Received 9 June 2023, accepted 25 June 2023, date of publication 30 June 2023, date of current version 18 July 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3291348

A Machine Learning and Optimization

Framework for the Early Diagnosis
of Bovine Respiratory Disease
ENRICO CASELLA 1 , (Member, IEEE), MELISSA C. CANTOR 2 ,
MEGAN M. WOODRUM SETSER 3 , SIMONE SILVESTRI 1 , (Senior Member, IEEE),
AND JOAO H. C. COSTA 4
1 Department of Computer Science, University of Kentucky, Lexington, KY 40506, USA
2 Department of Animal Science, Pennsylvania State University, University Park, PA 16802, USA
3 Department of Animal and Food Sciences, University of Kentucky, Lexington, KY 40546, USA
4 Department of Animal and Veterinary Sciences, University of Vermont, Burlington, VT 40546, USA

Corresponding author: Enrico Casella ([email protected])

This work was supported in part by NIFA under Grant 2021-68014-34139, in part by the NSF Smart and Connected Communities under
Grant 1952045, and in part by the U.S. Department of Agriculture National Institute of Food and Agriculture Hatch under
Project KY007100.
This work involved human subjects or animals in its research. The authors confirm that all human/animal subject research procedures and
protocols are exempt from review board approval.

ABSTRACT Bovine Respiratory Disease (BRD) is an infection of the respiratory tract that is the leading
reason for antimicrobial use in dairy calves and represents 22% of calf mortalities. The costs and effects
of BRD can severely damage a farm’s economy, since raising dairy calves is one of the largest economic
investments, and diagnosing BRD requires intensive and specialized labor that is hard to find. Precision
technologies based on the Internet of Things (IoT), such as automatic feeders, scales, and accelerometers,
can help detect behavioral changes before outward clinical signs of BRD. Such early detection enables
early treatment, and thus faster recovery, with less long term effects. In this paper, we propose a framework
for BRD diagnosis, its early detection, and identification of BRD persistency status using precision IoT
technologies. We adopt a machine learning model paired with a cost-sensitive feature selection problem
called Cost Optimization Worth (COW). COW maximizes prediction accuracy given a budget constraint.
We show that COW is NP-Hard, and propose an efficient heuristic with polynomial complexity called Cost-
Aware Learning Feature (CALF). We validate our methodology on a real dataset collected from 159 calves
during the preweaning period. Results show that our approach outperforms a recent state-of-the-art solution.
Numerically, we achieve an accuracy of 88% for labeling sick and healthy calves, 70% of sick calves are
predicted 4 days prior to diagnosis, and 80% of persistency status calves are detected within the first five
days of sickness.

INDEX TERMS Dairy calves, precision IoT technologies, machine learning, cost-sensitive optimization.

I. INTRODUCTION the cause of 22% of dairy calf deaths in the preweaning

Bovine Respiratory Disease (BRD) is an upper or lower period [3]. Furthermore, there is preliminary evidence that
infection of the respiratory tract in cattle [1] that is not only up to 20% of BRD calves develop BRD persistency status,
economically costly, but also associated with poor produc- which may lead to calfhood pneumonia, a painful disease
tivity in dairy calves [2]. Specifically, BRD was reported that compromises calves’ normal behavioral patterns [4].
to affect 15% of dairy calves, and it was attributed to be Calves with BRD persistency status also require an additional
antimicrobial treatment to recover, which can take several
The associate editor coordinating the review of this manuscript and days after the initial BRD diagnosis and treatment [4]. As a
approving it for publication was Daniel Augusto Ribeiro Chaves . result, they either relapse and eventually recover after the
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
71164 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

second antimicrobial intervention, or they become chronic trainee for a period of about 12 weeks (which includes
and may even die [5], further impacting the farm’s econ- turnover rates at 38.8% [15]), and $1.83 of daily payments
omy [4]. In addition, BRD persistency status must be per calf to perform the exams. On the other hand, a pedome-
managed on dairies to ensure the judicious use of antimi- ter only requires a purchase of $90 per calf and a daily
crobials, as classes of these drugs are becoming tightly maintenance of $0.38 per calf, while a grain feeder requires
managed [3]. Thus, BRD and BRD persistency status are a one-time purchase of $6,500 and a daily maintenance of
a major animal welfare and industry sustainability concern $1.62 [16].
that requires research strategies to identify their occurrence, In order to fully enable the benefits of precision tech-
possibly at early stages, in order to ensure effective treatment nologies, it is fundamental to design solutions that find the
and recovery. best tradeoff between cost of data collection and accuracy of
Dairy producers monitor their calves’ health daily for out- predictions. This results in analyzing the impact of varying
ward signs of disease [1]. However, assessing each calf for the start-up budgets and daily budgets on the prediction perfor-
cumulative signs of BRD is labor intensive [6], and extensive mance, taking into consideration both automatically collected
training is required to achieve agreement [7]. Moreover, many features by precision technologies, and manually collected
dairy farms struggle to find labor to raise and monitor their features through the intense labor of trained professionals.
calves, especially during the COVID-19 pandemic [8], [9]. In livestock medicine [17], most direct costs to detect and
Because of these factors, farms are incorporating precision resolve diseases in dairy cattle are incurred by the dairy
technologies based on the Internet of Things (IoT) to monitor farmer [18]. An investigation of the best trade-off between
their calves [10]. Examples of these technologies include accuracy and cost can guide the farmers to more efficient
wearable sensors, which record activity behavior data in dairy investments. Besides, existing works on cost-based feature
cattle (e.g., number of steps and lying time), and automatic selection oversimplifies feature costs in a way that is not real-
feeders, which can dispense milk and grain resources and istic in practice [19], [20]. For instance, costs are considered
record feeding behaviors (e.g., number of visits and liters of on a per-feature base, ignoring the fact that different features
consumed milk) [11]. Recent studies observed that there is coming from the same sensor incur the same one-time cost of
a negative statistical correlation between activity behavior, purchasing the sensor. Besides, costs need to be considered
feeding behavior, and BRD status in calves [12]. Thus, there in terms of start-up costs, for the initial purchase, along with
is a significant opportunity to exploit precision IoT technolo- daily costs, for maintenance.
gies, coupled with machine learning techniques, to identify In this paper we propose a framework for BRD diagnosis
BRD status in a timely manner. in calves, as well as the detection of early BRD status and
Additionally, research shows that relapsed calves depress BRD persistency status. This framework is composed by a
their feeding behavior and activity levels more than recovered machine learning model as well as by a cost-sensitive feature
ones [4]. This evidence suggests that behavioral monitor- selection problem called Cost Optimization Worth (COW).
ing through precision technologies can be used to detect COW selects the best features that maximize the prediction
BRD persistency status, i.e., BRD calves who relapse and accuracy, within a given a budget constraint.
require an additional antimicrobial to recover, or calves who We show that COW is NP-Hard and propose an efficient
become chronic and rarely recover from the disease [5]. heuristic with polynomial complexity called Cost Aware
Early evidence suggests that re-antimicrobial intervention Learning Feature (CALF). CALF is an algorithm based on
provides better response when administered as soon as pos- ε-greedy [21], which uses a mix of exploration and exploita-
sible on calves with BRD persistency status. In fact, delayed tion to test different subsets of features within the given
antimicrobial treatment for BRD can lead to poor growth start-up and daily budgets to maximize accuracy.
performance, reduced milk production, and chronic disease We validate our approach through extensive real-field
in dairy cattle [5], [13]. Thus, the early identification of BRD data collection over a period of 2 years from 159 calves
persistency status may help farmers make early antimicrobial for a total of 97 manual and automatic features collected
re-interventions to avoid chronic status calves who often need by trained researchers and precision livestock technologies,
to be euthanized. Diagnosis and early prediction of calves respectively. Our dataset, which is made publicly available
with BRD persistency status through precision IoT technol- with this paper at [22], is the first to collectively consider
ogy and machine learning is yet to be explored. a multitude of precision technologies to diagnose BRD in
The adoption of these technologies to monitor calves dairy calves, together with manual examination and ultra-
for signs of BRD and BRD persistency status may be a sonography. In addition, this dataset has the largest number of
viable alternative that also reduces farms’ economic bur- adopted precision technologies. Procedures of data analysis
den [10], [14]. However, precision IoT technology requires are carried out to prepare the data for feature extraction.
start-up investments as well as daily investments and main- These features are paired with real start-up and daily costs to
tenance that need to be evaluated together with the current perform feature selection. Post-processing methodologies are
state-of-the-art methodologies. For instance, health exams used to split the data into training and testing sets implement-
require an initial investment of $11,992 to pay trainer and ing a calf-based approach that offers more realistic results

