Byungho Chae, Nag-Jin Choi
Subclinical mastitis, identified through elevated somatic cell count (SCC), remains the most prevalent and costly health issue in dairy cattle, yet milk lactose, biologically linked to udder health and routinely measurable by Fourier-transform infrared (FTIR) analyzers, is rarely utilized in routine herd management. We investigated which routine dairy herd improvement (DHI) variables are most consistently associated with elevated SCC, using 72,547 test-day records from 3,562 Holstein cows across 37 Korean dairy farms (2021-2025), where lactose was not available in DHI records. Four interpretable models (logistic regression, generalized additive model, extreme gradient boosting, and random forest with SHapley Additive exPlanations [SHAP]) were compared under farm-level grouped cross-validation. The solids-not-fat minus protein residual (SNF% - protein%), an unvalidated lactose proxy, was the top-ranked contributor to elevated SCC, and the top-3 ranking (residual, parity, milk yield) was consistent across random forest SHAP values, logistic regression standardized coefficients, and extreme gradient boosting SHAP values, as well as across 4 SCC thresholds, 5 lactation stages, and 5 years of data. Farm-level grouped validation reduced discrimination relative to random cross-validation by an amount larger than the entire model-class difference in performance, indicating that validation design matters more than model choice. External corroboration on an independent European open data set with directly measured FTIR lactose supported the inverse lactose-SCC association and ranked directly measured lactose as the top SHAP contributor; given the small single-farm Belgian cohort, this external evidence is supportive rather than confirmatory. Because milk composition and SCC were recorded on the same test day, the results describe concurrent associations rather than prospective prediction; the positive predictive value of 0.530 also precludes stand-alone individual-cow screening.