Estimating risk probabilities for sickness from behavioural patterns to identify health challenges in dairy cows with multivariate cumulative sum control charts

Dairy cattle housing is characterised by increasing herd sizes and the need for assisting technical tools to monitor the cows’ health. This study investigated the combination of logistic regression models with multivariate cumulative sum (MCUSUM) control charts in healthmonitoring of dairy cattle. Sensor information of 618 cows with 791 lactations (138 438 cow days), nine behavioural variables were included as parts of the behavioural patterns: physical activity (“neck activity”, “leg activity”, “walking duration”), resting (“lying duration”, “standing duration”, “transitions from lying to standing”) and feeding (“feeding duration”, “rumination duration”, “inactivity duration”) behaviour. For each of these behavioural patterns, a logistic regression model with the health status (sick vs not sick) as a dependent variable was designed after a variable selection (herd level) based on the herd dataset with 618 cows (618 lactations; 115 547 cow days), which included the variables of each behaviour pattern and the stage of lactation nested in the number of lactations as explanatory variables. The explanatory variables were added stepwise to the model, with the final model being selected with respect to the lowest values of Akaike’s and Bayes’ information criteria. Each model was then applied to a dataset with 173 cows (22 891 cow days) at cow level, resulting in individual daily risk probabilities for getting sick. Thus, risk probabilities of each behavioural pattern were estimated and included in the MCUSUM control charts to identify cows at risk of disease. The performance of the MCUSUM control charts was cross-validated to identify the best fitting reference value k and the threshold value h. Alerts given within 5 days prior to diagnosis were counted as detected sicknesses. The performance resulted in a block sensitivity of 70.9–81.4%, specificity of 87.9–94.2% and a false-positive rate of 5.8–12.1%. The performance was confirmed while testing the entire algorithm resulting in a mean area under the receiver operating characteristics curve of 0.89. Calculating precision and the F₁-score resulted in a precision of 49.0–60.9% (training: 48.8–63.5%) and an F₁-score of 61.1–65.7% in testing (training: 61.0–67.0%). The precision-recall curve (PRC) was derived from precision and recall with an area under the PRC of 0.70 in training and testing. In summary, the present study was able to develop an algorithm showing good classification potential for the online monitoring of sickness behaviour.