VOLUME 11, 2023 71165

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

FIGURE 1. Outline of the proposed methodology.

than the current state of the art [23], [24], [25]. Finally, data processing, feature selection, and machine learning. Data col-
are cross validated and fed to machine learning models for lection is the phase where data was collected from the calves
classification. Results show that our approach outperforms using precision livestock technologies and by researchers
a state-of-the-art solution [24], with accuracy up to 88% in who performed physical health exams on the calves. Due
the diagnosis scenario, 70% in the presick scenario four days to the intrinsic characteristics of data collection methods,
prior to BRD, and 80% in the chronic scenario in the first five we identify automatic attributes and manual attributes.
days of illness. Specifically, automatic attributes consist of data collected
To the best of our knowledge, this is first work to by precision technologies including activity behavior, milk
study BRD persistency status through cost-effective machine feeding behavior, and grain feeding behavior, while manual
learning techniques, and the first study to publish such a com- attributes consist of variables collected by the researchers
prehensive dataset, which includes ultrasonography along such as health exams, body weights, and lung ultrasound
with health exams, and provides the highest number of preci- imaging. Note that, ultrasound imaging is visually catego-
sion technologies. In summary, the main contributions of this rized by domain experts in a value from 0-3 based on the size
paper are listed below: of lung consolidation. Hence, it is a numerical value and no
• We propose the first framework to ever perform early actual image processing technique is carried out. Precision
prediction of BRD persistency status; technologies are equipped to automatically send data to the
• We propose a framework for BRD diagnosis and its early cloud, while manual exams are manually uploaded to the
prediction in dairy calves; cloud by the domain experts who collected the data. More
• We perform real-field experiments with a 2-year long details on each feature group is given in Sec. V-A.
data collection, and make our dataset publicly available; The Data Processing phase includes several practices of
• Our dataset is the first to comprehensively include preci- data cleaning and data filling. A segmentation step aggregates
sion technology, manually performed health exams, and consecutive daily samples into windows of data, which allows
ultrasonography imaging to diagnose BRD; for feature extraction. The extracted automatic features and
• Our dataset has the largest number of adopted precision manual features are then organized in groups that share the
technology devices up to date; same start-up cost and daily cost. For automatic features,
• We formalize COW, a cost-aware feature selection prob- start-up and daily costs correspond, respectively, to the pur-
lem to maximize accuracy based on a certain budget; chase of the equipment and the daily maintenance costs.
• We prove that COW is NP-Hard and propose an efficient For manual features, costs reference personnel salary, which
heuristic called CALF. includes training as a start-up cost. It is worth mentioning
A preliminary version of this work appeared in [23]. This that, from a financial perspective, selecting a single feature
paper extends the conference version by proposing several in a group is equivalent to selecting all features in that group.
improvements and additions of our proposed methodology, Intuitively, this is because a sensor can either be used, or not
allowing us to perform accurate predictions of BRD per- used. Once a sensor is used, and thus their cost deducted
sistency status, as well as better early detection of BRD. from the available budget, then you have access to multiple
Other improvements include a more effective feature space attributes, whether or not they are all used. This is a unique
exploration, through an optimized selection of features within characteristic of our problem. In fact, previous cost-sensitive
a given budget, and a more realistic train/test split methodol- feature selection approaches do not take into account such
ogy. Finally, we collect additional data and, for the first time, group-based cost modeling, and they lack a refined definition
release the dataset to the public. of budget in terms of start-up cost and daily cost [19], [20].
The Feature Selection phase consists of the cost-aware
II. SYSTEM OVERVIEW feature selection problem called COW (Cost Optimization
The overview of our framework is depicted in Fig. 1. The Worth), further discussed in Section III. We show that COW
framework consists of 4 steps, namely, data collection, data is NP-Hard and propose an ε-greedy heuristic algorithm

71166 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

called CALF (Cost Aware Learning Feature). CALF relies on including noisy or uninformative features may decrease accu-
exploration and exploitation to navigate the high-dimensional racy. Eq. (1b) defines the set of selected features S through
feature space [21]. Note that, not necessarily all features in a a decision variable xij , which is equal to 1 if feature fij is
group should be selected, even though they may be selected selected, and 0 otherwise. In order to take into account the
at no extra cost. In fact, noisy and uninformative features may group nature of the incurred costs, we introduce an additional
worsen the accuracy of a scenario. For this reason, selecting decision variable zi for each group Gi . The variable zi defined
the best set of features within a start-up and a daily budget is in (1e) equals one if at least one feature in the group is
a challenging problem. selected, and thus the corresponding costs incurred, or zero
Finally, the Machine Learning phase is characterized by otherwise. Then, (1c) and (1d) enforce the budget constraints,
different methodologies to split the dataset in training and to ensure that the cost of all groups for which a feature is
testing sets, depending on the BRD scenario that we are inves- selected stays within the daily budget BD and start-up budget
tigating. Furthermore, we implement such splitting methods BS . Finally, (1f) defines the domains of the decision variables.
with a realistic calf-based approach that guarantees realistic In the following, we prove that COW is an NP-hard
and balanced sample shuffling. Then, the selected features problem, which motivates the development of an efficient
are used to train the adopted machine learning models, such heuristic solution to solve such problem.
as Gradient Boosting Classifier (GBC) and Support Vector Theorem 1: COW is an NP-hard problem.
Machines (SVM). Proof: [Sketch of proof] We provide a reduction from
In the following, we present further details of each step the 0-1 knapsack problem (KP). A generic instance of KP
of our framework. For ease of exposition, we first present considers a set of items I, where each item fi ∈ I has cost
COW, discuss its complexity, and introduce CALF. Then, ci and value ai . The objective is to find the subset S ⊆ I
we provide more details on the proposed scenarios of BRD of items which provide maximum value within a budget B.
development, followed by data collection, data processing, Given such generic instance, we set COW as follows. Each
machine learning, and finally the experimental results. group has only one feature, i.e., Gi = {fi }N i=1 (and Mi = 1).
We set the cost ci for Gi equal to the corresponding cost in KP.
III. PROBLEM FORMULATION Then, we define the function A() as a modular
P linear function
We consider a set of features F with N feature groups, of the value defined in KP, i.e., A(S) = fi ∈S ai . In addition,
defined as F = {fij |i = 1, . . . , N ; j = 1, . . . , Mi }, where we set BD = B and BS = ∞. Finally, since we have a single
fij is the j-th feature of the i-th group, and Mi is the number of feature per group, zi = xij .
features in group i. F can be also expressed as F = ∪N i=1 Gi ,
Solving such instance of COW corresponds to finding the
where Gi is the i-th group. Each group Gi has a start-up cost set of features S that maximizes accuracy within the daily
si , and a daily cost di . Thus, we define a start-up budget budget. This can be polynomially translated into the optimal
BS and a daily budget BD . In the following, we formulate solution of KP. Therefore, COW is at least as difficult as KP,
Cost Optimization Worth (COW), a cost-sensitive optimiza- and thus it is NP-Hard. □
tion that finds the feature set with highest accuracy1 within
budgets. A. COST-SENSITIVE FEATURE SELECTION
Since COW is NP-hard, we present CALF (Cost Aware
Learning Feature), a cost-aware feature selection heuristic.
maxS⊆F A(S) (1a)
[ 1) CALF
subject to S = {fij } (1b)
Our heuristic algorithm is inspired by the reinforcement
fij ∈F |xij =1
learning concept of exploration versus exploitation [21]. Due
G
X to the non-monotone nature and non-modular property of
zi di ≤ BD (1c)
the accuracy function A(), exploring the feature space and
i=1
discovering the accuracy of different subsets is essential to
G
X finding a solution for COW. CALF iteratively builds a solu-
zi si ≤ BS (1d)
tion by performing actions. An action consists of adding a
i=1
feature or removing a feature group from the current solution
zi ≥ xij ∀i = 1, . . . , N , j = 1, . . . , Mi (1e)
and testing the accuracy of CALF. In the following, we use the
zi , xij ∈ {0, 1}, i = 1, . . . , N , j = 1, . . . , Mi term ‘‘exploration’’ to perform an action in a random fashion,
(1f) while we use the term ‘‘exploitation’’ to perform an action
greedily, i.e., the best short-term action.
COW aims at finding the subset S ⊆ F that maximizes A recent approach to the exploration versus exploitation
the accuracy function A(), as defined in (1a). Note that A() tradeoff is the ε-greedy algorithm [21]. The idea of ε-greedy
is not necessarily a monotone nor a modular function, since is that a variable ε is properly tuned to prioritize more explo-
ration at the beginning of the algorithm execution, while
1 ‘‘Accuracy’’ here is used as a general term of a classification metric. slowly exploiting the current knowledge to perform the best

