Predicting dry matter intake in beef cattle

Abstract

Technology that facilitates estimations of individual animal dry matter intake (DMI) rates in group-housed settings will improve production and management efficiencies. Estimating DMI in pasture settings or facilities where feed intake cannot be monitored may benefit from predictive algorithms that use other variables as proxies. This study examined the relationships between DMI, animal performance, and environmental variables. Here we determined whether a machine learning approach can predict DMI from measured water intake variables, age, sex, full body weight, and average daily gain (ADG). Two hundred and five animals were studied in a drylot setting (152 bulls for 88 d and 53 steers for 50 d). Collected data included daily DMI, water intake, daily predicted full body weights, and ADG using In-Pen-Weighing Positions and Feed Intake Nodes. After exclusion of 26 bulls of low-frequency breeds and one severe (>3 standard deviations) outlier, the final number of animals used for modeling was 178 (125 bulls, 53 steers). Climate data were recorded at 30-min intervals throughout the study period. Random Forest Regression (RFR) and Repeated Measures Random Forest (RMRF) were used as machine learning approaches to develop a predictive algorithm. Repeated Measures ANOVA (RMANOVA) was used as the traditional approach. Using the RMRF method, an algorithm was constructed that predicts an animal’s DMI within 0.75 kg. Evaluation and refining of algorithms used to predict DMI in drylot by adding more representative data will allow for future extrapolation to controlled small plot grazing and, ultimately, more extensive group field settings.

We describe here a use of advanced technology to measure an animal’s individual dry matter intake without the need to measure actual dry matter consumption. Such an approach could lead to methods that could be applied to grazing cattle, where 96% of the global population is found.

Introduction

Animal agriculture is well suited to help meet global protein needs, mitigate climate change, and improve local ecologies (Gerber et al., 2013). However, animal agriculture’s environmental shortcomings are well-documented and include excessive land and water use for feedstuffs (Capper et al., 2014). Selection for feed efficiency, for example, residual feed intake (RFI) which addresses the efficiency of maintenance requirements (Koch, et al., 1963), can improve the overall efficiency of beef herds and therefore reduce land requirements, water use, feed inputs, and producer input costs (Koch et al., 1963; Kolath et al., 2007; Basarab et al., 2013). Residual feed intake may be the best measure of efficiency due to its moderate heritability and mathematical independence from phenotypic traits used to predict dry matter intake (DMI; Berry and Crowley, 2013; Kenny et al., 2018; Freetly et al., 2020). On these bases, the industry could have long-term increases in beef production while reducing its per capita ecological footprint via selection for RFI as a part of the overall selection program and proper management (Herd et al., 2003; Nkrumah et al., 2006; Hegarty et al., 2007; Bezerra et al., 2013). In fact, Klopatek and Oltjen, 2022 demonstrated that from 1991 to 2019, the U.S. beef herd produced 10% more boneless beef while shrinking the associated water footprint by 29%, partly due in part to increases in animal efficiency. Currently, the collection of individual efficiency data is limited to animals fed in controlled dry lot environments where individual animal feed intake can be measured by delivering feed to automated intake monitoring bunks. However, most beef cattle, including the cow herd, are maintained in nonconfined pasture settings (Sheaffer et al., 2009) and many of those in confinement are not in systems with individual feed intake monitoring systems. Measuring individual animal feed intake in nonconfined pasture settings is highly challenging and confined to relatively small plots (Walker, 1995; Cottle, 2013; Greenwood et al., 2014). Predicting DMI using proxies and machine learning (ML) approaches will allow for the development of management strategies based on individual animal feed and water intake (Hubbart et al., 2023). The use of big data and ML to address questions of importance to modern agricultural production systems is becoming more prevalent. However, compared to cropping systems or animal health, ML applications in animal production have been limited (Liakos et al., 2018; Ellis et al., 2020; Benos et al., 2021). In this work, we pilot the use of ML to predict individual beef cattle DMI in confinement from proposed DMI proxies such as water intake, drinking behavior, animal performance, and climatic conditions to lay the foundations for future ML work predicting DMI of animals in systems without individual feed intake monitoring equipment and in pasture systems, thereby creating avenues for ensuring the sustainability of animal agriculture (Herd et al., 2002; Brew et al., 2011; Chang-Fung-Martel et al., 2021).

Materials and Methods

Animal care and use committee statement

The animal work described herein was approved by the West Virginia University Animal Care and Use Committee as protocol 1608003693. The West Virginia University Animal Care and Use Committee uses the Guide for the Care and Use of Agricultural Animals for its Agricultural Animal Program.

Animal breeds and sourcing

Animals used for this study were housed at West Virginia University Reymann Memorial Farm (Wardensville, WV). The animals consisted of two groups, one of bulls that were part of a performance test and one of steers being evaluated for performance as part of a separate research project. Bulls examined in this study originated from 22 farms across West Virginia and Virginia. Bulls (152) entered the test with a breed composition: 126 Angus, 3 Charolais, 10 Hereford, 10 SimAngus, and 3 Simmental. Upon arrival, bulls were treated with Safe-Guard (Merck Animal Health, Madison, NJ, US), a de-wormer, a dosage of Draxxin (Zoetis, Parsippany-Troy Hills, New Jersey, US) for prevention of bovine respiratory disease, an intranasal dose of pasteurella vaccine (Bovilis Once PMH, Merck Animal Health, Omaha, Nebraska) chute weighed, and ear tagged as per Wardensville Bull Test guidelines. Steers (n = 53) were sourced from the Reymann Memorial Farm herd (Angus) and the West Virginia Department of Agriculture herd (crossbred). Mean weight of all 152 bulls on day 0 was 351 ± 3 kg, and of 53 steers 262 ± 4 kg. Mean age of bulls and steers upon arrival was 282.54 ± 1 and 370 ± 3 d, respectively. After exclusions of 26 bulls of low-frequency breeds and one severe (<3 standard deviations, Angus bull) outlier, the final number of animals used for statistical modeling was 178 (125 bulls, 53 steers).

