Biofluid and cross-farm...

Lameness has an unsustainably high prevalence in dairy cattle, causing a major welfare problem for the dairy industry that requires early detection and, ideally, prevention to reduce prevalence. Timely treatment not only improves animal welfare through better treatment outcomes but also reduces production and economic losses, thereby improving the sustainability of dairy herds. Novel methods for classifying current lameness status are needed to enable earlier and more objective detection of lame cows. Although the accuracy of metabolic biomarkers for lameness classification has been investigated in previous studies, most have been conducted within a single herd and have typically focused on a single biofluid. To the best of our knowledge, comparative studies across commercial farms to evaluate metabolic biomarker classification performance between biofluids and across herds have not yet been conducted. In this study, 425 cows from 10 commercial dairy farms were mobility scored once by trained observers (n = 3) using the AHDB scoring scale, with cows scoring ≥ 2 classified as lame and those scoring < 2 as non-lame. This study aimed to investigate how well metabolic biomarkers could classify current lameness status within farms and to compare classification performance across multiple commercial dairy herds, using 2 non-invasively collected biofluids. Samples (424 urine and 425 milk samples) were analyzed using untargeted liquid chromatography-high-resolution mass spectrometry (LC-HRMS). LC-HRMS data generated included 4 machine learning algorithms (Elastic Net, Multivariate Adaptive Regression Spline, Partial Least Squares and Random Forest). Results showed that urine metabolomics achieved higher balanced accuracy for lameness classification across all 10 commercial farms (65.0%) than milk lipidomics (57.7%). Substantial variability in accuracies at the farm level was demonstrated, ranging from 56.6% to 80.1% for urine and 47.5% to 69.8% for milk. Three farms demonstrated notably higher accuracies with urine samples, while 5 farms showed similar performance between biofluids, and 2 farms showed a trend toward higher balanced accuracies for the milk lipidome. Partial Least Squares machine learning algorithm showed the most consistent performance across all 10 farms. While our findings demonstrate superior classification performance of the urine metabolome in this data set, practical considerations such as sample collection ease and storage stability may be relevant when considering suitable biofluids for diagnostic tools. While some farms achieved classification accuracies similar to previous single-herd studies (71-100%), the overall lower accuracies observed in this cross-farm study likely reflect the challenge of model transferability across farms with different management practices, environmental conditions, lameness etiologies and differences in outcome classification. Future multi-herd studies with robust classification of lameness outcomes, including lesions, could further improve accuracy. Larger multi-herd studies could also help understand the role of herd-level factors on model generalizability and performance. These findings provide a multi-herd perspective on the comparative utility of urine and milk metabolomics for lameness classification and illustrate the extent to which diagnostic performance varies across commercial farms. This cross-farm evaluation offers important insights into model generalizability beyond single-herd studies, informing the future development of more robust and transferable approaches to support intervention and improved animal welfare.