Technical note: validation of a system for monitoring individual behavior in beef heifers

Abstract

The objectives of the 2 studies conducted were to validate the accuracy of an automated monitoring device (AMD; HR-LDn tags, SCR Engineers Ltd., Netanya, Israel) for different types of behaviors or cow-states (side lying, resting, medium activity, high activity, rumination, grazing, walking, and panting) in beef heifers and to determine if the total time per cow-state recorded by the AMD corresponds to the total time per cow-state recorded by instantaneous observations. Cow-state is recorded every second and, within 1 min, the most prevalent cow-state is considered to be the behavior presented by the animal during that interval. Study personnel (n = 2) observed heifers (n = 10) for 20 min from 0800 to 1140 h and 10 min from 1500 to 1640 h during 4 consecutive days and recorded continuously each cow-state at started and ended. Thus, study personnel were able to determine within a 1-min interval, which cow-state was most prevalent and represented the heifer’s behavior. Because the proprietary machine learning algorithm prioritizes certain behaviors over others based on their contribution to the understanding of generalized bovine behavior patterns, we also determined the most prevalent behavior observed in 5-min intervals. Test characteristics (sensitivity, specificity, accuracy, and negative and positive predicted values) were calculated using the observer as the gold standard. In study 2, heifer behavior was scanned by observers (n = 2) every 5 min from 0800 to 1100 h and 1500 to 1800 h for 3 consecutive days. Total minutes per cow-state according to the observer were compared with the total minutes per cow-state according to the AMD during the same period to determine the correlation coefficient. In study 1, test characteristics were high (low ≤ 40%, moderate = 41 to 74%, high ≥ 75%) for rumination (≥ 89.7%), grazing (≥ 76.5%), and side lying (≥ 81.8%), and moderate for resting (≥ 48.8%). In study 2, the correlation coefficient for rumination (R² = 0.92) and grazing (R² = 0.90) were high and the correlation coefficient for resting (R² = 0.66) and walking (R² = 0.33) were moderate. We conclude that the AMD used in this study showed high accuracy when measuring rumination and grazing, but it was subpar when measuring resting and walking. The algorithms employed by the AMD used need to be improved for determination of walking and resting behaviors of beef cattle.

Introduction

Behavioral monitoring in cattle is a tool for diagnosing disease (Theurer et al., 2013). Visual observation may be used to monitor animals but it is labor intensive and subject to human error (Theurer et al., 2013). Automated monitoring devices (AMD) have been validated for measuring activity (Elischer et al., 2013; Molfino et al., 2017), rumination (Schirmann et al., 2009; Molfino et al., 2017), lying time (Swartz et al., 2016), respiration rate (Milan et al., 2016), and feeding (Borchers et al., 2016). Automated systems that measure several behaviors simultaneously are available and may prove to be a formidable tool for monitoring dairy and beef cattle.

The HR-LDn tags (SCR Engineers Ltd.) are composed of a neck mounted device consisting of an accelerometer sensor, a processing unit, and wireless communication functionality. On the basis of the continuous data sensed by the 3-axis accelerometer, machine learning algorithms determine the cow-state. Using an older version of this system (Hi-Tag, SCR Engineers Ltd.), Schirmann et al. (2009) demonstrated that the accuracy of rumination time recorded by the system was high when using lactating dairy cows. Other researchers using < 9-mo-old dairy heifers (Burfeind et al., 2011) and beef cattle (Goldhawk et al., 2013) observed only a moderate correlation between the AMD recorded and observed rumination time. Thus, additional studies are necessary to validate new technologies that measure more than rumination time, such as the HR-LDn tags.

The objectives of these studies were to evaluate 1) the test characteristics of the cow-state recorded minute-by-minute by an AMD compared with continuous visual observation and 2) whether the total time per cow-state is in agreement with the total time per cow-state according to instantaneous recording at 5-min intervals recorded by an AMD.

Materials and Methods

Animal care and handling were in accordance with the Guide for the Care and Use of Agricultural Animals in Research and Teaching (Federation of Animal Science Societies, 2010) and the studies were approved by the Institutional Animal Care and Use Committee (protocol #201709952). The studies were conducted at the North Florida Research Educational center in Marianna, FL. The studies were conducted during 7 consecutive days during the month of October, 2017. The beef heifers were housed in 6,000 m² pens (36 × 166 m) and had ad libitum access to grass (Paspalum notatum) and water during the entire study with no feed supplementation. The herbage mass during the study was 1,150 kg/ha. Nine days before the beginning of study 1, 10 Angus heifers, 21 to 23 mo of age, and average body weight 425.0 ± 15.0 kg were dye painted with roman numbers (I-X) on both lateral parts of the body to facilitate their recognition. On the same day of the dyeing, each heifer was fitted with a collar on the proximal third of the neck with an AMD (HR-LDn tags, SCR Engineers Ltd.) immediately behind the left ear.