VOLUME 11, 2023 71167

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

next action over time. CALF uses two of such variables, In lines 4-18 we have a loop of L iterations. In line 5,
making it a ε2 -greedy algorithm. A variable εEE regulates we define F̂ which corresponds to the set of available features
exploration and exploitation, as intended in the original F̂, organized by group. Similarly, in line 6 we define F ∗ as
ε-greedy algorithm, while another variable εAR regulates the the set of selected features, organized by group. At each iter-
action of adding or removing features, by considering the ation, four possible actions can be taken as a combination of
utilized budgets and the number of available features that exploration (RND) or exploitation (GREEDY), and addition
could be added at no cost to prioritize either action. The (ADD) or removal (REM). Each action is selected with a
pseudo-code of CALF is given in Alg. 1. probability that depends on the value of εEE and εAR . Lines
7-8 define the action (RND − ADD), corresponding to explo-
ration and addition, which occurs with probability εEE εAR .
Algorithm 1 Cost Aware Learning Feature (CALF)
We select a random feature f from a randomly selected group
Algorithm
Ĝi ∈ F̂, and add it to the current solution set S. When
Input : F = ∪N i=1 , Gi , BD , BS , L, α, β action (RND − REM) occurs in lines 9-10 with probability
Output: A feature set SB and its accuracy score aB εEE (1 − εAR ), a random group is removed from the current
1 S ← ∅, CD = 0, CS = 0 ; // Loop variables solution set S. Note that, while the addition occurs for a single
2 εAR = 0, εEE = 1 ; // Tuning variables feature of a group, the removal occurs for a whole group. This
3 Let F̂ ⊆ F \ S be the set of available features within approach helps to perform a thorough exploration because
budget removing a group frees up budget that allows exploration
4 Pick one of the following actions while εEE > 0: of other groups. On the contrary, removing one feature at a
5 Let F̂ = {Gi ∩ F̂}N i=1 , set of available features by time would force us to rely on the random removal of all
group features from the same group before freeing up budget, thus
6 Let F ∗ = {Gi ∩ S}N i=1 , set of selected features by preventing an issue of local maximum.
group Action (GREEDY − ADD), in lines 11-12, is executed
7 RND - ADD with probability εEE εAR : with probability (1 − εEE )εAR . This action picks the feature
8 S ← S + f where f ∈ Ĝi = rnd_pick(F̂) that, once added to the current solution S, provides the highest
9 RND - REM with probability εEE (1 − εAR ): accuracy score among the available features F̂. In lines 13-14,
10 S ← S − G∗i where G∗i = rnd_pick(F ∗ ) action (GREEDY − REM), performed with probability
11 GREEDY - ADD with probability (1 − εEE )εAR : (1 − εEE )(1 − εAR ), removes the feature group that provides
12 S ← S + f where f = argmaxf ∈F̂ A(S + f ) the minimum ranking criterion, defined as
13 GREEDY - REM with probability A(S) − A(S − G∗i )
(1 − εEE )(1 − εAR ): rank = (2)
αs∗i + βdi∗
14 S ← S − G∗i where G∗i = min_ranking(F ∗ )
which prioritizes the removal of groups that either increase
15 Update costs CD and CS
accuracy the most (when numerator is negative) or reduce
16 SB , aB = update_if _best(S) ; // update
accuracy the least (when numerator is positive), while pri-
output
oritizing the removal of the most expensive ones (thus large
17 Re-calculate F̂
denominator). Then, we update the current daily cost CD , and
18 εAR = 1−(α CBDD +β CBSS )( |S| ), εEE = εEE − L1 the start-up cost CS in line 15. Line 16 checks if the current
|S|+|F̂|
19 return SB , aB solution in S provides better accuracy than the best solution
encountered so far, SB , and updates it accordingly.
Finally, calculating F̂ in line 17 allows us to update the
The input variables and parameters of CALF are: the set value of εAR in line 18, where εEE is updated as well. The
of features F = ∪N probability εAR of adding or removing a feature is updated
i=1 Gi , the total available budgets BS and
BD , the total number of exploration/exploitation loops L, and according to the remaining available budgets (properly scaled
the normalizing variables α and β. The output is SB , i.e., by α and β) weighed by the ratio |S| , which reduces the
|S|+|F̂|
the subset of features that had the highest accuracy score aB . impact of the budgets if many features are still available at no
In line 1, we initialize the solution set S, as well as the costs cost. The probability εEE of exploration over exploitation is
CD (daily cost) and CS (start-up cost) to keep track of how reduced by L1 at each iteration.
much budget has been used. Then, in line 2, we initialize
εEE = 1, which regulates exploration and exploitation, and 2) ACCURACY FUNCTION A()
εAR = 0, which regulates addition and removal of features. CALF is designed to be used with any metric function A().
These variables are set to favor exploration and addition of This may be accuracy, f1-score, or any other evaluation metric
new features during the first part of the execution. In line 3 we for classification tasks. The metric function used in our imple-
create F̂, the set of available features, i.e., the set of features mentation is the harmonic mean of the two labels’ accuracy,
that can be selected at this iteration without violating budgets i.e., sick and healthy. This choice lies behind the idea that an
(CD ≤ BD and CS ≤ BS ). arithmetic mean may return a high score in situations where

