Towards on-site automatic detection of noxious events in dairy cows

Successful detection of pain in cows could circumscribe the therapeutic window for treatment before the cow's condition deteriorates further. While severe clinical cases that are characterized as painful have clear behavioral and physiological manifestations, mild pain may go unnoticed in cows due to their stoic nature. This work presents the first step in developing a warning system that will enable the identification of mild pain in dairy cows. In a set of three experiments, a topical application of 10 % capsaicin cream was used to elicit a noxious sensation. Experiment 1 was aimed at establishing capsaicin as a noxious model for bovines (n = 11). Each cow was treated with neutral cream on day one and the noxious cream on day two. Since the duration of capsaicin effects on bovine skin is unknown, Experiment 2 was designed to evaluate capsaicin's impact on bovines 30 min after application (n = 17). Physiological signs were collected in response to the capsaicin cream application and were compared to the application of the neutral cream. In Experiment 3 physiological signs and continuous behavioral data were recorded (n = 22, four cows participated in Experiment 1). Each cow was treated with neutral cream on day one and the noxious cream on days two and three, i.e., repeated exposure to the noxious stimulus. Heart and breathing rates were elevated soon after the noxious treatment but not for the neutral cream. Blood oxygen saturation was inconclusive. Changes in daily activity patterns consecutive to the noxious challenge included a decrease in rumination time and an increase in lying bouts. These results are in line with what would be expected for physiological and behavioral effects of pain in cows. Additional data are required to rule out habituation or sensitization to the procedure. The resulting database was then used to develop a machine-learning algorithm to detect noxious sensations by applying random forest classifiers trained with two different approaches. The learning herd approach, in which a specialist labels a set of observations and uses them to derive a classifier for new observations from the same herd, achieved 82 % ±9% accuracy. The unlearning herd approach, in which a single database is used to train a classifier that can be applied to members of other herds, resulted in an accuracy of 86 %±18 %. The data discussed in this study meet the requirements of an automatic on-site noxious detection system; real-time on-farm measurements, informative of negative high-arousal states.