Characterizing drinking behavior from reticular temperature with artificial neural networks.
A. E. Pape
Events
06-23-2020
Abstract:
T13
Characterizing drinking behavior from reticular temperature with artificial neural networks.
A. E. Pape*, C. S. Ballard. William H. Miner Agricultural Research Institute Chazy, NY.
The use of reticular temperature (RT) to assess drinking behavior would be a useful tool to evaluate animal health and well-being. Our objective was to compare the performance of a temperature threshold and an artificial neural network (ANN) for identifying drinking bouts. Four commercial dairy farms were monitored from June through September, 2019. Thirty high producing (52 � 8 kg/d) focal animals were selected on each farm and administered a RT bolus. Five days of data were selected randomly for each cow, yielding 150 cow-days for each farm that were manually labeled for drinking bouts. Two algorithms were experimented with in R 3.6.2. The temperature threshold algorithm consisted of applying a threshold at a specified quantile of RT for each cow-day, with any point falling below the threshold being classified as a drinking bout. Alternatively, 10 ANN each with a single hidden layer containing 20 units used RT observations over a 70-min time window as predictors, with the label (bout vs. non-bout) for the window being the label assigned to the time point at its center. Both algorithms were evaluated by 3 metrics: (1) area under the receiver operating characteristic curve (AUROC) using time point as the unit of observation; (2) mean absolute error (MAE) of predicted number of bouts using cow-day as the unit of observation; and (3) MAE of predicted mean interbout duration, also using cow-day as the unit of observation. Of 150 cow-days for each farm, approximately 60% were used for training and approximately 40% for validation. ANN outperformed the temperature threshold on all 3 metrics and for all 4 farms (Table 1). For mean interbout duration in particular, ANN provide enough sensitivity (MAE <4 min) to monitor even mild behavioral responses to changes in management and environmental conditions.Table 1.
Farm | Algorithm | AUC | MAE, n bouts | MAE, interbout duration (min) | |
A | Temperature | 0.983 | 0.51 | 11.30 | |
ANN | >0.999 | 0.08 | 1.52 | ||
B | Temperature | 0.986 | 0.47 | 9.69 | |
ANN | >0.999 | 0.12 | 3.28 | ||
C | Temperature | 0.987 | 0.36 | 7.62 | |
ANN | 0.998 | 0.14 | 2.10 | ||
D | Temperature | 0.988 | 0.47 | 10.94 | |
ANN | >0.999 | 0.02 | 0.32 |
Keywords: drinking behavior, reticular temperature, neural networks.