302 Incremental Variables in Predicting Beef Cattle Dry Matter Intake

Abstract

Abstract The great strain of increasing climate variability on animal agriculture necessitates improvement of current production processes, the development of methods to measure individual feed and water intakes, and the improvement of feed and water use efficiency in cattle. The use of machine learning (ML) approaches in predicting beef cattle dry matter intake (DMI) have provided groundwork for improvement of beef cattle production using big data and ML. However, despite wide variation between daily feed intakes, typical variables in animal intake-based ML approaches are not incremental, not allowing for daily prediction of intake. Our work introduces first-differenced intake, climate, and animal performance variables into a ML algorithm to predict beef cattle DMI and increase the immediacy of prediction applications. A total of 745 animals were evaluated in eight test groups from 11/25/2019 to 9/2/2021 in a dry lot equipped with In-Pen Weighing Positions and Feed and Water-Intake Nodes. Relationships among average daily gain (ADG), DMI, residual feed intake (RFI), water intake (WI), residual water intake (RWI), animal performance variables, and environmental variables at the individual animal level were investigated on a first test group of 125 Angus bulls and 53 crossbred steers. Out-Of-Bag root mean square error (OOB RMSE), an internal measure of Random Forest (RF) prediction accuracy, was used as a measure of error between model-predicted and observed DMI. Input variables were first-differenced by differencing a variable’s value between sequential test days for the entire test period. Two models were run using the RF package in R (ntree = 500, mtry = 11), with the first model (M1) using a standard test-train data split and the second model (M2) using all first-differenced and non-incremental observations to train the model. The model incorporating the test-train split performed marginally better than M2 (OOB RMSE = 1.39 and 1.44, respectively), indicating that first-differenced input variables can be used to predict daily individual DMI by within 1.39 kg by using M1. Using percent-increase in MSE (%IncMSE), first-differenced variables were found to be of less overall predictive value than non-incremental variables, with incremental daily gain and first-differenced water intake being ~55 and 45% less predictive than ADG and overall water intake in both M1 and M2. Though first-differenced variables were less predictive than non-incremental variables, inclusion of first-differenced variables allowed for prediction of individual DMI by within 1.39 kg. Our work demonstrates critical preliminary steps in the development of a deployable algorithm for the efficient prediction of DMI and the improvement of beef cattle intake efficiencies for the continued sustainability of animal agriculture.