Prediction of 24-h and 6-h Periods before Calving Using a Multimodal Tail-Attached Device Equipped with a Thermistor and 3-Axis Accelerometer through Supervised Machine Learning

Abstract

Simple SummaryRoutine visual observation of signs of imminent calving, such as softening of ligaments around the tailhead and udder distension, is time-consuming, and the resulting calving predictions are relatively unreliable. To address this issue, we used a multimodal tail-attached device (tail sensor) and developed calving prediction models through supervised machine learning. The tail sensor is equipped with a thermistor and 3-axis accelerometer, and can monitor tail skin temperature, activity intensity, lying time, posture changes (standing to lying or vice versa), and tail raising behavior. Using the sensor data with a non-sensor-based data (days to the expected calving date), we developed calving prediction models for 24-h and 6-h periods before calving and evaluated their predictive ability under two distinct housing conditions, tethering (tie-stall) and untethering (free-stall and individual pen). Our results demonstrated that calving prediction models based on tail sensor data with supervised machine learning have the potential to achieve effective calving prediction, irrespective of the cattle housing conditions.In this study, we developed calving prediction models for 24-h and 6-h periods before calving using data on physiological (tail skin temperature) and behavioral (activity intensity, lying time, posture change, and tail raising) parameters obtained using a multimodal tail-attached device (tail sensor). The efficiencies of the models were validated under tethering (tie-stall) and untethering (free-stall and individual pen) conditions. Data were collected from 33 and 30 pregnant cattle under tethering and untethering conditions, respectively, from approximately 15 days before the expected calving date. Based on pre-calving changes, 40 features (8 physiological and 32 behavioral) were extracted from the sensor data, and one non-sensor-based feature (days to the expected calving date) was added to develop models using a support vector machine. Cross-validation showed that calving within the next 24 h under tethering and untethering conditions was predicted with a sensitivity of 97% and 93% and precision of 80% and 76%, respectively, while calving within the next 6 h was predicted with a sensitivity of 91% and 90% and precision of 88% and 90%, respectively. Calving prediction models based on the tail sensor data with supervised machine learning have the potential to achieve effective calving prediction, irrespective of the cattle housing conditions.