Jump to Main Content
Estrous detection by continuous measurements of vaginal temperature and conductivity with supervised machine learning in cattle
- Higaki, Shogo, Miura, Ryotaro, Suda, Tomoko, Andersson, L. Mattias, Okada, Hironao, Zhang, Yi, Itoh, Toshihiro, Miwakeichi, Fumikazu, Yoshioka, Koji
- Theriogenology 2019 v.123 pp. 90-99
- cows, decision support systems, estrus, herds, luteinizing hormone, neural networks, ovulation, progesterone, support vector machines, temperature, tie stalls, ultrasonography, variance
- This study aimed to evaluate the effectiveness of estrous detection technique based on continuous measurements of vaginal temperature (VT) and conductivity (VC) with supervised machine learning in cattle. The VT and VC of 17 cows in tie-stalls were measured using our developed wearable vaginal sensor from Day 11 (Day 0 = ovulation day) to Day 11 of the subsequent estrous cycle at 15-min interval. After the maximum VT and VC were extracted hourly, their changes were expressed as residual VT (rVT = actual VT − mean VT for the same time on the previous 3 days) and as VC ratio (VCr = actual VC/mean VC for the same time on Day 11–13), respectively, and were used for analysis. Trans-rectal ultrasonography was performed to monitor ovarian structure changes. The plasma concentrations of reproductive hormones (progesterone: P4, estradiol-17β: E2, and LH) were measured in the experimental period. Standing estrus was confirmed by testing with herd mates at 3-h interval. The rVT decreased transiently, which coincided with decreasing P4 a few days before estrus, and a sharp increase was associated with LH surge during estrus. The VCr increased as estrus approached, corresponding with decreasing P4 and increasing E2 and LH. After noise reduction, features, possible to follow-up estrus-associated changes in rVT and VCr, were extracted and used for developing estrous detection models; 9 models were developed with 3 feature sets (features extracted from rVT alone, VCr alone, and combination of rVT and VCr) and 3 machine learning algorithms (decision tree: DT, support vector machine: SVM, and artificial neural network: ANN). Cross-validation showed that models using the features from the combination of rVT and VCr showed better performance in terms of sensitivity and precision than those using features from VCr alone, and precision than those of using features from rVT alone. Within the models using the features from the combination of rVT and VCr, sensitivity and precision of the model generated by ANN were numerically, but not statistically, higher than those generated by DT and SVM. Of 17 estruses, 16 were detected, with one false positive, when the best model was used. Furthermore, both mean and variance of the interval from the beginning of the estrous detection alert to ovulation (27.3 ± 6.7 h, mean ± SD of 16 estruses) were not significantly different to those from the onset of standing estrus to ovulation (30.8 ± 5.8 h, n = 17), indicating that the estrus can be detected real-time by the present technique. Hence, the estrous detection technique based on continuous measurements of VT and VC with supervised machine learning has a potential for efficient and accurate estrous detection in cattle.