Before the start of studies 1 and 2, observers (n = 2) were trained to reach a consensus about the beginning and end of each cow-state. The behaviors adopted to characterize each cow-state are reported in Table 1. In studies 1 and 2, the observers were located immediately outside the pen with an unobstructed view of the heifers. The observers were equipped with laptop computers and clocks that were synchronized with the internal clock of the computer in which the cow-states generated by the AMD were recorded. Independently, the observers recorded manually the cow-states in an Excel sheet (Microsoft Corp., Redmond, WA) using unique alphabetic characters per cow-state for all heifers in the studies 1 and 2. The level of agreement between the 2 independent observers was evaluated using the concordance correlation coefficient (CCC) of MedCalc (MedCalc software, Mariakerke, Belgium).

Table 1.

Descriptive ethogram of the observed cow-states by study personnel

Cow-state	Description
Side lying	Heifer is lying laterally with any side of the body touching the ground
Resting	Heifer can be either lying or standing and it is not moving, ruminating or grazing
Rumination	The onset of the activity is when the heifer starts to chew a regurgitated bolus and it ends when the bolus is swallowed
Walking	Heifer is moving within the pen at a regular pace, with 3 or more consecutive steps without stopping
Grazing/Eating	Heifer starts to actively search for food and it also includes biting or chewing pasture
Panting	Heifer presents heavy breathing

In study 1, 10 beef heifers were divided in 2 groups of 5 heifers each. On the first (day 1) and second (day 2) days of the validation, heifers I to V were observed and during day 3 and 4 of the validation, heifers VI to X were observed. The validation started at 0800 h with the first animal being observed for 10 min continuously. Every 10 min, the animal being observed would change according to the numerical order. Each animal was observed individually making it feasible to record the exact seconds when the cow-state began and ended. After the 5 heifers were observed once, the cycle of 5 observations restarted from animal number I (day 1 and 2) or VI (day 3 and 4) to complete 2 cycles of observation from the same 5 heifers in the morning between 0800 h and 1139 h. In the afternoon, the same cycle of observations was adopted but only carried out once from 1500 h to 1639 h. In total, each animal was observed for 60 min in 2 consecutive days.

The HR-LDn tags determine cow-state (side lying, resting, high and medium activity, rumination, grazing, eating, walking, and panting) every minute. The proprietary machine learning algorithm prioritizes certain behaviors over others based on their contribution to the understanding of generalized bovine behavior patterns. For research purposes, SCR Ltd. personnel, blinded to the observers’ evaluation, decoded the algorithm and sent the data to the investigators for analysis.

Medium and high activities were cow-states not clearly specified to the observers before the start of the study; thus, observers only recorded side lying, resting, rumination, grazing, walking, and panting. At the conclusion of the study, the most prevalent cow-state observed within 1 min was considered to be the cow-state the heifers presented at that interval. We then summarized the data in 1-min intervals such that the most prevalent cow-state was considered to be the behavior presented by the animal. When 2 cow-states had equal mode within a minute, the cow-state that better explained generalized bovine behavior patterns was recorded for the minute in question.

To evaluate the sensitivity (Se), specificity (Sp), accuracy, negative (NPV), and positive predicted values (PPV), the investigators compared the observer’s recorded cow-state (gold standard) within a minute with the AMD recorded cow-state within the same minute. Test characteristics were calculated as follows:

$${\displaystyle \begin{array}{l}{\displaystyle Se=\frac{truepositives}{truepositives+falsenegatives}}\end{array}}$$

$${\displaystyle \begin{array}{l}{\displaystyle Sp=\frac{truenegatives}{truenegatives+falsepositives}}\end{array}}$$

$${\displaystyle \begin{array}{l}{\displaystyle Accuracy=\frac{truepositives+truenegatives}{total}}\end{array}}$$

$${\displaystyle \begin{array}{l}{\displaystyle NPV=\frac{truenegatives}{truenegatives+falsenegatives}}\end{array}}$$

$${\displaystyle \begin{array}{l}{\displaystyle PPV=\frac{truepositives}{truepositives+falsepositives}}\end{array}}$$

As explained previously, the proprietary machine learning algorithm prioritizes certain behaviors over others based on their contribution to the understanding of generalized bovine behavior patterns. Therefore, determining the predominant cow-state on a minute-by-minute basis is not strictly based on which behavior occupies the majority of the minute but also on the significance of the behavior. Considering that most behaviors recorded by the observers do not occur sporadically within a day (e.g., rumination, resting, grazing), we determined the most prevalent behavior observed in 5-min intervals and evaluated the test characteristics and agreement between the observer and the AMD.

