Answers are grounded in this paper and cite their sources.
Accelerometers worn by beef calves can record activity, allowing researchers and producers to potentially monitor neonate health and behaviour. The objectives were to develop a machine learning algorithm that predicts behaviour using accelerometer data and compare pre-processing and model-level strategies for improving model performance. Fifteen newborn calves were fitted with a collar containing a 3D-accelerometer with a sampling rate of 1.0 Hz. Three cohorts of 5 cow-calf pairs were recorded with a camcorder 5 d/wk for the first two weeks of life between 0800–2000h. Observers recorded lying, standing, walking, and nursing behaviours using focal animal sampling, resulting in 644 hrs of observations. Observations were synced with the accelerometer data using the animal ID and timestamp, cleaned, and segmented into time-windows of 20 s. Data from 3 calves were separated as independent test sets. From the training set composed of the remaining 12 calves, 132 features were calculated. Fourteen classifiers were developed, including both hierarchical and non-hierarchical models, as well as variations that employed over- and/or under-sampling techniques to manage the imbalanced dataset. A Leave-One-Animal-Out cross-validation (LOAO) was conducted for each classifier that calculated the precision, sensitivity, and the Matthews Correlation Coefficient (MCC). These metrics were averaged across calves and a Friedman test identified differences between models which was followed by pairwise comparisons using a Nemenyi Test and Critical Difference plot visualizations. The final model with the highest predictive performance was then tested with 3 independent test sets of 1 animal each. The dataset was imbalanced because calves spent the majority of their time lying down. The Non-Hierarchical Model (NHM) with no pre-processing sampling techniques performed best overall at LOAO and was further tested in the test sets. Their MCC was 0.78 ± 0.05 in the LOAO and 0.77 ± 0.00 (mean±SD) in the test set. Precision and sensitivity for lying and standing were ≥ 0.79, and both were ≥ 0.65 for nursing for LOAO and the test sets. Walking had the lowest predictive performance (precision: LOAO = 0.26 ± 0.38; sensitivity: LOAO = 0.02 ± 0.04), and the NHM was unable to correctly classify walking in the test set. These preliminary results suggest that hierarchical model and pre-processing sampling techniques attempted could not compensate for the severity of imbalance in the dataset. The final classifier predicted lying, standing, and nursing behaviour reliably, but walking behaviour may not be developed enough in the first two weeks of life to consistently identify it at a sampling rate of 1.0 Hz.
No takes yet. Share an insight, caveat, or question.
Coon et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: