Predicting metritis events in dairy cattle using machine learning classifiers on multiple data streams under nowcasting and forecasting frameworks

In recent years an increasing number of precision dairy farming technologies (PDFTs) have been incorporated into the management of dairy operations. Recently, research has been centered on the use of sensors to quantify animal behaviors such as activity level, rumination time, or lying time, and their potential for disease detection. In dairy cows, the transition period around parturition is considered the time when most diseases occur, being hypocalcemia, metritis, or hyperketonemia the most common. Whether the combination of different behaviors registered by sensors can better diagnose diseases during transition period, their sensitivity and specificity to detect diseases, or what is their predictive ability or how far in advance they can detect disease is not known. Our goal was to develop, test and validate a workflow for disease surveillance in dairy cattle with emphasis on metritis, using a combination of feature selection strategies and machine learning algorithms. The long-term goal was to provide a framework where high-frequency time series behavioral data registered by multiple PDFTs could be used as a tool for early detection of dairy cattle health problems during the transition period. Data from 35 dairy cows that either did not experience any disease postpartum or were only diagnosed with metritis were retrospectively selected from a dataset containing behavioral, production, and clinical data from 138 lactating cows during the first 21 days postpartum at the University of Kentucky Coldstream Dairy (Lexington, KY, USA). Metritis events were created based on changes in metritis scores recorded during clinical examination. After a review of PDFTs and machine learning approaches (Chapter 1), Chapters 2 and 3 study the classification performance of three classifiers (k-nearest neighbors, random forest, and support vector machines) when predicting metritis events by using behaviors registered by two different 3-axis accelerometers. Chapter 4 studies the classification performance of a random forest classifier to predict metritis events when multiple inputs from multiple data streams were combined. Multiple time windows, time lags, and classification thresholds were compared under nowcasting (Chapter 2, 3, and 4) and forecasting frameworks (Chapter 4). Random Forest had the greatest F1 score across time windows and time lags, but best behaviors for classification changed depending on the combination of time window and time lag. Furthermore, forecasting metritis events 2 and 3 days forward had similar performance results compared to the nowcasting framework. Based on our findings, machine learning classifiers can aid in the identification of animals at higher risk of being sick before traditional diagnosis is performed