71168 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

one label has extremely high accuracy, while the other is

just moderate. However, the harmonic mean requires that,
by definition, both accuracy scores need to be high, which
better suits the task of finding the best performing set of
features for all labels. On the other hand, we did consider and
experimented using the mean of per-class F1-scores, which
is a harmonic mean by definition. We tested its use and
its cross-validated performance, and found that the features
sets selected by F1-score did not perform as well. Besides,
we found F1-score to be a less intuitive metric. Note that,
our experiments to evaluate the use of F1-score and harmonic FIGURE 2. Visual representation of BRD stages over time.
mean of accuracy for the metric function A, the evaluated
cross-validation performance was always assessed using both
accuracy and f1-score, and results were compared with each
other. we identify two persistency scenarios, namely the relapse
scenario and the chronic scenario. Specifically, the relapse
3) COMPUTATIONAL COMPLEXITY scenario captures calves who do respond positively to the
The computational complexity of CALF is dominated by antibiotic treatment on day 14, i.e., ‘‘AT 3’’, and thus result
the machine learning algorithm employed to compute the healthy before day 21. However, the chronic scenario cap-
accuracy score, since the loop is a costant value (L). In this tures those calves who receive such treatment and still do
paper, as discussed in Section VI we adopt Support Vector not recover even after 21 days from BRD diagnosis, after
Machines (SVM) and Gradient Boosting Classifier (GBC). which they may be given additional treatment, be euthanized,
As a result, the overall complexity depends on the adopted or sold for beef once antimicrobials are no longer circulat-
algorithm, and specifically its training. Given T as the number ing in the calf’s system [5]. Fig. 2 shows a comprehensive
of training samples, and F as the number of features, the representation of these stages. Note that, relapse and chronic
complexity of CALF using SVM is O(T 3 ) [26], while using calves are labeled as such starting from day 0 because the
GBC the complexity is O(T × F) [27]. objective of this paper is to detect these calves during the first
few days of sickness, i.e., capture those behavioral difference
IV. PRE-SICK AND BRD PERSISTENCY between relapse or chronic calves against sick calves at an
STATUS SCENARIOS early stage. The ability to perform such early prediction of
In this section we present additional prediction scenarios. BRD persistency status allows domain experts to provide
So far, we have talked about labeling calves as either sick with additional appropriate treatment which, we suggest, can help
BRD or healthy. We refer to such prediction as the diagnosis farmers’ decision making and allow fast BRD recovery.
scenario, because the algorithm aims at finding sick calves
when they show clinical outward signs of disease. However, B. ADAPTING CALF AND A() to BRD SCENARIOS
in the following we present three more scenarios that capture In the presick scenario and both persistency scenarios, addi-
the other stages of BRD. tional steps are added to the calculation of CALF’s accuracy
score A(). The main difference of these scenarios is that
A. CHARACTERISTICS not only do we want to have accurate predictions, but it
Firstly, we present the pre-sick scenario. Although a formal also matters how accurate they are with respect to the day
definition of this stage does not exist, it represents what is of BRD diagnosis. In fact, the goal is to make an accurate
commonly known as the sickness behavior stage. This is prediction as early as possible, i.e., several days before BRD
when the immune system has encountered a pathogen in diagnosis in the presick scenario, and as soon as BRD is
the body, and the activation of inflammatory cytokines are diagnosed in the persistency scenarios. However, accuracy is
present, which makes the animal adapt behavior to conserve a single score that does not account for such time component.
energy, and can occur before the calf is clinically outwardly Thus, we envision A() as a linear combination of accuracy
sick [28]. In fact, the goal is to perform the early prediction of scores and weights, where weights can be properly tuned to
a calf incurring BRD a few days prior to the clinical diagnosis. prioritize the days where high accuracy is most important.
A successful prediction of this stage is important because it By doing so, we are instructing the algorithm to give us an
would allow the first Antibiotic Treatment (AT), displayed overall high score A() when individual high accuracy values
in Fig. 2 as ‘‘AT 1’’, to occur before day 0, which often are obtained in the prioritized days.
results in faster recovery [29]. Some calves may, at times, Therefore, we split a window of 10 days in two sub-
be treated again on day 7, as indicated by the ‘‘AT 2’’ windows that can be properly tuned to fit our weight
grey arrow in Fig. 2, and they are always treated again on requirements. One window is characterized by a low start
day 14 [4], if BRD persists. If calves are still sick after day 14, and a quick growth. The other window starts high and

VOLUME 11, 2023 71169

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

a: ACTIVITY BEHAVIOR DATA

Each calf wears a commercial pedometer sensor, IceQube,2
with a three-axis accelerometer strapped to the rear left leg
at 24 hours of age. Data are collected at a frequency of 4Hz,
and a summary of the collected data are sent to a cloud server
every 15 minutes. Daily summaries of the following attributes
are then generated: total step count (steps/day), lying time
(hours/day), lying bouts (bouts/day), and motion index
(a scalar score).
FIGURE 3. Weights of A().

b: FEEDING BEHAVIOR DATA

Each calf has a commercial unique RFID tag attached to the
decreases slowly. These behaviors can be expressed, respec-
left ear. On average, at 3 days of age, calves are trained to
tively, by an exponential function and a logarithmic function,
drink up to 10 L/d of milk replacer from a commercial auto-
as follows:
mated calf feeder3 with a maximum 3L meal size allocated
across a 24 hour interval. A separate commercial automated
E(d) = e(d−q) , L(d) = ln(−d − (p + 1)) (3) feeder4 dispenses grain, but no limits are set for grain meal
size. When a calf approaches the automated milk feeder, the
where p and q are properly set based on the day in which we software recognizes each RFID to identify if a calf qualifies
want our weight functions to go to 0, and d is the day in which for a milk meal. After each new visit, data are sent to the
the weight is calculated. cloud server, and at the end of each day, data are summarized
Specifically, in the presick scenario, L(d) is applied for into daily attributes for each calf: milk intake (L/day), per-
−5 ≤ d ≤ −1 with p = 0, while E(d) is applied for centage of milk consumed (%), daily average drinking speed
−10 ≤ d ≤ −6 with q = −11, thus prioritizing the (ml/min), 12-day average drinking speed (ml/t), rewarded
highest accuracy on day -5. In fact, while trying to get a visits (visits/day), and unrewarded visits (visits/day). The
high accuracy any earlier would be unrealistic [23], a high grain smart feeder records the daily grain intake (g/day) for
accuracy on day -5 intrinsically gives good performance in each calf.
the days closer to diagnosis as well (−5 ≤ d ≤ −1).
Then, weights are normalized. Specifically, each sub-window 2) MANUAL ATTRIBUTES
is individually normalized first, and then all weights properly Conversely, manual attributes are manually collected by
scaled to have an overall sum equal to 1. Scaling parameters trained researchers and consist of the following 3 sub-groups:
are set to 0.3 for the logarithmic weights and 0.7 for the health exam data, weight data, and lung ultrasound data.
exponential weights, which give equal maximum values in As opposed to automatic features, which require a one-time
each sub-window. On the other hand, in both persistency setup of devices and maintenance of the technology, highly
scenarios, L(d) is applied for 2 ≤ d ≤ 9, with p = −12, skilled and intensive labor is required to retrieve manual
while E(d) is applied for 0 ≤ d ≤ 2 with q = 0, thus setting features about the calves.
the highest weight on day 2, i.e., the third day of BRD [4].
After individual normalization, further scaling is applied with
a: HEALTH EXAM DATA
weights of 0.2 and 0.8 respectively. Fig. 3 shows a visual
representation of the final weights for each scenario. A health exam is performed on each calf daily using a
validated veterinary system for identifying Bovine Respira-
tory Disease, the Wisconsin scoring system [6]. Specifically,
V. DATA COLLECTION AND DATA PROCESSING
a score from 0 (no signs) to 3 (severe signs) is assigned
A. DATA COLLECTION
to each of the following attributes: nasal discharge score,
We collected data from 159 calves through a mix of mod- eye discharge score, ear tilt score, cough score, and a rec-
ern precision technologies for automatic tracking, as well as tal temperature score. A BRD total score is also created,
manual health exams as supported by state-of-the-art dairy where abnormal scores in each category are summed. Respi-
veterinary research [10], [30], and ultrasonography imaging. ration rate for each calf is assessed daily as either normal or
Our dataset is publicly available at [22]. labored [31]. Finally, each calf’s umbilical health, and fecal
score are collected as part of the daily health exam since these
1) AUTOMATIC ATTRIBUTES
Precision technology sensors deployed at the farm are 2 https://www.icerobotics.com/researchers/#sensors
employed for data collection of automatic attributes and con- 3 https://www.foerster-technik.com/calf-feeding/automatic-calf-feeder/
sist of the following 2 sub-groups: activity behavior data, 4 https://www.foerster-technik.com/calf-feeding/calf-feeder-
and feeding behavior data (divided in milk and grain). accessories/maxiflex-feeding-box/

