Analysis of sequential ruminal temperature sensor data from dairy cows to identify cow subgroups by clustering and predict calving through supervised machine learning

Abstract

The present study investigated the applicability of a calving prediction model based on supervised machine learning of ruminal temperature (RT) data in dairy cows. The existence of cow subgroups for prepartum RT changes was also examined, and the predictive performance of the model was compared among these subgroups. RT data were collected from 24 Holstein cows at 10 min intervals using an RT sensor system. The average hourly RT was calculated and data were expressed as residual RTs (rRT = actual RT − mean RT for the same time on the previous three days). The mean rRT decreased beginning at approximately 48 h before calving to a low of −0.5°C at 5 h before calving. However, two cow subgroups were identified: cows with a late and small rRT decrease (Cluster 1, n = 9) and those with an early and large rRT decrease (Cluster 2, n = 15). A calving prediction model was developed using five features extracted from the sensor data (indicative of prepartum rRT changes) through a support vector machine. Cross-validation showed that calving within 24 h was predicted with a sensitivity of 87.5% (21/24) and precision of 77.8% (21/27). A significant difference in sensitivity was observed between Clusters 1 and 2 (66.7 vs. 100%, respectively), while none was observed for precision. Therefore, the model based on RT data with supervised machine learning has the potential to efficiently predict calving, although improvements for specific cow subgroups are required.

Keywords

• cattle
• core body temperature
• parturition prediction
• rumen sensor
• support vector machine