Collection of intake data

Feed Intake Nodes (GrowSafe 8000) were used to collect feed intake data during the tests. Feed Intake Nodes collect feed data via a feed bunk on load cells. To access the feed bunk, an animal must pass its head through a gate that limits access to one animal at a time and an antenna in the lip of the bunk reads an animal’s radio frequency identification (RFID) tag. In-Pen-Weighing (IPW) Positions (GrowSafe Beef) were used to weigh animals daily (Wells et al., 2021). The IPW Positions consist of a weighing platform positioned in front of the water trough, designed to collect partial body weights (BWs) paired with an antenna to read the animal’s RFID tag. Each time an animal visits the water trough, its unique RFID number and partial BW are recorded. Full BWs, used as the BW variable, are calculated via imputation from front-end weight collected when the animal is in the IPW with a resolution of 30 g. The accuracy of full BW data based on IPW-collected front-end weight has been validated by Wells et al (2021). Ritchie Ecofount waterers were used as attractants, equipped with custom flow meters (JLC International, Inc, New Britain, PA, USA) to collect water intake data during the test duration. Water was continuously available as cattle accessed the IPW Positions. In cases where an individual animal weight at an IPW Position was not recorded BW was imputed using prediction from linear regression of day and weight, as the RMRF cannot post-process missing data. This occurred one to three times in all 53 steers for a total of 57 occurrences, 0.45% of all 12,056 observation rows. The study occurred within a drylot facility containing five double pens measuring 29 m × 51 m (Figure 1A). Each pen included six Feed Intake Nodes and two IPW Positions (Figure 1B). Bulls were grouped based on consignor group, size, and breed.

Figure 1.

(A) Image of drylot facility with 5 pens (pairs of pens are managed together) and (B) close up of in-pen-weighing station with metered water units.

The first performance evaluation period (P1) consisted of bulls that were on-test for 88 d (25 November 2019 to 20 February 2020), and the second evaluation period (P2) consisted of steers that were on-test for 50 d (19 March 2020 to 7 May 2020), each following a 14-d diet acclimation period. Bulls were fed a total mixed ration ad libitum. The diet was formulated for a target ADG of 1.47 kg/d using standards for growing bulls (NASEM, 2016) and consisted of 65% corn silage, 22.5% supplement (cotton seed hulls, dried distillers grains, soybean hulls, soybean meal, peanut hulls, and wheat middlings), 12.5% mixed grass hay (2.5 to 5 cm particle length), and a commercial vitamin and mineral mix containing selenium. Rumensin and Tylan were added at the labeled dosage rate. The ration contained 13% crude protein (CP) and 68% total digestible nutrients (TDN) on dry matter (DM = 52%) basis. Calculated net energy for maintenance (NE_m) and net energy for gain (NE_g) were 6.53 and 4.06 MJ/kg, respectively. Steers were fed a total mixed ration ad libitum. The diet consisted of 80% corn silage, 14.5% supplement, and 5.5% mixed grass hay (2.5 to 5 cm particle length), including a commercial vitamin and mineral mix containing selenium. Rumensin and Tylan were added at the labeled dosage rate. The ration contained 11% CP on a dry matter (DM = 44%) basis. Calculated NE_m and NE_g were 6.99 and 4.51 MJ/kg, respectively. Ration samples were collected into quart bags from the feed truck after mixing and prior to dispensing into the feed bunks as per Cumberland Valley Analytical sample submission guidelines. Collected samples were refrigerated and overnight shipped to Cumberland Valley Analytical Services (Waynesboro, PA, US). Nutrient content of the ration was analyzed via Near Infrared Reflectance spectroscopy. Supplementary white salt was provided ad libitum in each pen during the data collection window. Bulls were weighed using a conventional livestock scale every 2 wk while on the test. Data were collected from 125 Black Angus bulls and 53 crossbred steers.

Collection of climate data

Climate data were recorded at 30-min intervals throughout the study (P1 and P2). A single representative climate station located approximately 275 m southwesterly from drylot pens (coordinates: 39°6’12.73”N, 78°35’8.19”W) was appropriate for this study given the relatively flat and open landscape (i.e., no local orographic weather influence). Air temperature and humidity were measured with a Campbell Scientific EE181-L probe (error ± 0.2 °C; Campbell Scientific, Inc.; Logan, Utah, USA). Precipitation totals were recorded with a Texas Electronics TE525MM-L tipping bucket rain gauge (error 1%, up to 50 mm/h; Texas Electronics; Dallas, Texas USA). Wind speed was measured using a Met One 034B Wind Set instrument, and a Campbell Scientific NR01 Four-Component Net Radiation Sensor was used to monitor radiation.