71170 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

non-respiratory parameters have shown to be predictive of days of data into a window allows us to extract features
mortality in calves [32]. from each attribute. For each of the 25 attributes within a
window, four features are extracted, namely, mean, minimum,
b: WEIGHT AND ULTRASOUND DATA maximum, and sum. An exception is made for the ‘‘age’’
Body weight by scale5 and ultrasound attributes by lung attribute, for which we only calculate the maximum. Thus,
ultrasound6 are collected once every three days in each calf. we have a total of 97 features. After feature extraction, each
Note that ultrasound imaging is not processed by means window is labeled based on BRD status of the last sample,
of image processing techniques. Instead, similarly to health which corresponds to the output label.
exams, domains experts assign a score from 0 to 3 based on In the diagnosis scenario, samples with BRD status ≥ 0 are
the size of lung consolidation. labeled as sick (S (+) ), and samples with BRD status ≤ −10 as
Finally, two more attributes are considered, namely, the healthy (H (−) ).In the presick scenario, additional samples
age of the calf and the Immunoglobulin G (IgG) status at with −10 < BRD status < 0 are labeled as pre-sick (PS).
48 to 72 hours of age, the IgG is a measurement of successful Furthermore, in the persistency scenarios, a portion of the
transfer of passive immunity status from the cow’s colostrum sick calves are labeled as relapse if they are sick for longer
to the postnatal calf. than 14 days and up to 20 days, while calves labeled as
chronic are sick for longer than 21 days. Note that, as further
B. DATA PROCESSING AND OUTPUT LABELS mentioned in Sec. IV-A and shown in Fig. 2, relapse and
Due to the nature of the data collection processes and the chronic samples are labeled as such from day 0, despite their
corresponding technologies, different datasets were initially eligibility being related to days 14 and 21. This labeling is
created during the data collection process. In fact, each auto- necessary in order to train models so that they can early
matic attribute group has its own cloud server, due to the predict such persistency status in the first days of sickness.
fact that the adopted precision technologies are not built Finally, we perform standardization, which consists of
in-house, but rather purchased from commercial livestock scaling each individual feature to have mean equal to 0 and
farming companies. Because the development of signs of standard deviation equal to 1, thus creating homogeneous dis-
BRD in calves occurs over several days [30], daily summaries tributions that machine learning models can fairly interpret.
are created for each dataset and they are merged together
into one dataset. Reducing the data into daily samples, rather VI. MACHINE LEARNING
than, for example, hourly summaries, may seem like a loss In this section we provide further details on the training of
of detail. However, it is important to note that calf feeding our models under each scenario, such as how the dataset
behavior peaks in the morning and the evening, thus certain was split into training and testing, which samples and calves
information are not actually perceived at a finer detail [33]. were included in the model, and what algorithms were used.
However, weight and ultrasound, are collected once every Specifically, we first provide general information about our
three days. Hence, weight data are filled in by means of a realistic calf-based train/test split methodology. Then, we dis-
linear interpolation to simulate the natural growth of the calf’s cuss the details of the training phase of each scenario.
weight, while ultrasound data are filled in by propagating
each collected data point forward.
A. TRAIN/TEST SPLIT
An output label named BRD status defines the ground
truth. BRD positive samples are represented by two cat- The classic and most common approach of splitting the
egories with abnormal scores according to the Wisconsin dataset into training and testing is to set a certain ratio of
Scoring System [6], and lung consolidation at 3cm2 observed training and testing data, and then randomly sample values to
by ultrasound [13]. Thus, for each calf that develops BRD, fill each set. However, in this specific context, such approach
the first day of diagnosis is labeled with 0. Samples preceding is unrealistic, even though it has erroneously been proposed
day 0 are labeled from −1 up to −10. Similarly, the days after in similar works [23], [24], [25]. This would add samples of
day 0 are labeled with increasing numbers, until the calf is no a single calf to both training set and testing set. As a con-
longer sick with BRD. Finally, healthy data points are labeled sequence, the trained model would classify a calf’s sample
with −∞. almost perfectly, because it may already have information
In order to gather how behavior changes over a certain about its past and its future. On the other hand, while it may
time frame, we perform a segmentation step to represent the be realistic to split your dataset in such a way that only past
trend of each attribute within a window of consecutive days. knowledge is exploited to make inferences about the future,
Thus, a few consecutive days are summarized into one single this approach might never use information of calves beyond
sample, and then such window is shifted by one day at a time a certain age, which is undesirable. Thus, in order to produce
to generate the next sample. Aggregating these consecutive realistic experiments, we propose a methodology according
to which a calf belongs to either training or testing, but not
5 https://www.brecknellscales.com/products/veterinary-scales/ps1000-/- both. At the same time, samples from a calf may have more
ps2000-veterinary-scales.html than one label. Hence, selecting calves in a way that the
6 https://www.eimedical.com/pro training set is properly balanced is non-trivial.