In study 2, the 10 Angus heifers from study 1 were housed in the same paddock to be observed at the same time on day 5, 6, and 7. Behavior was collected using instantaneous recording conducted with 5-min intervals, from 0800 h until 1100 h and from 1500 h until 1800 h, totaling 740 observations per day. The cow-state observed at each scan (e.g., every 5 min from 0800 h to 1100 h) was considered to have been preserved during the entire 5 min until a new scan was performed. The CCC analysis, the coefficient of determination (R²), and the regression equation between the total minutes per cow-states according to the observer and the AMD were determined using MedCalc.

Results and Discussion

We observed a high level of agreement between the 2 independent observers (single measures = 0.97 [95% confidence interval (CI) = 0.97 to 0.97]; average measures = 0.98 (95% CI = 0.98 to 0.99). Hence, for data analyses, the observations recorded by one of the observers, chosen at random, were considered as gold standard and were used for comparison with the data recorded by the AMD. The AMD validated in this study is able to record various cow-states, which may be an important tool to improve farm management at the herd and individual animal levels. It is important to mention that the algorithms used by the AMD used in the current study were developed mainly using confined dairy cattle, fed total mixed ration diets. Therefore, inherent differences in behavior of confined dairy cattle to grazing beef cattle may play a significant role in our findings.

In Table 2 (study 1), we present the total minutes per cow-state according to the observer and the AMD and the test characteristics. A total of 1,200 cow-state minutes was recorded by the observer and by the AMD. No panting was observed, whereas the AMD recorded 7 min of panting. The overall CCC was 0.70 (95% CI = 0.67 to 0.72). In Table 3 (study 1), we present the evaluation of cow-state in intervals of 5 min and the test characteristics. Walking was not a predominant cow-state within any of the 5-min intervals according to the observer, but it was predominant in 4 intervals according to the AMD. Although the overall CCC for walking was 0.80 (95% CI = 0.74 to 0.84), there was a very limited number of observations for this behavior, limiting its external validity. It is important to note that grazing, defined as the start of active search for food and biting and/or chewing pasture, inherently involves walking. Because the search for feed and eating may be considered biologically more important, events of grazing that also involved some walking or were interspaced by walking were recorded only as grazing. Therefore, walking may have been underestimated since walking within grazing would have been ignored.

Table 2.

Comparison of observer¹ and automated monitoring device (AMD) recorded cow-states according to minute-by-minute evaluation (study 1)

Cow-state	Observer, min	AMD, min	Se1, %	Sp1, %	Accuracy, %	PPV1, %	NPV1, %
Panting	0	7	NA	NA	NA	NA	NA
Rumination	300	319	95.3	96.3	96.1	89.7	98.4
Grazing/Eating	576	461	76.5	96.9	87.1	95.9	81.6
Side Lying	10	11	90.0	99.8	99.8	81.8	99.9
Resting	301	164	48.8	98.1	85.8	89.6	85.1
Walking	11	30	45.5	97.9	97.4	16.7	99.5
Medium activity2	NA	104	NA	NA	NA	NA	NA
High activity2	NA	104	NA	NA	NA	NA	NA

¹Observer was considered the gold standard for all test characteristics. Se, sensitivity; Sp, specificity; PPV, positive predicted value; and, NPV, negative predicted value.

²Cow-states measured by the automated monitoring device but not clearly specified to the observer before the start of the study; thus, not recorded by the observer.

Table 3.

Comparison of observer¹ and automated monitoring device (AMD) recorded cow-states according to 5-min intervals (study 1)

Cow-state	Observer, intervals	AMD, intervals	Se1, %	Sp1, %	Accuracy, %	PPV1, %	NPV1, %
Rumination	61	66	95.1	95.5	95.4	87.9	98.3
Grazing/Eating	122	102	82.8	99.2	90.8	99.0	84.8
Side Lying	2	2	100	100	100	100	100
Resting	55	34	52.7	97.3	87.1	85.3	87.4
Walking	0	2	NA	0	0	0	NA
Medium activity2	NA	10	NA	NA	NA	NA	NA
High activity2	NA	24	NA	NA	NA	NA	NA

¹Observer was considered the gold standard for all test characteristics. Se, sensitivity; Sp, specificity; PPV, positive predicted value; and, NPV, negative predicted value.

²Cow-states measured by the automated monitoring device but not clearly specified to the observer before the start of the study; thus, not recorded by the observer.