Animal and climate data were assembled using R 4.1.1, RStudio 2022.07.2 + 576 “tidyverse” ecosystem of packages, including the “dplyr” package, version 1.0.7 (Wickham et al. 2019, 2021; RStudio Team, 2020; R Core Team, 2021). The data were joined by their common field, “date”, such that each animal on test for a given date was assigned the climate values collected for that date.

Inputs for predictive algorithms

The data matrix with 12,056 observations (rows) and 20 variables (columns, including ID) had specified daily measurements of bulls (P1) and steers (P2). Specific predictor variables (x) included class (bulls, steers), BW (kg), age (d), ADG (kg), water intake (L), duration of IPW episodes (s), frequency of IPW episodes, test day, minimal (min) daily (d) air temperature (Ta) (°C), maximal (max) d T_a (°C), average (avg) d T_a (°C), min d relative humidity (Rh) (%), max d Rh (%), avg d Rh (%), min d wind speed (Wsp.) in (m/s), max d Wsp (m/s), avg d Wsp (m/s), W-wave radiation (MJ), and d total precipitation (PPT; mm). Ration DM was used to convert feed intake into DMI.

We used three analytical approaches: Random Forest Regression (RFR), Repeated Measures Random Forest (RMRF), and a mixed-effects Repeated Measures Analysis of Variance (RMANOVA). For all three approaches, the initial full algorithm included 18 continuous variables and one classification variable (class-either bull or steer). In addition, a core algorithm without any climate variables was constructed. Prior to analysis, the dataset (178 animals; N = 12,056 observations) was divided at random, based on animal ID, into a training subset (m, 70%, 125 beef [mix of bulls and steers]; m=8,443 observations) and an external validation subset composed of the remaining 30% (v, 53 beef; v = 3,613 observations) (Figure 2A). The predicted daily DMI and the averages for the entire testing period (animal means) were plotted in separate figures against the actual DMI on the external validation subset. This enabled assessing the three methods based on the smallest root mean squared error (RMSE) of each actual vs. predicted DMI. Both the full and core algorithms are presented.

Figure 2.

Approaches to data analysis (A) and Random Forest Regression (RFR) algorithm (B). The clean dataset including 178 animals and 12,056 observations was partitioned based on animals at random into the training subset (70%, m), on which the analyses were done, and the external validation subset (30%, v), on which the results from m were tested, validated and reported, using root mean square error (RMSE). The algorithm of RFR includes further partitioning of m—training subset—into bagging subset (B) and out-of-bag (OOB) subsample for the internal validation. On the subset B, a 500 (default) regression trees are implemented in steps 1 to 4. In a hypothetical regression tree at the step 3, six independent variables (x_n) are tested in series of simple regressions, and the “best” selected variable, for example, “a = the body weight at x_a = 400 kg” leads to a binary split in the data into two daughter nodes, based on the smallest sums of squares (SS) in DMI, such that the left daughter node (left circle) would have observations with lower DMI and the right node (right circle) observations with higher DMI, with the means represented in the circles. In those daughter nodes the same process is repeated, searching for the regressor that leads to the split with the lowest SS until the terminal node. The specific x variable, which was selected the most often as the best in each node across all 500 trees, would have the highest “importance”.

ML approaches

The RFR and RMRF analyses were built in R version 4.1.0. RFR was run using the “randomForests” package, and repeated k-fold cross-validation (CV) was conducted using the “caret” package. However, R has no “ready to go” package for RMRF. Thus, the code from Calhoun et al. (2020) was retrieved from the depository at GitHub.com (2022), deciphered, and adapted to our project. The code from Calhoun et al. (2020) makes use of the “rpart”, “partykit”, “geeM”, and “pROC” packages for R. The R script produced the RFR, RMRF, and the repeated k-fold CV in a single execution on West Virginia University’s Thorny Flat High Performance Computing cluster (HPC) and utilized parallel computing via the “parallel” R package.

Random forest regression

The workflow utilized to predict DMI is summarized in Figure 2B, using the terminology of Breiman (1984, 2001). In brief, the training set m was divided at random (bootstrapped), by observation, into a 70% learning subset or “bagged (B)” sample while the remaining 30%, the “out-of-bag (OOB)” sample, was set aside for internal validation. From within the continuum of one of the numbers of explanatory random variables (mtry), the value that most successfully separated the high response (DMI) values from the low (Breiman et al., 1984), based on least-squares bivariate fit, is chosen. This “split” divided the animals (observations) into the first “tree branches” or “child nodes” based on their weight. These groups then served as the new starting subgroups of observations and the processes of recursive partitioning, regressions, and mean square error (MSE) calculations until the terminal node. To use this tree as a predictor, data from a new animal (all 18 variables) are “dropped down the tree” until it rests in a terminal node. Each tree’s predicted values of the DMI are averaged across the forest number of trees (ntree) to obtain the RFR prediction of DMI (Breiman et al., 1984). The remaining 30% of the subset m was then used to calculate an unbiased prediction error (OOBerror). This measure serves as an internal validation and is used to test the predictive algorithm’s accuracy and calculate variable importance (Xu et al., 2019). At the end of the process, variables were ranked based on their significance in the entire forest, with the percent increase in the MSE of each variable enumerated. The trained forest was then used to generate predicted values of the response variable (DMI) for the observations in the v dataset as external validation.