VOLUME 11, 2023 71171

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

Firstly, our train/test split methodology finds the number of additional presick label was added to sick and healthy labels,
samples of the minority label corresponding to a certain train- the strength of our models would be damaged by the bal-
ing percentage, e.g., 75%. Then, for each label, we randomly ancing step, since the number of presick samples is much
pick calves and add the portion of their samples with such lower than sick and healthy ones. However, intuitively, it is
label, until the calculated 75% threshold is reached. Finally, expected that presick samples would have a certain similarity
we randomly remove the excess samples of the majority to healthy samples when they are chronologically farther
label, if such situation occurs, in order to ensure balance. All from BRD diagnosis, and a similarity to sick samples oth-
remaining data from calves that have not been added to the erwise. Hence, we expect our model to classify our presick
training set are simply added to the testing set. samples as sick when they are closer to BRD diagnosis, and
As a practical example, consider having 720 sick samples healthy otherwise. In light of this, our training set for the
from 35 calves, and 1030 healthy samples from 25 calves. presick scenario consists of only sick and healthy samples.
Note that, this imbalance is not uncommon, since many sick Presick samples are only added to the testing set, which also
calves still contribute to the healthy samples at some point includes the remaining healthy samples not used for training.
in their lifetime. The 75% threshold of the minority label Furthermore, sick samples of the training set are limited to the
consist of 540 sick samples (720 × 0.75=540). A portion first five days of sickness, since symptoms may start changing
of the 35 sick calves is randomly selected, and added one with prolonged BRD and not be representative enough of the
by one to the training set until a total of at least 540 sam- presick stage.
ples is reached. Then, the same exact process is carried out Finally, both persistency scenarios evaluate sick calves to
for healthy samples. After this step, we may end up with assess whether they will relapse or become chronic. Specif-
542 sick samples from the first selection, and 551 healthy ically, each scenario consists of a binary machine learning
samples (samples are added selecting all samples from each model that includes labels from one BRD persistence status,
calf, rather than specifically selecting a number of samples, e.g., chronic, evaluated against sick calves, i.e., they recover
so we may end up with excess samples). Hence, we apply a within 14 days. Training is carried out using a 3-day moving
further balancing step by discarding 9 of the excess healthy window centered around the day of prediction. In fact, simi-
samples at random, which are simply removed and not used in larly to the presick scenario, due to the individual variation of
either training or testing. All remaining samples, i.e., 178 sick sickness over time, a small training window helps at perform-
samples and 479 healthy samples are added to the testing set, ing a more accurate detection since it is more indicative of the
without any need for balancing. sample being classified, despite losing training samples from
In the rest of the paper, when referring to a 75/25 train/test other days and thus reducing the number of training samples.
split, we refer to 75% of the minority label, as just described. By doing so, we are able to perform the early prediction
Intuitively, this tends to be less than the 75% of the whole of BRD persistency status in these calves within the first
dataset due to 1) 75% is applied to the minority label of a few days after sick diagnosis. Such contribution would prove
potentially imbalanced dataset, depending on the scenario, extremely helpful to domain experts, as this may increase the
and 2) a further balancing step is applied at the end which likelihood of faster BRD recovery and more efficient decision
leads to discarding a few extra samples. making, such as flagging the calf to be sold for beef once
antimicrobials are no longer circulating [5].
B. TRAINING MODELS
The methodology used to train a model has a high impact on VII. EXPERIMENTS
the prediction performance of a classification task. In addi- In this section we present the results of our solution and
tion, the definition of labels in our dataset is complex and compare its performance against the comparison approach.
dependent on a time component that needs to be taken into
account. Thus, we now present how our models were trained A. EXPERIMENTAL SETUP
under each scenario, and the labels being classified. We ran our experiments on real-field collected data consisting
First, the diagnosis scenario detects whether a calf is sick, of 159 calves over a period of 24 months, namely, a period
with outward signs of BRD, or healthy. In this scenario, from June 2018 to September 2019, and another period from
we simply split the dataset into training and testing consid- February 2020 to November 2020. Due to the pen size, only
ering all sick and healthy samples, as defined in Sec. V-B. 20 calves at a time were monitored simultaneously. Overall,
However, we remove presick samples to make sure that our 58 calves are healthy and never incurred BRD, 101 calves
models are not trained with ambiguous data, since these incurred BRD, of which 23 relapse and 47 become chroni-
samples may exhibit sickness without having outward clinical cally ill. Note that sick calves still contributed to the dataset
signs of BRD. with a portion of healthy samples, since they are not born
The presick scenario detects if a calf is healthy or presick, sick, and may recover after being sick. The dataset is divided
i.e., the calf is going to get sick but does not yet show in 7 cost groups, one per data collection method. For each
clinical signs. A clear and formal definition of the time range group, we have assigned a start-up cost and a daily cost. Each
that defines the early stage of BRD status does not exist, cost is considered on a per-calf basis. Daily costs include
which makes labeling these calves challenging. Besides, if an maintenance and operation costs where applicable. All costs

71172 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

are from [16], while costs associated with lung ultrasound accuracy, which would heavily consider the dominant label
exams are from [34]. in the testing set, this metric fairly represents both, while
For manual labor, we consider the start-up cost as 12 weeks resulting more intuitive than F1-scores.
of training, which means paying employee labor and man-
agement labor at $14/hour and $22/hour respectively, for B. PRELIMINARY RESULTS
20 hours a week. We also take into account turnover rates In Fig. 4, we show how CALF behaves and performs. For
at 38.8% [15]. For the daily cost, we consider management this experiment, the start-up budget was set to $15, 000, while
labor at $22/hour. Daily costs are considered per day and the daily budget was set to $2.5, hence only a portion of all
per calf. Hence we divide hourly rate by 12 (as a result features are available. Both plots show the accuracy obtained
of 5 minutes of manual labor per calf), for each manual by the algorithm and the respective size of the selected feature
feature group. Daily cost related to weight and ultrasound set across CALF’s iterations. In Fig. 4a the results of the full
were further divided by 3 since they were collected once execution are shown. Here, the trend of the length goes up
every three days. Monthly costs of maintenance were divided and down quite consistently throughout the 1000 iterations,
by 30 (i.e., the average number of days in a month), and showing that our algorithm effectively explored the features
included as daily costs. Then they were divided again by set, increasing and decreasing its size to largely explore
20 (number of calves in pen) to consider costs on a per- a wide variety of combinations. However, we see a small
calf basis. In summary, we reported costs for start-up and gap between iteration 500 and 600. Intuitively, this corre-
daily costs as follows: $89.50 and $0.38 for activity behavior, sponds to the stage where greedy actions are more likely to
$13500 and $4.35 for milk feeding, $6500 and $1.62 for occur, as opposed to random actions. In fact, after this stage,
grain feeding, $890.5 and $0.73 for weight, $11, 992 and the unstable trend of accuracy starts converging and a high
$1.83 for health exam, $0.0 and $6.23 for ultrasound, as well accuracy is maintained. In addition, Fig. 4b demonstrates a
as $0.0 and $0.0 for calves’ age and Immunoglobulin G (IgG), zoomed-in version of the full execution and thus allows us
respectively. to present further details of the CALF’s behavior. Here, it is
We tested a variety of machine learning algorithms in each clear that a high feature solution size does not necessarily
adopted scenario, which led us to pick out the algorithms lead to a higher accuracy, and thus further highlights the
to use. Specifically, we use are Gradient Boosting Classi- non-trivial complexity of the problem. Specifically, while
fier (GBC) for the diagnosis scenario, and Support Vector the solution size has a local maximum between iteration
Machines (SVC) for all other scenarios, with a window size 450 and 500, the local maximum of the accuracy score is
of 3 for segmentation. Furthermore, the following parameters found between iteration 300 and 350.
are set for CALF: L = 1000, α = 0.5, and β = 0.5. We tested
a few of combinations of these parameters, which we do not C. COMPARISON APPROACH—BOWEN ET AL.
include as they are out of scope. We believe that setting α We also compared our results against a recently published
and β to 0.5 allows us to perform fair evaluations where daily state-of-the-art approach that performs BRD prediction in
budget and start-up budget are considered in equal measure. calves with machine learning techniques [24]. We imple-
Different budgets are considered as varying parameters of mented the authors’ feature selection, which consisted
our experiments, namely, start-up budgets of $7, 500 and of removing and adding back each feature, and evalu-
$15, 000, and daily budgets of $0.5, $1.5, $2.5, $5.0, and ated their importance, calculated as their mean accuracy
$9.0. The results are reported after performing a 10-fold cross decrease/increase. Each feature was assigned an importance
validation with a 75/25 train/test split in diagnosis and presick score, and those with a value higher than or equal to 0.09 were
scenarios, and 70/30 split in the persistency scenarios due to utilized. We further extended their feature selection to include
the small number of samples. Finally, we provide some of the budgets. Specifically, we selected features, starting from fea-
machine learning models’ parameters used in the experiments tures with the highest importance score, until the maximum
which were set by performing a grid search to evaluate the allotted budget was met. Train/test split was performed in
best performing ones. For GBC, we used a binomial loss each scenario using the same methodology used for CALF,
function, the learning rate was 0.1, criterion to measure the and it was fixed at a 70/30 ratio, as indicated by the
quality of a split was the mean squared error with Friedman authors [24]. The adopted machine learning algorithm uses
improvement, the minimum number of samples to split a Random Forest, with a window size of 3. Some of the param-
node was 2, and the maximum an individual estimator had eters provided in the manuscript include the number of trees
was 3. For SVC, we use a radial basis function kernel, with a set to 3001, and the number of variables to use at each split
1
scalable value of gamma equal to N ×var (N is the number of set to 3. More details can be found at [24].
training samples and var their variance), a tolerance of 0.001,
and a regularization parameter of 1.0. Finally, results are dis- D. DIAGNOSIS SCENARIO
played using the metric of unweighted mean accuracy, which The first set of results on Fig. 5 shows a visual representation
means accuracy values were calculated individually for each to qualitatively appreciate how CALF performs, which uses
class, and then simply averaged. As opposed to a weighted feature selection CALF and prediction algorithm GBC. In this

