Development and evaluation of a milk yield-based indexing framework for mastitis monitoring in dairy cows.

OBJECTIVE: Mastitis is a major disease affecting dairy cows that requires timely detection to enable effective management and reduce substantial economic losses. In this study, we developed a practical mastitis monitoring framework based exclusively on daily milk yield, a routinely collected and noninvasive variable readily available from automated milk-recording systems. METHODS: Daily milk yield records (n = 121,692) from 555 Holstein cows on four Korean dairy farms were analyzed. Lactations were reconstructed, and mastitis datasets were matched 1:1 with non-mastitis controls by farm, date, and days in milk. After completeness screening, 63 matched pairs were divided into development (44 pairs) and test datasets (19 pairs). Expected yield was estimated using exponential smoothing and a modified Wilmink equation. The index combined milk yield reduction, a day-specific 90% lower prediction threshold, and consecutive negative deviations. Index changes were compared across 7-, 9-, and 11-day intervals, and alert rules based on index level, relative decline, or both were optimized in the development dataset and evaluated in the test dataset. RESULTS: In mastitis datasets, mean index scores progressively decreased from the pre-diagnosis to post-diagnosis intervals across all window sizes, with significant differences among all three intervals (p<0.05). No significant interval-dependent differences were observed in matched non-mastitis datasets. In the development dataset, the highest F1-score was 0.800 at day +7 using a 3-day evaluation window and a 5% relative-decline rule. In the test dataset, the highest and most balanced performance was observed from days +4 to +6, with accuracy, sensitivity, specificity, and F1-score all reaching 0.790. CONCLUSION: Sustained deviations from expected daily milk production patterns can be converted into an interpretable alert signal for mastitis-associated changes. The proposed framework may support practical and noninvasive identification of cows requiring closer observation or confirmatory testing without additional sensors or invasive sampling.