The observer and the AMD recorded 300 and 319 min of rumination, respectively. The test characteristics for rumination were high (Tables 2 and 3). A total of 576 and 461 min of grazing were recorded by the observer and the AMD, respectively. Similar to rumination, the test characteristics for grazing were high (Tables 2 and 3). Molfino et al. (2017) used grazing dairy cattle to validate rumination and grazing cow-states measured by the same device used in our study and observed high agreement between the AMD and visual observations. Other researchers using an older AMD also observed high agreement between the AMD recorded rumination time and visual observation (Schirmann et al., 2009; Elischer et al., 2013; Borchers et al., 2016). Therefore, the AMD tested in this study may facilitate the recording of behaviors generally associated with health and welfare (Hixson et al., 2018).

Only 10 min of side lying were observed and the AMD recorded 11 min of side lying. Within each minute of side lying recorded by the observer the sensitivity, specificity, accuracy, PPV and NPV were high (Tables 2 and 3). The observer recorded 301 min of resting, whereas the AMD recorded 164 min. According to the minute-by-minute comparison between the observer and the AMD, there were high specificity, accuracy, PPV, and NPV (≥85.1%) but the sensitivity was moderate. Similarly, sensitivity was moderate and all other test characteristics were high (≥85.3%) when we compared observer and AMD according to the 5-min intervals. Of the 301 min recorded by the observer as resting, the AMD only recorded 147 min as resting and 156 min as other cow-states such as medium and high activities (91 min), rumination (29 min), grazing/eating (19 min), walking (9 min), panting (4 min), and side lying (2 min). The specificity, accuracy, and NPV of walking recorded by the AMD were high (≥ 97.4%), but the sensitivity was moderate and PPV was low. Of the 11 min recorded by the observer as walking, the AMD only recorded 5 min as walking and 3 min each as medium and high activities. It is important to note that the very limited number of observations for walking in this study limits the external validity of our findings.

In Fig. 1, we depicted the differences of total minutes per cow-state recorded by the observer and the AMD on study day 5, 6, and 7 (study 2). There was a high correlation between total grazing time recorded by the observer and the AMD (Table 4), but the AMD slightly underestimated the total grazing time (Fig. 1A). Similarly, the correlation between observer and AMD for total rumination time was high (Table 4) and the differences between observer and AMD recorded total rumination time were more homogenous (Fig. 1B). The CCC for total resting time recorded by the observer and AMD was low but the correlation of determination between observer and AMD recorded resting time was moderate (Table 4). Although in general the AMD underestimated the total resting time, there was a positive correlation between observer recorded total resting time and the underestimation by the AMD (Fig. 1C). The correlation between observer and AMD recorded walking time was moderate according to the CCC and low according to the correlation of determination (Table 4).

Figure 1.

Differences between the observed and automated monitoring device (AMD) recorded total time (min/d) for grazing (A), rumination (B), and resting (C) in the morning and afternoon of study day 5, 6, and 7 (study 2).

Table 4.

Correlation between visual observation and automated monitoring device regarding total time of cow-state (study 2)

Cow-state	Observed, min/d	AMD, min/d	CCC (95% CI)a	Goodness of fit
				R2 b	Equation
Rumination time	1,862	1,883	0.96 (0.94, 0.98)	0.92	y = 0.8239 + 0.9626x
Grazing/Eating time	6,754	6,268	0.92 (0.88, 0.95)	0.90	y = 19.64 + 0.8895x
Resting time	1,967	896	0.48 (0.35, 0.58)	0.66	y = 14.911 + 1.1968x
Walking	267	210	0.53 (0.34, 0.68)	0.33	y = 1.6187 + 0.8089x

^aConcordance correlation coefficient and 95% confidence interval.

^bCorrelation of determination.

Continuous behavioral observation is a tool that may provide accurate measurements of rumination and grazing; however, it may be an unfeasible technique when there are too many animals being observed simultaneously. Previous research has validated time sampling using instantaneous recording with short intervals (5 min) for collection of a range of activity and feeding behaviors in feedlot cattle (Mitlöhner et al., 2001). The high correlations between the 5-min scanning technique and the AMD relative to total rumination and grazing times reiterate the finding of study 1, which is that the AMD used in this study is an accurate tool to monitor rumination and grazing behaviors. In this study, however, we only observed a moderate correlation between the 5-min scanning technique and AMD for total resting and walking time. As explained previously, the minute-by-minute accuracy of resting and walking cow-states recorded by the AMD proved to be moderate, which may explain the moderate correlation in total resting and walking time between the 5-min scanning technique and the AMD.

We concluded that the AMD used in this study is an accurate tool to measure cow-states such as rumination and grazing but it is subpar when detecting resting and walking. The accuracy of measurements referent to side lying was high; however, very few side lying events were recorded in this study making necessary additional studies to confirm our findings. It is important to emphasize that the AMD used in this study employs algorithms developed for dairy cattle; nonetheless, it seems to be a good tool to measure behaviors (rumination and grazing) associated with health and welfare.