RFR algorithms’ hyperparameters (ntree, mtry) were optimized using Repeated k-Fold CV, demonstrated in Figure 3. The ntree is the number of trees generated in the random forest, and mtry is the number of explanatory variables used to find the best split at each node of each tree in the forest. The Repeated k-Fold CV procedure divides the training data set m into k nonoverlapping folds. Then, one of the k-folds is randomly selected to serve as a validation set, while the other folds are used as ‘training’ data. In all, k algorithms are fit to the training data and evaluated on the k validation sets. This procedure is repeated multiple times, and the average performance for each unique set of hyperparameters across all folds and all repeats is reported. The current study included three repeats and k = 10. Thus, 30 distinct validation sets (3-repeats * 10-fold) were used to estimate the RFR’s performance for each unique combination of the 19 possible levels of mtry and four levels of ntree.

Figure 3.

Optimization of Random Forest Regression. Mtry represents number of variables (19) being screened at each split in the tree for the forest and ntree is number of trees that could be used in the RFR. The colored cables of the parallelogram correspond to 76 combinations of 19 mtry (1 to 19) and 4 ntrees levels (500 blue, 1,000 red, 1,500 green, and 2,000 fuchsia) we explored using the repeated k-fold cross-validation that generated the measures of accuracy, such as root mean square error (RMSE), Rsquared and additional four (not shown) to aid in decision for optimal combination of mtry and ntree. For this particular algorithm (full) the 1,500 ntree at 15 mtry reached the asymptote in minimum RMSE, meaning that RFR with these parameter settings will provide desired accuracy.

Repeated measures random forest

Repeated measures random forests is a variation of RFR, specifically applicable to data where observations are repeated measurements on the same experimental unit, as in our experimental approach. The main algorithm is the same for both procedures; the main difference is that the “bagging” (bootstrap aggregating) in RFR is based on a random sample of all observations, while in RMRF, the bagging is based on the subjects (animal ID) (Calhoun et al., 2020). This approach facilitates more control of inner correlations within each animal.

Repeated measures analysis of variance

The RMANOVA proceeded in two steps. First, a suitable covariance structure [autoregressive of the first degree; AR(1)] was selected for the within-subject (within animal) approach. Then the entire algorithm was fit using the AR(1) structure in a marginal probability approach. The RMANOVA was evaluated using JMP software (JMP, Version Pro 16.0.0, SAS Institute Inc., Cary, NC, Copyright 2021). Specifically, fit model platform was selected, with mixed model personality, (equivalent to PROC MIXED of SAS), casting the ADG (kg), water intake (L), daily duration of IPW episodes (s), daily frequency of IPW episodes, full BW (kg) and class (bull or steer) as fixed effects; all of the climate variables (T_a, Rh, Wsp, W-wave radiation, and PPT) as the random effects (covariates). The repeated structure factor was time (test day), the subject was the animal ID, and repeated covariance structure option was AR(1). The significance criterion alpha for all tests was 0.05.

Results

The 178 animals used during the study periods and analysis are described in Table 1, with bulls and steers partitioned into the training and testing groups, 70% and 30%, respectively. The bulls (n = 125) began the test at 238 ± 1 d-of-age with an average BW of 353 ± 4 kg and gained an average of 1.65 kg/d. The steers (n = 53) began their test at 371 ± 3 d-of-age with an average BW of 259 ± 4 kg and gained an average of 1.73 kg/d.

Table 1.

Descriptive details of the animals used in the current evaluation. Bulls were evaluated from 25 November 2019 to 20 February 2020 and steers from 19 March 2020 to 7 May 2020

Class	Statistic	DMI, kg		ADG, kg		Water Intake, L
		Train	Test	Train	Test	Train	Test
Bull	Mean	8.92	9.10	1.65	1.65	20.13	20.67
	Max	17.41	16.43	2.43	2.10	54.56	45.71
	Min	1.01	2.54	0.96	1.21	0.19	3.44
	Median	8.88	9.08	1.63	1.65	19.77	20.37
	SD	1.79	1.68	0.28	0.23	5.71	5.74
Steer	Mean	8.31	8.60	1.72	1.77	22.77	21.74
	Max	13.78	14.76	2.01	2.19	51.40	47.04
	Min	3.34	3.16	1.37	1.49	2.62	4.32
	Median	8.32	8.53	1.77	1.73	22.64	21.55
	SD	1.56	1.58	0.18	0.20	6.22	6.13

Climate during both study periods was variable (Table 2). The average air temperature during P1 was 3.0 °C, while the average air temperature during P2 was 10.2 °C. The total precipitation for P1 was 178.9 mm, and the total precipitation for P2 was 168.6 mm. Average relative humidity and wind speed were 76.1% and 1.2 m/s for P1 and 72.9 and 1.4 m/s for P2. Solar radiation (shortwave) was 6.0 and 14.7 MJ/m² for P1 and P2, respectively (Table 2). The end of trial predicted total BWs were highly correlated with the end of trial chute weights (r² = 0.983).

Table 2.

