Adsa Logo White Adsa Title White

Performance of different machine learning methods for prediction of the health status of lactating dairy cows.

M. M. Pérez

Events

06-22-2020

Abstract:

73
Performance of different machine learning methods for prediction of the health status of lactating dairy cows.
M. M. Pérez*1, Y. You2, Y. Wang2, K. Q. Weinberger2, D. V. Nydam3, J. O. Giordano1. 1Department of Animal Science, Cornell University Ithaca, NY, 2Department of Computer Science, Cornell University Ithaca, NY, 3Department of Population Medicine and Diagnostic Sciences, Cornell University Ithaca, NY.

Our objective was to evaluate the performance of different machine learning methods (ML) to predict the health status of dairy cows based on multiple sensor and non-sensor data. The clinical health status (clinical or no clinical disease) of lactating Holstein cows (n = 1,211) was determined based on daily clinical examination from 1 to 30 DIM. Disorders recorded were: metritis, mastitis, ketosis, indigestion, displaced abomasum, and pneumonia. Cows were considered to have a clinical disorder for all days at which any of these conditions were recorded. Sensor data offered to ML models were: physical activity, resting behavior, reticulo-rumen temperature, rumination, eating behavior, and environmental temperature and humidity from −21 to 30 DIM. After calving data was also available for BW and daily milk volume, conductivity, and components (fat, protein, lactose). Non-sensor data used were: previous health and reproductive events, production records, and pen stocking density. Models created and evaluated using Python included: XGBoost, Multi-Layer Perceptron (MLP) and Recurrent Neural Network (RNN). Data available was split to use 80% for training and 20% for testing. Sensitivity (Se) and specificity (Sp) for predicting the occurrence of clinical health disorders were estimated. For MLP Se and Sp for the training set were 100% and 98% and for testing set were 43% and 96%. For RNN Se and Sp for the training set were 99% and 67% and for testing set were 70% and 67%. For XGBoost Se and Sp for the training set were 95% and 90% and for testing set were 88% and 88%. Models for MLP and RNN tended to overfit the training data and were not able to generalize to the testing data, likely due to the limited training outcomes and unbalanced ratio of positive to negative outcomes. Thus, some predictive models created with ML methods may be effective for predicting the health status of cows when including multiple cow behavioral, physiological, and performance sensor parameters, environmental sensor data, and health, reproductive, and performance records.

Keywords: prediction, disease, dairy cow.