(Sub)clinical hypocalcaemia occurs frequently in the dairy industry, and is one of the earliest symptoms of an impaired transition period. Calcium deficiency is accompanied by changes in cows’ daily behavioural variables, which can be measured by sensors. The goal of this study was to construct a predictive model to identify cows at risk of hypocalcaemia in dairy cows using behavioural sensor data. For this study 133 primiparous and 476 multiparous cows from 8 commercial Dutch dairy farms were equipped with neck and leg sensors measuring daily behavioural parameters, including eating, ruminating, standing, lying, and walking behaviour of the 21 days before calving. From each cow, a blood sample was taken within 48 h after calving to measure their blood calcium concentration. Cows with a blood calcium concentration 2.0 mmol/L were defined as hypocalcemic. In order to create a more context based cut-off, a second way of dividing the calcium concentrations into two categories was proposed, using a linear mixed-effects model with a k-Means clustering. Three possible binary predictive models were tested; a logistic regression model, a XgBoost model and a LSTM deep learning model. The models were expanded by adding the following static features as input variables; parity (1, 2 or 3＋), calving season (summer, autumn, winter, spring), day of calcium sampling relative to calving (0, 1 or 2), body condition score and locomotion score. Of the three models, the deep learning model performed best with an area under the receiver operating characteristic curve (AUC) of 0.71 and an average precision of 0.47. This final model was constructed with the addition of the static features, since they improved the model’s tuning AUC with 0.11. The calcium label based on the cut-off categorization method proved to be easier to predict for the models compared to the categorization method with the k-means clustering. This study provides a novel approach for the prediction of hypocalcaemia, and an ameliorated version of the deep learning model proposed in this study could serve as a tool to help monitor herd calcium status and to identify animals at risk for associated transition diseases.

中文翻译：

---

利用行为传感器数据与深度学习相结合的奶牛低钙血症预测模型

（亚）临床低钙血症在乳制品行业中经常发生，是过渡期受损的最早症状之一。缺钙伴随着奶牛日常行为变量的变化，这些变化可以通过传感器测量。本研究的目的是构建一个预测模型，利用行为传感器数据识别奶牛存在低钙血症风险。在这项研究中，来自 8 个荷兰商业奶牛场的 133 头初产奶牛和 476 头经产奶牛配备了颈部和腿部传感器，用于测量日常行为参数，包括产犊前 21 天的进食、反刍、站立、躺卧和行走行为。在产犊后 48 小时内从每头牛身上采集血样，以测量其血钙浓度。血钙浓度2.0mmol/L的奶牛被定义为低钙血症。为了创建更加基于上下文的截止值，提出了将钙浓度分为两类的第二种方法，即使用具有 k 均值聚类的线性混合效应模型。测试了三种可能的二元预测模型；逻辑回归模型、XgBoost 模型和 LSTM 深度学习模型。通过添加以下静态特征作为输入变量来扩展模型；胎次（1、2或3＋）、产犊季节（夏、秋、冬、春）、相对于产犊的钙采样日（0、1或2）、体况评分和运动评分。在这三个模型中，深度学习模型表现最好，受试者工作特征曲线下面积 (AUC) 为 0.71，平均精度为 0.47。最终模型是通过添加静态特征构建的，因为它们将模型的调整 AUC 提高了 0.11。事实证明，与使用 k 均值聚类的分类方法相比，基于截止分类方法的钙标签更容易预测模型。这项研究提供了一种预测低钙血症的新方法，并且本研究中提出的深度学习模型的改进版本可以作为帮助监测牛群钙状态并识别有相关过渡性疾病风险的动物的工具。