Combining multivariate cumulative sum control charts with principal component analysis and partial least squares model to detect sickness behaviour in dairy cattle

The present study investigated the suitability of latent variables, generated by principal component analysis (PCA) and partial least squares regression (PLS), for the early detection of behavioural changes due to developing sickness. Therefore, behavioural information was collected from 480 milking cows between September 2018 and May 2019 on a German dairy farm. All animals were equipped with two sensor systems delivering information about the behavioural patterns resting, activity, feeding and rumination. In addition, performance parameters were provided by the milking parlour. The sensor information was combined in seven different ways to create scenarios (C1-C7) that are potentially available on-farm. Diagnoses, treatments and claw trimmings were provided by the farm's veterinarian and claw trimmer. 298 cows with 44,865 days of observations were selected from all the milking cows in consideration of different data restrictions such as missing values; hence a data set was created that included 154 healthy and 144 sick cows with 300 sickness events. For the analyses, the data set was subdivided into ten training data sets (90% of the cows) which were used to set the necessary number of principal components (PCs) and PLS-factors, respectively, and ten testing (10% of the cows) data sets. After selecting PCs and PLS-factors from each scenario, the training data sets were used to train the reference value (k) and threshold value (h) of the multivariate cumulative sum control charts (MCUSUM). The best performing combination of k and h was then used for testing accuracy of the approaches. Hence, 2 (C1) to 6 (C7) PCs were chosen that jointly explained > 70% of the data's variance. Within the PLS approach, 3 (C1) to 10 (C7) PLSfactors were selected that explained the variation of the health status. The PCA-MCUSUMs showed consistent sickness detection as the block sensitivities showed a range from 69.9% to 77.2% (training: 71.0% to 75.8%) and specificities varied from 85.3% to 89.3% (training: 85.2% to 89.4%). The PLS-MCUSUMs showed some irregularities. Here, scenarios C5 and C7 detected > 83% and > 94% sickness events in training and testing, thus causing a decrease of specificities and therefore increased false positive rates of > 20%. In summary, both approaches could be applicable in practice, although the results of the PCA are more consistent and could be more reliable in comparison to the PLS approach.