Hoof lesion detection in lactating Holsteins: Part II. Development and on-farm validation of a predictive...

Hoof lesion detection remains a challenge in lameness management on dairy farms. Recent studies have proposed locomotion score (LS)-based thresholds using an autonomous camera system (AUTO), which generates BCS and LS. However, these thresholds have limited ability to distinguish affected cows. In the first phase, the primary objective was to develop and tune classification models optimized for the F_0.5 scores, which emphasize positive predictive value (PPV) over sensitivity. In the second phase, which was a prospective, live implementation design, the primary objective was to evaluate the positive classifications generated by the top-performing approach identified in phase 1, with a secondary objective of comparing its PPV to that of the farm's existing passive surveillance approach, defined as on-farm identification of cows for hoof trimming independent of structured locomotion scoring. In phase 1, a total of 511 Holstein cows between 8 and 100 DIM with no typically treated hoof lesions and no more than one prior noninfectious lesion were enrolled from 2 sites. Cows were classified based on hoof trimming outcomes as typically treated (TT), which included hoof lesions typically requiring treatment (e.g., digital dermatitis, sole ulcer, white line disease), or as typically not treated (TNT), which included cows without hoof lesions or without lesions typically requiring treatment. Five machine learning algorithms incorporating AUTO BCS and LS, along with health and milk production data, were tuned and evaluated across multiple feature sets using cross-validation. For the on-farm evaluation, cows similar to the enrollment criteria were processed by the model, and those classified as positive were trimmed by the on-site hoof trimmer. The PPV was calculated against on-site hoof trimmer-reported nonhealthy outcomes and compared with PPV from cows flagged through passive surveillance using the same subset. During development, the top-performing model was a random forest model incorporating AUTO-derived BCS and LS features from the 14-d period before hoof trimming, selected based on the highest F_0.5 score (59.1), with PPV values of 100% (95% CI: 100%, 100%) in training and 92.3% (95% CI: 76.4%, 100%) in testing. During live implementation, this random forest model yielded a lower PPV of 13.7% (22/187; 95% CI: 9.9%-17%). Meanwhile, passive surveillance yielded a greater PPV of 48.6% (34/70; 95% CI: 40%-57.9%). Both passive surveillance and random forest identified a total of 78 cows with an outcome at hoof trimming. Of these, 12.8% (10/78) were flagged by both random forest and passive surveillance, 39.7% (31/78) by the random forest model only, and 47.4% (37/78) by passive surveillance only. Although the random forest model demonstrated lower PPV than passive surveillance, each approach identified cows that were not detected by the other. Although the model resulted in potentially more unnecessary trims, it also flagged cows that would have otherwise gone undetected. These findings highlight the potential complementary value of the model for hoof lesion detection; however, increased unnecessary trims may raise costs or occupy trimming capacity, delaying treatment for cows requiring intervention.