Descriptive daily environmental statistics for the period from 25 November 2019 to 20 February 2020 (bulls) and From 19 March 2020 to 7 May 2020 (steers)

Class	Statistic	Rh, %			T a, °C			USP, m/s			W-wave Rad, Mj/m2
		Avg.	Max	Min	Avg	Max	Min	Avg	Max	Min
Bulls	Mean	76.1	96.7	47.6	3.0	9.7	−2.8	1.2	5.0	0.0	6.0
	Max	99.8	100.0	96.9	17.5	22.3	13.3	2.8	10.0	0.0	14.3
	Min	49.6	70.6	18.9	−7.0	−0.9	−13.0	0.3	1.7	0.0	0.4
	Median	75.1	100.0	44.8	3.0	9.2	−2.3	1.1	4.7	0.0	6.0
	SD	12.9	6.0	17.4	4.8	5.7	5.0	0.6	2.0	0.0	3.4
	Total ppt, mm = 178.9
Steers	Mean	72.9	97.6	41.8	10.2	17.2	3.7	1.4	5.8	0.0	14.7
	Max	97.0	100.0	88.8	18.8	30.2	11.7	2.7	11.7	0.3	29.0
	Min	37.7	59.8	15.2	4.6	7.3	−4.2	0.5	3.3	0.0	4.5
	Median	73.8	100.0	35.8	9.3	17.0	3.6	1.3	5.2	0.0	14.0
	SD	13.9	7.1	17.3	3.8	5.4	4.1	0.5	1.9	0.1	7.1
	Total ppt, mm = 168.6

Based on the Repeated k-Fold CV, we determined that 1,500 was the optimal ntree developed in RFR, with 15 mtry for the full algorithm, while 2000 ntree and 5 mtry for the core algorithm. Further increase of ntree did not remarkably decrease the OOBerror. The summary of all ntree and mtry and the performance of RF methods (MSE and R²) is in Table 3. As indicated in Table 3, the repeated approaches resulted in smaller MSE and larger pseudo-coefficient of determination (R²). The smallest of all MSE (0.98 kg) while the largest R² (0.68) were achieved with the Core RMRF algorithm, comprised of BW (kg), water intake (L), test days, age (days), ADG (kg), class (bull, steer), duration of drinking (s), and frequency of drinking.

Table 3.

Optimal hyperparameters of RF variants and performance for Full and Core Algorithms using DMI in kg, generated on the OOB (internal validation) subsets

Predictive algorithmic approach	mtry	ntree	MSE	R 2
Full RFR	15	1,500	1.71	0.45
Core RFR	5	2,000	1.73	0.44
Full RMRF	1	2,000	1.09	0.65
Core RMRF	1	2,000	0.98	0.68

Figures 4 and 5 list the importance of variables based on the full algorithm RFR and RMRF, ranked from the strongest to the weakest predictor. It demonstrated that both procedures selected similar, but not identical, variables as important. The “importance” is defined as the variable with the highest percentage increase in MSE (% IncMSE) when that variable was not included in a given tree. In both procedures, the BW, ADG, and water intake (kg), in addition, to test day and age, were amongst the most influential variables on DMI. Figures 4 and 5 demonstrate that the RMANOVA, based on the effect P-values, had a similar ranking of the variable importance as the RF procedures, with BW (kg), ADG, water intake (L), and duration of drinking (s) significant in both full and core RMANOVA approaches.

Figure 4.

Variable importance in the full algorithms from Random Forests Regression (RFR), colors - overlayed with P-values from repeated measures ANOVA (RMANOVA). Variables used in the full algorithm on combined datasets of bulls and steers were ranked in RFR by the percent increase in mean squared error (% IncMSE) represented by the length of each bar. It corresponds to how much would total MSE increase if the variable would be eliminated from the OOB (out-of-bag) observations within each approach. The more the prediction error increases, the more important the variable is. The % IncMSE is computed for every tree, averaged, and divided by the variability over the entire ensemble (Xu et al., 2019). The colors of the bars relate to the P-value of fixed effects in the RMANOVA according to the legend.

Figure 5.

Variable importance in the full algorithms from Repeated Measures Random Forest (RMRF), colors – overlayed with P-values from repeated measures ANOVA (RMANOVA). Variables used in the full algorithm on combined datasets of bulls and steers were ranked by the percent increase in mean squared error (% IncMSE) represented by the length of each bar. It corresponds to how much would total MSE increase if the variable would be eliminated from the OOB (out-of-bag) observations within each approach. The more the prediction error increases, the more important the variable is. The % IncMSE is computed for every tree, averaged, and divided by the variability over the entire ensemble (Xu et al., 2019). The colors of the bars relate to the p-value of fixed effects in the RMANOVA according to the legend. In the RMRF variables were more uniformly important based on the % IncMSE than in RFR (see Figure 4), with similar high importance variables significant in RMANOVA, such as the body weight, ADG, daily water intake as well as duration of in-pen-weighing (IPW) episodes. Variables that ranked differentially between RFR and RMRF were the duration and frequency of IPW episodes. Overall, a similar subset of variables was determined important with the repeated methods (RMRF and RMANOVA).

