Sørensen L P, Bjerring M, Løvendahl P
Department of Molecular Biology and Genetics, Center for Quantitative Genetics and Genomics, DK-8830 Tjele, Denmark.
Department of Animal Science, Research Centre Foulum, Aarhus University, DK-8830 Tjele, Denmark.
J Dairy Sci. 2016 Jan;99(1):608-20. doi: 10.3168/jds.2014-8823. Epub 2015 Nov 5.
This study presents and validates a detection and monitoring model for mastitis based on automated frequent sampling of online cell count (OCC). Initially, data were filtered and adjusted for sensor drift and skewed distribution using ln-transformation. Acceptable data were passed on to a time-series model using double exponential smoothing to estimate level and trends at cow level. The OCC levels and trends were converted to a continuous (0-1) scale, termed elevated mastitis risk (EMR), where values close to zero indicate healthy cow status and values close to 1 indicate high risk of mastitis. Finally, a feedback loop was included to dynamically request a time to next sample, based on latest EMR values or errors in the raw data stream. The estimated EMR values were used to issue 2 types of alerts, new and (on-going) intramammary infection (IMI) alerts. The new alerts were issued when the EMR values exceeded a threshold, and the IMI alerts were issued for subsequent alerts. New alerts were only issued after the EMR had been below the threshold for at least 8d. The detection model was evaluated using time-window analysis and commercial herd data (6 herds, 595,927 milkings) at different sampling intensities. Recorded treatments of mastitis were used as gold standard. Significantly higher EMR values were detected in treated than in contemporary untreated cows. The proportion of detected mastitis cases using new alerts was between 28.0 and 43.1% and highest for a fixed sampling scheme aiming at 24h between measurements. This was higher for IMI alerts, between 54.6 and 89.0%, and highest when all available measurements were used. The lowest false alert rate of 6.5 per 1,000 milkings was observed when all measurements were used. The results showed that a dynamic sampling scheme with a default value of 24h between measurements gave only a small reduction in proportion of detected mastitis treatments and remained at 88.5%. It was concluded that filtering of raw data combined with a time-series model was effective in detecting and monitoring mastitis status in dairy cows when based on IMI alerts, and by using a dynamically adjusting sampling scheme almost full performance was still obtainable. However, results were less desirable when based on new alerts most likely because of the used gold standard for mastitis, which may not necessarily reflect the onset of and IMI case in contrast to a new alert.
本研究提出并验证了一种基于在线细胞计数（OCC）自动频繁采样的乳腺炎检测与监测模型。首先，使用自然对数变换对数据进行过滤，并针对传感器漂移和分布偏斜进行调整。将可接受的数据传递给使用双指数平滑的时间序列模型，以估计奶牛水平的水平和趋势。OCC水平和趋势被转换为连续的（0 - 1）尺度，称为乳腺炎风险升高（EMR），其中接近零的值表示奶牛健康状态，接近1的值表示乳腺炎高风险。最后，纳入一个反馈回路，根据最新的EMR值或原始数据流中的错误动态请求下一次采样时间。估计的EMR值用于发出两种类型的警报，新的和（正在进行的）乳房内感染（IMI）警报。当EMR值超过阈值时发出新警报，对于后续警报发出IMI警报。新警报仅在EMR低于阈值至少8天后发出。使用时间窗口分析和不同采样强度下的商业牛群数据（6个牛群，595,927次挤奶）对检测模型进行评估。记录的乳腺炎治疗情况用作金标准。在接受治疗的奶牛中检测到的EMR值显著高于同期未治疗的奶牛。使用新警报检测到的乳腺炎病例比例在28.0%至43.1%之间，对于旨在测量间隔24小时的固定采样方案最高。IMI警报的比例更高，在54.6%至89.0%之间，当使用所有可用测量值时最高。当使用所有测量值时，观察到最低误报率为每1000次挤奶6.5次。结果表明，测量间隔默认值为24小时的动态采样方案在检测到的乳腺炎治疗比例上仅略有降低，仍为88.5%。得出的结论是，基于IMI警报，对原始数据进行过滤并结合时间序列模型在检测和监测奶牛乳腺炎状态方面是有效的，并且通过使用动态调整的采样方案几乎仍可获得全部性能。然而，基于新警报时结果不太理想，这很可能是因为所用的乳腺炎金标准，与新警报相比，它不一定能反映IMI病例的发生情况。