VOLUME 11, 2023 71173

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

FIGURE 4. Trends of CALF’s accuracy and solution size. Dashed line is the solution size; continuous line is the accuracy.

following, we dive deeper into these results with quantitative

experiments for the detection of BRD diagnosis as well as
subsequent BRD stages.
Fig. 6 further evaluates the diagnosis scenario, looking at
the performance of CALF against the comparison approach
Bowen et al. by means of unweighted mean accuracy. Each
figure has a fixed start-up budget, and daily budgets vary
across the x-axis. Results show that CALF outperforms
Bowen et al. across all scenarios. Although the adopted
machine learning algorithms are different, CALF’s better per-
formance is to be attributed to its feature selection abilities.
In fact, GBC (CALF’s adopted machine learning models) and
Random Forest (Bowen’s adopted model) are intrinsically
very similar algorithms. However, CALF shows significant
improvements with a higher start-up budget, as shown in the
FIGURE 5. Visual representation of diagnosis prediction.
differences between Fig. 6b and Fig. 6a, with a significant
boost between daily budgets $2.5 and $5.0. The improvement
at these particular values resides on the fact that feature
groups such as feeding behaviors and health exams become
graph, we use one of the cross-validated trained models, available, and CALF is able to take full advantage of the
and a randomly selected pool of calves from the testing set, additional information, as opposed to the more traditional
using a start-up budget of $15, 000 and a daily budget of feature selection proposed by Bowen et al.
$5.0. The x-axis shows time information with daily precision, These figures also provide additional information.
while each line on the y-axis represents predicted labels and We notice a big performance increase when a daily budget
ground truth of a specific calf. Overall, it’s clear that our of $9.0 is reached, which occurs with both start-up budgets.
predictions fairly match the ground truth. Besides, often the Intuitively, this is due to the availability of ultrasound data,
algorithm predicted evidence of sickness behavior before which cost over $6.0 daily and thus can only be selected with
the ground truth, or outward symptoms of BRD, such is this budget. In fact, lung consolidation detected by ultrasound
the case of calf ID 1966 and 303. This agrees with animal tends to be extremely informative of BRD, since it represents
science research, calf behavior often changes before outward how pneumonia is diagnosed with the greatest sensitivity [7].
symptoms are visible, a phenomena known as the sickness We also noticed that very small or very high daily budgets do
behavior response [23], [35]. Furthermore, in the case of calf not benefit greatly from a higher start-up budget, as opposed
1080, the algorithm is able to grasp a slight improvement to intermediate daily budgets. On one hand, low daily budgets
which is due to the antibiotic shot given on the 7th day of do not gain more information with higher start-up budgets.
BRD (as described in Sec. IV-A), which results in a relapse On the other hand, ultrasound data, available at a high daily
of the disease itself that CALF is able to detect. In the budget with $0.0 start-up cost, is able to perform well on its

71174 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

FIGURE 6. CALF performance of the presick scenario against Bowen et al.

own without further information. However, most likely only Finally, CALF improves performance with a higher bud-
very large farms could afford this expertise at such a high get by effectively selecting features, and is able to create a
daily budget by diluting the cost through economies of scale. bigger gap with respect to the lower budget scenario than
Bowen et al., which are unable to further improve perfor-
E. PRESICK SCENARIO mance due to their limited feature selection.
In this experiment, we evaluate the presick scenario under
two different budget scenarios. The low-budget scenario
shown in Fig. 7a consists of a start-up budget of $7, 000 and F. PERSISTENCY SCENARIOS
daily budget of $1.5, while the high-budget scenario shown Fig. 8 shows the performance of the relapse and chronic
in Fig. 7b consists of a start-up budget of $15, 000 and a scenarios within the first five days of BRD, i.e., the prediction
daily budget of $5.0. Each data point shows the accuracy of detecting calves who are sick for more than 14 days and
of the algorithm regarding predictions in the days prior to 21 days respectively, within the first five days of sickness.
BRD diagnosis in the calves. Results are shown with respect In this experiment, we used a start-up budget of $15,000 and
to the maximum prediction accuracy of the algorithm in the a daily budget of $2.5. The relapse prediction of CALF has
diagnosis scenario. Hence, the 0.79 accuracy at one day prior moderate accuracy, showing the ability to predict about 2 out
to BRD diagnosis in Fig. 7a means that the accuracy is 79% of of 3 relapse calves. In addition, CALF outperforms the com-
the maximum prediction of BRD diagnosis. Thus, the actual parison approach. It is worth mentioning that our dataset only
accuracy of each data point can simply be multiplied by the has 23 relapse calves, which may not be enough to properly
maximum diagnosis at 0.88. Finally, note that each data point train a prediction model. Thus, our performance may improve
also includes the prediction accuracy of healthy samples and when more data is available. In addition, relapse calves tend
the displayed accuracy is the unweighted mean accuracy of to show behaviors that are quite similar to sick calves. Thus,
pre-sick and healthy labels. detecting 2 out of 3 calves within the first five days of sickness
The low-budget scenario in Fig. 7a involves activity behav- is a significant achievement. On the other hand, the prediction
ior data, weight, and calf’s age. Although prediction accuracy of chronic status calves is much higher, and performs signif-
is not very high, our model can still classify 2 out of 3 calves icantly better than the comparison approach. In addition, this
who are developing BRD up to 4 days prior to BRD diag- is the first work of its kind that demonstrates that relapse
nosis. In addition CALF outperforms Bowen et al. on each status BRD calves and chronic status BRD calves can be
day prior to BRD diagnosis. In the high-budget scenario identified within a few days after the initial antimicrobial
shown in Fig. 7b, CALF uses weight, calf’s age, and health intervention. In fact, with an overall accuracy of 0.8 within
exams, showing significantly improved algorithm perfor- the first 5 days after BRD diagnosis, a similar model would
mance. In fact, 4 days prior to BRD diagnosis the accuracy prove extremely helpful to farmers, who usually re-treat their
increased to 0.81 w.r.t. diagnosis prediction, corresponding relapsed calves once clinical signs of BRD re-emerge [25].
to a final accuracy of 0.72. On the other hand, 1 day prior This is even more so for potential chronic cases, as these
to BRD diagnosis, the prediction accuracy is 0.85 w.r.t. calves likely endure pain, have compromised productivity,
diagnosis prediction, corresponding to 0.76 final accuracy. and long BRD bouts with several antimicrobial interventions

VOLUME 11, 2023 71175

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

FIGURE 7. CALF performance of the presick scenario against Bowen et al.

prior to diagnosis [10]. While some studies have shown the

ability of machine learning paired with precision technology
to detect BRD and its early symptoms [23], [25], we propose
that this is the first approach to perform the prediction of BRD
persistency status in dairy calves. Additionally, our dataset
uniquely combines various precision technologies, along
with manual examination and ultrasonography, to diagnose
BRD in dairy calves. It is also worth noting that our dataset
includes a larger number of adopted precision technologies
compared to any other dataset available. Furthermore, since
a study has been published that uses machine learning to
indicate BRD in calves, we implement their solution and use
it as a comparison in our experiments [24].
Managing dairy cattle is a challenge as farms increase
in size and labor access becomes more limited. Thus, it is
FIGURE 8. CALF performance of the persistency scenario against important to consider the economic value of data collected
Bowen et al. by precision technology and the budget allotted per calf since
replacement dairy calves are the largest economic investment
on the dairy farm [16]. These sensors come with a collection
which are emotionally and fiscally taxing for dairy of features. However, researchers seldom consider the collec-
producers [36]. tive potential of multiple sensors to indicate disease status in
calves [37], [38]. Instead, each individual feature is assessed
VIII. RELATED WORK for associations with disease status in calves [10]. Hence,
Research on veal calves has observed that relapsed BRD we conducted our work to consider the cumulative impact of
calves had the slowest growth performance, and a reduced groups of features on BRD prediction. Furthermore, in order
carcass weight compared to their peers [36]. It is possible that to fully enable the benefits of precision technologies, it is
the reason for such reduced performance in relapsed BRD fundamental to design solutions that find the best tradeoff
calves is because there is a delay between the re-emergence between cost and accuracy.
of clinical signs of BRD and the re-intervention of antimi- In recent years, cost-sensitive feature selection has gained
crobials to resolve the disease. This is likely because calves attention in the medical field as a mean to reduce diagnos-
initially sick with BRD require time to approach convales- tic costs as well as to gain a greater understanding of the
cence and respond to the first antimicrobial intervention. problem [39]. However, previous approaches have not con-
Researchers have established that Bovine Respiratory Dis- sidered precision technologies with an abstract cost model;
ease in calves is associated with changes in feeding behavior in fact, research has oversimplified the important differences
and activity patterns recorded by sensors up to four days between feature selection and the cost to obtain those features