In the full approach, RMANOVA, the important variables for DMI were BW, ADG, and daily water intake ­(Figures 4 and 5), followed by the duration of IPW episodes (P < 0.05). The RMANOVA prediction equation from the core approach is

$${\displaystyle \begin{array}{l}{\displaystyle \begin{array}{llll}& {\displaystyle \mathrm{D}\mathrm{M}\mathrm{I}=2.1340+1.1128\times \mathrm{A}\mathrm{D}\mathrm{G}\left[kg\right]+0.0346\times \mathrm{W}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{I}\mathrm{n}\mathrm{t}\mathrm{a}\mathrm{k}\mathrm{e}\left[L\right]}& {\displaystyle +0.0002\times \mathrm{D}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{I}\mathrm{P}\mathrm{W}\mathrm{e}\mathrm{p}\mathrm{i}\mathrm{s}\mathrm{o}\mathrm{d}\mathrm{e}\left[\mathrm{s}\right]+0.01022\times \mathrm{F}\mathrm{u}\mathrm{l}\mathrm{l}\mathrm{B}\mathrm{W}\left[\mathrm{k}\mathrm{g}\right]}& {\displaystyle +(0.18462\mathrm{i}\mathrm{f}\mathrm{a}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{l}\mathrm{w}\mathrm{a}\mathrm{s}\mathrm{a}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{r}-0.18462\mathrm{i}\mathrm{f}\mathrm{a}\mathrm{b}\mathrm{u}\mathrm{l}\mathrm{l}}\end{array}}\end{array}}$$

The measured DMI was plotted against the predicted DMI from RFR, RMRF, and RMANOVA (Figures 6 and 7) as an external validation and a comparison of methods on the 30% of animals (v). If the predicted DMI closely matched the actual DMI, the data points would fall near the diagonal line through the origin, and the RMSE would be minimal (Figures 6 and 7). Figure 6 demonstrates that in the Full algorithm, the predictions using the RMRF had the smallest predicted error (RMSE 1.43 daily DMI). In the core algorithms, RFR and RMANOVA generated very similar errors of ­prediction (1.55 and 1.56 kg, respectively), with RMRF having the best RMSE (1.51 kg). When the actual and predicted DMI from the full algorithm were averaged across all testing days for the v subset, and each animal was represented by one datapoint (Figure 7) RMSE of RMANOVA was 0.82 kg, of RFR 0.83 kg, and the RMRF outperformed both again with 0.75 kg.

Figure 6.

Plots of actual vs. predicted daily DMI using Random Forests Regression (RFR), Repeated Measures (RM) Random Forest (RMRF) and RMANOVA utilizing Full and Core algorithms. The line through the origin represents the perfect DMI prediction. Residuals (RMSE) characterize the spread of the data from the diagonal line. Full algorithms (left panels) included all 18 continuous and one classification variable, while the core algorithms (right) were without any climate variable. Actual DMI and predictions from daily values with corresponding RMSE represent remarkable similarity in predictive precision, with the repeated procedures (RMRF and RMANOVA) demonstrating the smallest predictive errors.

Figure 7.

Plots of actual vs. predicted DMI for individual animals averaged across an entire testing period using Random Forests Regression (RFR), Repeated Measures (RM) Random Forest (RMRF) and RMANOVA utilizing Full and Core algorithms. The line through the origin represents the perfect DMI prediction. Residuals (RMSE) characterize the spread of the data from the diagonal line. Full algorithms (left panels) included all 18 continuous and one classification variable, while the core algorithms (right) were without any climate variable. Averaged DMI and predictions from individual animals across entire testing periods with corresponding RMSE represent remarkable similarity in predictive precision, with the repeated procedures (RMRF and RMANOVA) demonstrating the smallest predictive errors.

Interestingly, for the core algorithm, RMANOVA had the smallest RMSE (0.76 kg) compared to RFR and RMFR (0.83 and 0.79 kg RMSE). The ability to better predict average dry DMI across the intake period is partly due to the variation in day-to-day DMI for any individual (Figure 8). In summary, RFR and RMRF had similar results to the mixed effects RMANOVA when the same group of variables was used for these approaches. The RMRF produced the smallest (the best) prediction error for the Full daily algorithm, including all climate variables, indicating that repeated component is the most suitable approach when using RF for this type of study. Use of the RMRF ML full algorithm allowed for the prediction of individual cattle DMI across the test period within 0.75 kg of the actual DMI. Notably, the range in actual DMI was much greater than the daily predicted DMI, likely due to the observed daily fluctuation in actual DMI and water intake (Figure 8).

Figure 8.

Representative data from an individual animal for the 88 d of DMI demonstrating the fluctuation in individual daily DMI and water intake.

Discussion

By using three ML approaches, RMANOVA, RFR, and RMRF, we were able to build a predictive algorithm that can predict individual cattle DMI across the test period within 0.75 kg of the actual DMI. While the best prediction accuracy resulted from averages of predictions based on the RMRF full algorithm, all three ML approaches differed in ranking predictive variables. Full algorithm Random Forest Regression predicted individual cattle DMI across the test period within 0.83 kg of the actual DMI and ranked algorithmic variables in order of predictive importance as BW, ADG, test day, age, and water intake. Of those variables, only BW, ADG, and water intake were significant factors, according to RMANOVA. The duration of IPW episodes was significant (P < 0.05) but was ranked 10th in predictive performance by the RFR algorithm. Full algorithm RFR performance was lower when daily DMI values were used instead of test mean values and allowed for the prediction of DMI within 1.5 kg of the actual DMI. The full and core RFR algorithms also demonstrated differences in predictive accuracy. When using test mean values, the performance of the RFR core algorithm was identical to the full (RMSE = 0.83 kg). However, the RFR full algorithm’s performance was better than the RFR core algorithm when daily DMI values were used (RMSE = 1.5 and 1.55 kg, respectively). Regardless, improvements in the prediction of 0.05 kg between full and core algorithms may not be relevant given the advantage in reduced processing requirements of the core algorithm.

