Answers are grounded in this paper and cite their sources.
Effective calving monitoring is a critical component of modern dairy farm management, and accurate forecasting of parturition timing provides a foundational framework for implementing optimal calving surveillance protocols. In recent years, deep learning approaches leveraging behavioral data have emerged as promising tools for predicting parturition in dairy cattle. To more effectively capture amplitude variations and temporal dependencies in tail acceleration data, this study assessed the performance of a standalone Transformer model and two hybrid models—Convolutional Neural Network-Transformer (CNN-Transformer) and Long Short-Term Memory-Transformer (LSTM-Transformer). Tail acceleration data were collected from 28 cows for model training and an additional 17 cows for model testing. Given the inherent sample imbalance between prenatal and non-prenatal acceleration data, Time-series Generative Adversarial Networks (TimeGAN) was employed to augment prenatal-stage data, ensuring balanced class distribution for model training. Performance results showed that: the standalone Transformer with three Transformer blocks (Transformer_3layer) achieved optimal performance, with an Accuracy of 0.935, Precision of 0.952, Recall of 0.918, F1-score of 0.935, and Loss of 0.076; The hybrid CNN-Transformer and LSTM-Transformer, each with one Transformer block (CNN-Transformer_1layer and LSTM-Transformer_1layer), achieved optimal performance with respective metrics: Accuracy of 0.944 and 0.939, Precision of 0.940 and 0.945, Recall of 0.950 and 0.933, F1-score of 0.945 and 0.939, and Loss of 0.064 and 0.069, respectively. Bootstrap resampling revealed no significant differences in classification metrics (Accuracy, Precision, Recall, F1-score) across the three models, while Transformer_3layer and CNN-Transformer_1layer showed lower Loss value than LSTM-Transformer_1layer model, with Mean difference ± std: -0.042 ± 0.013 (95% CI: [-0.074, -0.011]) and Mean difference ± std: -0.051 ± 0.014 (95% CI: [-0.080, -0.024]). DeLong’s test indicated that the Transformer_3layer had a significantly smaller Area Under the ROC Curve than the CNN-Transformer_1layer (z = -2.19, p = 0.005). These preliminary findings demonstrate the feasibility of using standalone Transformer, hybrid CNN-Transformer, and LSTM-Transformer models with tail acceleration data to predict calving time in dairy cattle. This work thereby offers practical approaches to optimize calving management and enhance animal welfare through improved monitoring.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: