Adsa Logo White Adsa Title White

Predicting feed intake and heath events using sensor data in lactating Holstein cows.

C. J. Siberski



Join C. J. Siberski on this page for a live text chat!
6:00 PM - 8:00 PM GMT


Predicting feed intake and heath events using sensor data in lactating Holstein cows.
C. J. Siberski*1, M. S. Mayes1, P. J. Gorden2, A. Copeland2, M. Healey1, B. M. Goetz1, H. Beiki1, L. M. Kramer1, L. H. Baumgard1, P. Dixon3, J. E. Koltes1. 1Department of Animal Science, Iowa State University Ames, IA, 2Veterinary Diagnostic & Production Animal Medicine, Iowa State University Ames, IA, 3Department of Statistics, Iowa State University Ames, IA.

Feed intake and cow health are highly interrelated, and impact dairy sustainability. The objective of this study was to evaluate the utility of wearable sensor data to predict feed intake and health events [n = 107 Holstein cows; parity 1—4; day in milk (DIM) 50—270; season = summer (n = 47) and fall (n = 60)]. Three automated sensors were evaluated: 2 ear tags (n = 107 and 41) and a rumen bolus (n = 60). Sensors recorded animal activity, ear temperature, rumination, rumen temperature and pH. Additional traits collected included milk yield and components (fat, protein, and lactose), BW, BCS, and health events from veterinary records. Temperature-humidity index (THI) was calculated from a local weather station located near the dairy. Adjusted daily dry matter intake (ADMI) was calculated utilizing PROC GLIMMIX in SAS 9.4, where DMI was adjusted for milk yield and components, BW0.75, DIM, parity and contemporary group. Missing data were imputed utilizing the mice package in R. Prediction was conducted using the random forest algorithm (R-caret package). Feed intake (ADMI) models included sensor, health, BCS and THI. Health models included these variables plus contemporary group, DIM, BW0.75, milk components and yield, parity and ADMI. Models with and without cow were considered. Leave-one-out validation was utilized, in which all records across all DIM from each cow were removed. The caret package also applied a 10-fold cross-validation within each leave-one-out data set. The average coefficient of determination for prediction of ADMI was 0.19—0.23 depending on the sensor measures included. Temperature measurements appeared to have the highest variable importance values for ADMI. The average accuracy of predicting health events was 0.94—0.96 (kappa values 0.72—0.82) for all sensors. For health events, activity and ear temperature had the highest variable importance values for sensor measures. Although validation is needed, sensor measures may be useful in commercial settings to detect health events or fluctuations in feed intake.

Keywords: feed intake, health events, prediction.