Repeated Measures Random Forest generated accurate predictions of DMI while accounting for the lack of independence of data within individual animal records. Reranking of the importance of predictive variables occurred between RFR and RMRF algorithms, but important variables remained largely the same between the two algorithms. Notably, the predictive significance of water intake greatly increased in the RMRF algorithm, with water intake and full BW sharing nearly identical—and primary—predictive importance. Duration of drinking events also greatly increased in the ranking of predictive importance between the RFR and RMRF algorithms. Daily frequency of water visits and class (bull or steer) were ranked within the 10 most important variables but were insignificant. Like the RFR algorithm, the RMRF algorithm differed in performance between the full and core algorithm and daily DMI values and test mean values predictions. The test mean values generated from the full RMRF algorithm provided the smallest error of all six algorithms (RMSE = 0.75 kg), with the core algorithm performing marginally worse (RMSE = 0.79 kg) than the full algorithm. The performance of the RMRF was lower when using daily DMI values, with the full and core algorithms having RMSEs of 1.43 and 1.51 kg, respectively. It is unclear if the marginal improvement in algorithmic prediction is of higher value than the reduction in processing requirements allowed by using the core algorithm.

Of all the algorithms, outputs of the RMANOVA showed the most variation, with the RMANOVA core approach of the test means performing only marginally worse (RMSE = 0.76 kg) than the RMRF full algorithm on test means. The RMANOVA full algorithm using the daily DMI values performed the poorest of all algorithmic iterations (RMSE = 1.58). Unlike RMRF, the core RMANOVA approaches performed better than the full RMANOVA approaches for both test means and daily DMI values. The chief benefit of using RMANOVA for the prediction of DMI is the development of an equation from which coefficients and relationships among variables can be elucidated.

Previous research resulted in approaches to predict the DMIs of beef cattle (Minson and McDonald, 1987; Rook et al., 1990). However, many of these lacked precision and consistently suffered from issues with overprediction of feed intakes of livestock (Liakos et al., 2018). Traditional, standardized approaches to predicting feed and water intake in cattle do exist (NASEM, 2016). However, these equations have been reported to unreliably predict DMI in calves and yearling steers (McMeniman et al., 2009 ; Patterson et al., 2000; Block et al., 2001). Additionally, these models require the calculation and use of several adjustments for environmental factors, diet composition, forage availability, use of growth implants and additives, animal age, and class. Some of these equations can be used to predict forage intake in the pasture; however, the committee of the National Academies of Science Engineering and Medicine caution against the use of an NDF-based method or guidelines published by Lalman (2004) for the prediction of forage intake in beef cattle as those methods have not been thoroughly evaluated (NASEM, 2016). A growing number of authors have used ML techniques to predict DMI; however, few have used repeated measures in ML techniques and most lack the ability to be expanded into grazing systems. In the effort most similar to our approach, Davison et al. (2021) predicted DMI from eating behavior variables from animal-mounted accelerometers in 80 Limousin–Aberdeen Angus steer crosses, employing random forest (RF) and support vector regressor (SVR). Their results in RF were comparable to ours, with their RMSE of 1.61 and 1.45 kg DMI for concentrated and mixed ration, respectively, while our full algorithm RMSE yielded 1.5 and 1.43 kg DMI for RFR and RMRF, respectively. The author argued that RMSE represents about 10% of the average DMI and such error seems not satisfyingly accurate (Davison et al., 2021). Input variables in the Davidson et al. (2021) RF model, except for age and mean daily temperature, were confined to feeding behavior-based variables, limiting the expansion of their approach into other classes of livestock (bull and heifer) and into grazing systems, where collection of feeding behavior data is extremely limited. Achour et al., 2020 used Support Vector Regression (SVR) coupled with a Convolutional Neural Network (CNN) to identify feeding behavior in individual dairy cows. Identifying individual animals and feeding behavior was possible, though estimates of individual feed intakes were not derived. Other neural-network efforts have been similarly behavior-based (Peng et al., 2019; Shen et al., 2022). Saar et al. (2022) expanded the prediction of individual cow intake via use of computer vision and ML approaches. Using Transfer Learning models based on CNN’s paired with a real-time computer vision system, Saar et al., (2022) were able to accurately predict (RMSE = 0.19 kg) individual feed intake of dairy cattle using multiple types of confinement-fed feed. Though their prediction is highly accurate relative to other CNN-based approaches, it is not readily scalable to grazing systems. Williams et al. (2019) focused on predicting DMI in grazing lactating beef cows, using 94 variables of body measurements, linear type scoring, grazing behavior from and thermal imaging in a multivariable regression. Width at pins, full body depth, ruminating mastications, central ligament, and rump width score, were retained in the approach in addition to milk yield, BW, parity, calving day, and maternal origin, with the predictability of DMI with R² of 0.68 in training and 0.59 in a testing herd. However, the wearable technology was only recorded for two periods of 24 h. In these studies, no consideration is given to individual water intake. Our results and those outside the ones presented in this manuscript, share the common challenge of highly-varied animal DMIs. Therefore, it would be valuable to get the daily measurements that can be employed to predict DMI of individual beef cattle. The use of ML in agricultural technologies and research has increased. However, much of ML’s use in agriculture is relegated to cropping systems (Liakos et al., 2018). Those ML-based predictive approaches designed for animal agriculture are widely used in animal welfare, genomics, and the prediction of BW (Dutta et al., 2015; Weihao et al., 2021; Hakem et al., 2022). ML, particularly RMRF, allows for close prediction of beef cattle DMI at the scale of the individual animal, while considering the repeated measurements data structure (Calhoun et al., 2021). Animals in this study were fed in confinement and had feed measured during every feeding event. However, applications of this algorithm to livestock regimes in which measurement of feed is not feasible, such as a feedlot or pasture, present new avenues of innovation for both beef cattle producers and researchers. To expand our work into pasture-based systems, in which ~96% of livestock will spend most of their lifecycle, ground-validated forage intake data is needed to build a predictive ML algorithm (Sheaffer et al., 2009). Currently, the exceedingly high labor and financial costs associated with grazing research limit the development of an ML algorithm that can predict the DMI of pasture-grazed animals (Ramoelo et al., 2014).

