Estrus detection in tie-stall housed cows through supervised machine learning using a multimodal tail-attached device

•New tail-attached device with thermistor and 3-axis accelerometer was developed.•Estrus associated physiological and behavioral changes were revealed.•Machine learning based estrus detection models were developed.•The best model could detect estrus with 92% sensitivity and 55% precision. In this study, the authors developed a novel multimodal tail-attached device with a thermistor and 3-axis accelerometer that can monitor physiological (body surface temperature (ST)) and behavioral (activity intensity, standing time, and posture changes: standing to lying or vice versa) parameters in cattle. The applicability of the device for estrus detection was also evaluated based on supervised machine learning using the data obtained by continuous measurements of these parameters in tie-stalled cows. Data were collected at 3-min intervals for a total of 25 estrous cycles of 13 cows from Day 10 of the cycle (Day 0 = ovulation day) until Day 11 of the subsequent cycle. Estrus was confirmed based on the standing-to-be-mounted behavior by testing with herd mates at 6-h intervals. After the end of estrus, transrectal ultrasonography was performed every 2 h to confirm ovulation time. The maximum ST, average activity, standing time, and number of posture changes were calculated per hour. For analysis, the calculated data were expressed as residual values and ratios, namely, residual ST (rST = actual ST − mean ST for the same time in the previous three days) and ratios of activity intensity, standing time, and posture change (each ratio = total value during the last 24 h / total value during the last 24–48 h). The mean rST and ratios of activity intensity and standing time gradually increased to approximately 60 h before ovulation, with a peak at estrus (approximately 24 h before ovulation). The posture change ratio started to decrease from approximately 48 h before ovulation and reached its lowest level at the onset of estrus (approximately 30 h before ovulation). A total of 20 features (5 physiological and 15 behavioral features) that could to follow-up the estrus-associated changes were extracted and used for developing nine estrus detection models based on 3 feature sets (physiological features alone, behavioral features alone, and a combination of these features) and 3 machine learning algorithms (decision tree, artificial neural network, and support vector machine (SVM)). Cross-validation showed that the sensitivity of the models tended to be high when using a combinat