Development and validation of a predictive model for calving time based on sensor measurements of ingestive behavior in dairy cows

Calving is an event with major impact on working routines in dairy farming and highly affects the physiological state of dairy cows. Hence, it is in the interest of livestock farmers to have information on approaching calving events to ensure a sound birth, health and welfare of the dairy cow and calf for profitable and sustainable milk production. Changes in the ingestive behavior of dairy cows due to the onset of calving have been revealed in several studies. Therefore, sensor data of these behaviors may be useful for automated prediction of calving time. The current study used sensor data of a novel monitoring device for ingestive behavior (RumiWatch noseband sensor, Agroscope, Ettenhausen, Switzerland and Itin + Hoch GmbH, Liestal, Switzerland) of 35 dairy cows for development and validation of a predictive model for calving time. Sensor data obtained from calving events on three farms were used as one training dataset and two independent validation datasets to evaluate the predictive performance of a Naive Bayes classifier model for calving prediction at 1 h before the start of calving. The model performance was evaluated on an hourly basis for 168 h prior to the start of calving. Combined sensor variables with highest predictive performance were ruminating chews, ruminating boluses, and eating chews (sensitivity = 0.82, specificity = 0.87, positive predictive value = 0.04) in Validation Dataset 1, and ruminating chews per bolus, ruminating chews per minute, eating chews, other activity time, other chews (sensitivity = 0.69, specificity = 0.86, positive predictive value = 0.03) in Validation Dataset 2. These results indicate, that the sensitivity and specificity of the predictive model were satisfying, but the positive predictive value was low and the amount of false positive alerts was considerably high. Although the developed model is therefore not suitable for application in practice, we found that particularly variables of rumination behavior have predictive value and should be taken into consideration for future research on calving detection models. The findings of this study demonstrate that an assessment limited to the terms of sensitivity and specificity may be misleading, as these variables may achieve high values and suggest adequate performance, while the model is not appropriate in the light of its expected use.