Limitations of our current approach exist, and improvements must be made for the tool to benefit producers. Variables were ranked in order of predictive importance within our two RF approaches (RFR and RMRF) (Figure 4). Climate-based variables ranked unimportant in the RMRF algorithm and nearly negligible in the RFR algorithm. It is known that livestock intakes are heavily affected by climate variables (Gates, 1980; Yousef, 1985; Mader, 2003). We believe that the low-importance ranking of climate variables is due to the test period (25 November 2019 to 20 February 2020 for bulls and 19 March 2020 to 7 May 2020 for steers). We expect that water and DMI data collected during the summer months, when temperature and humidity are likely more significant limiting factors (Brosh et al., 1998), will be more impacted by climatic conditions. Additionally, full BW, which was ranked as a highly important predictive factor in both the RFR and RMRF algorithms, drives maintenance metabolic demand and has historically been viewed as a primary predictor of DMI. Full BW is also a function of both water and feed intake. Interestingly, the range in daily DMI is much wider, regardless of ML method or full versus core algorithm, than in predicted daily values. This observation may not be surprising given the fluctuation observed in individual daily DMI and water intake (Figure 8).

As stated above, these intakes are modulated by climatic stress factors—particularly extremes of cold and heat. Incorporating data from summer DMI tests in which individual animal feed and water intakes are known is necessary to improve our algorithm. Within our system, daily full BW is an average of all daily weights collected during each drinking event, allowing us to monitor fluctuations in individual animal weight closely. Our data has made it clear that weight gain in livestock is not a linear function. Although ADG is the standard metric by which livestock growth is measured, ML approaches and high-resolution feed and water intake monitoring have allowed the exploration of more nuanced data, necessitating a more incremental standard for animal growth. Our approach was done using a relatively small population of animals (N = 178) consisting of a limited number of breeds (Angus bulls and crossbred steers) during cool months. Collection and algorithmic incorporation of data from more cattle of varying breeds, from all classes of cattle, and during a broader range of months, particularly from summer months, would improve algorithm accuracy and increase its applicational scope.

Conclusion

Using Random Forest-based ML approaches, we developed an algorithm that can predict individual cattle DMI across a test period within 0.75 kg of the actual DMI when using test mean DMI values. This approach allows for the measurement of individual DMI without the need for automated bunk technology and feed measurements. Opportunities to further improve the algorithm primarily lie within the collection and addition of more and more varied (e.g., more breeds, increased test seasonality, heifers) data and the incorporation of an incremental growth metric such as incremental daily gain. Long-term improvement of the algorithm would require the collection of feed and water intake data in different regions of the U.S. Improvement of the algorithm would increase its ability to be used by beef cattle producers in production systems where measurement of feed is not feasible. Perhaps the greatest potential of our algorithm lies in its expansion to the prediction of DMI intake in pastured animals, which account for ~96% of all ruminant livestock. While the use of the algorithm in grazing systems is currently limited by the extraordinary cost and effort of ground-validation data collection, our ML approaches to improving beef cattle production have already laid the groundwork for efficient improvement of beef management and production.

Glossary

Abbreviations:

ADG

average daily gain

B

bagged

BW

body weight

CNN

convolution neural network

CP

crude protein

CV

cross validation; d, daily

DM

dry matter

DMI

dry matter intake

RFR

random forest regression

RMRF

repeated measures random forest

RMANOVA

repeated measures analysis of variance

HPC

high performance computer

IPW

in-pen-weighing position

m

training data subset

max

maximum

min

minimal

ML

machine learning

MSE

mean square error

mtry

number of explanatory variables

NE_g

net energy for gain

NE_m

net energy for maintenance

ntree

number of trees

OOB

out-of-bag

OOBerror

unbiased prediction error

P1

period 1

P2

period 2

PPT

total precipitation

RFI

residual feed intake

RFID

radio frequency identification

RFR

random forest regression

Rh

relative humidity

RMANOVA

repeated measures ANOVA

RMRF

repeated measures random forest

SVR

support vector regression

T _a

air temperature

TDN

total digestible nutrients

v

validation data subset

Wsp

wind speed

x

predictor variables

Conflict of interest statement

The authors declare no conflict of interest.