Evaluation of Approaches to Cattle Calving Prediction Based on Vaginal Temperature: Machine Learning-Versus Threshold-Based Models.

Prediction of calving among cattle based on body temperature is used to prevent calving accidents. However, conventional threshold-based methods often miss or misdetect signs of calving. We quantitatively compared a machine learning-based model with conventional threshold-based methods for calving prediction using vaginal temperature data from wearable devices. Vaginal temperature was recorded at 5-min intervals from approximately 7 days before the expected calving date until parturition in beef (n = 77) and dairy cattle (n = 78). Twenty-six time-series features were extracted, and a machine learning-based calving prediction model was developed using gradient-boosted decision trees. The model predicting calving within 24 h achieved 84.4% and 94.9% sensitivity and 60.2% and 56.1% precision in beef and dairy cattle, respectively, compared with 70.1% and 76.9% sensitivity and 38.3% and 51.7% precision for a conventional threshold-based model using fixed threshold values derived from previous literature. Furthermore, the model predicting calving within 12 h achieved performance comparable to that of the 24-h model (81.8% and 84.6% sensitivity; 50.0% and 42.3% precision, respectively), suggesting the feasibility of customizable alert timing across multiple time windows. Further improvements, including larger datasets and refined feature engineering and model optimization, are needed to reduce missed and false detections.