71176 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

in human models [19], [20]. Cost-sensitive feature selection Overall, we see that our solution outperforms a recent state-
is a problem where the term ‘‘cost’’ usually refers to either of-the-art approach under all scenarios and budgets. In fact,
misclassification costs, or test cost. While misclassification we achieved accuracy up to 0.88 in the diagnosis scenario,
costs is a well studied topic based on the impact of wrong and an early prediction of more than 2 out of 3 calves up
predictions [40], test cost is gaining attention, especially in to 4 days prior to BRD diagnosis, even under a low bud-
the medical field [39]. In fact, this approach allows to not only get scenario. This is a finding that greatly encourages the
reduce the cost of collecting the necessary data to diagnose adoption of precision livestock technology, such as pedome-
a disease, but also to gain a deeper understanding about the ters, since early treatments have shown to promote faster
problem itself [39]. Test cost generally refers to financial recovery. Furthermore, we are able to predict relapse calves
effort of data collection [19], as well as time to collect data and chronic calves within the first five days of diagnosis of
or overall required effort [20]. However, the current literature respiratory disease. In fact, 2 out of 3 relapse calves can be
considers abstract models that only consider a single general found by our machine learning algorithm, a result that we
budget, rather than a more refined and realistic model with a believe can be improved with the availability of a bigger
start-up budget and daily budget. Furthermore, existing works dataset. As a matter of fact, with a larger number of chronic
only consider an unrealistic cost per feature, rather than per calves, we observed that chronic calves were labeled with an
feature group. This is impractical, since paying the cost of a accuracy of 0.8 within the first five days of diagnosis.
feature from one sensor means that purchasing other features In conclusion, these encouraging results provided further
from the same sensor would not incur the cost of purchasing incentive to investigate methodologies for the early detection
the sensor again. of chronic calves in dairy farms. In our future work, we plan
A preliminary version of this work appeared in [23]. to implement a cost-based Pareto optimization, which is well
This paper extends the conference version by proposing sev- suited to navigate the feature space and improve the accuracy
eral improvements and additions. In fact, we implement the of our predictions.
early prediction of BRD persistency status, namely, relapse
calves and chronic calves, using our machine learning based REFERENCES
methodology, which has not been previously addressed in [1] W. J. Love, T. W. Lehenbauer, P. H. Kass, A. L. Van Eenennaam, and
the state of the art. Besides, the adopted methodology has S. S. Aly, ‘‘Development of a novel clinical scoring system for on-farm
diagnosis of bovine respiratory disease in pre-weaned dairy calves,’’ PeerJ,
been thoroughly revised and updated to better navigate the vol. 2, p. e238, Jan. 2014.
feature space when removing features, as demonstrated by [2] U. Dairy, ‘‘Dairy cattle management practices in the United States,’’
Fig. 4. Our accuracy function A() has been thoroughly USDA–APHIS–VS–CEAH–NAHMS, Central Jakarta, Indonesia,
Tech. Rep., 2014.
revised for persistency status predictions, proving to pro-
[3] N. J. Urie, J. E. Lombard, C. B. Shivley, C. A. Kopral, A. E. Adams,
vide more accurate presick scenario predictions as well. T. J. Earleywine, J. D. Olson, and F. B. Garry, ‘‘Preweaned heifer manage-
Besides, all experiments are based on a newly implemented ment on U.S. dairy operations: Part V. Factors associated with morbidity
realistic train/test methodology that was overlooked in pre- and mortality in preweaned dairy heifer calves,’’ J. Dairy Sci., vol. 101,
no. 10, pp. 9229–9244, Oct. 2018.
vious works [23], [24], [25]. Finally, we collected a lot [4] M. C. Cantor, D. L. Renaud, H. W. Neave, and J. H. C. Costa, ‘‘Feeding
more data, which we make publicly available through this behavior and activity levels are associated with recovery status in dairy
manuscript at [22]. calves treated with antimicrobials for bovine respiratory disease,’’ Sci.
Rep., vol. 12, no. 1, p. 4854, Mar. 2022.
[5] V. Welling, N. Lundeheim, and B. Bengtsson, ‘‘A pilot study in Sweden on
IX. CONCLUSION efficacy of benzylpenicillin, oxytetracycline, and florfenicol in treatment
In this work, we proposed the first work in the literature to of acute undifferentiated respiratory disease in calves,’’ Antibiotics, vol. 9,
no. 11, p. 736, Oct. 2020.
perform a cost-aware feature selection coupled with machine
[6] S. M. McGuirk and S. F. Peek, ‘‘Timely diagnosis of dairy calf respiratory
learning to guide the prediction of all stages of bovine res- disease using a standardized scoring system,’’ Animal Health Res. Rev.,
piratory disease in dairy calves, i.e., healthy, presick, sick, vol. 15, no. 2, pp. 145–147, Dec. 2014.
and BRD persistency status (both relapse, and chronic). Our [7] S. Buczinski, C. Faure, S. Jolivet, and A. Abdallah, ‘‘Evaluation of inter-
observer agreement when using a clinical respiratory scoring system in
approach relies on a cost-based feature selection that adopts pre-weaned dairy calves,’’ New Zealand Veterinary J., vol. 64, no. 4,
a realistic financial component, by using both start-up costs pp. 243–247, Jul. 2016.
and daily costs, and assigning costs to groups of features [8] A. Weersink, M. Massow, B. McDougall, and N. Bannon, ‘‘Re-examining
that belong to the same data collection process. We carried the implications of COVID-19 on the Canadian dairy and poultry sectors,’’
Can. J. Agricult. Econ./Revue canadienne d’agroeconomie, vol. 69, no. 2,
out experiments based on real-field data from 159 calves pp. 215–224, Jun. 2021.
collected over a period of 2 years. To this day, this is the [9] P. Rahimi, M. S. Islam, P. M. Duarte, S. S. Tazerji, M. A. Sobur,
first dataset that comprehensively considers health exams and M. E. El Zowalaty, H. M. Ashour, and M. T. Rahman, ‘‘Impact of the
COVID-19 pandemic on food production and animal health,’’ Trends Food
ultrasonography along with the highest number of precision Sci. Technol., vol. 121, pp. 105–113, Mar. 2022.
technologies. This dataset is also made publicly available [10] J. H. C. Costa, M. C. Cantor, and H. W. Neave, ‘‘Symposium review:
through this publication at [22]. We demonstrate the pre- Precision technologies for dairy calves and management applications,’’
J. Dairy Sci., vol. 104, no. 1, pp. 1203–1219, Jan. 2021.
diction performance of each scenario, and evaluated how
[11] J. I. Gargiulo, C. R. Eastwood, S. C. Garcia, and N. A. Lyons, ‘‘Dairy
performance is impacted by different start-up budgets and farmers with larger herd sizes adopt more precision dairy technologies,’’
daily budgets for data collection. J. Dairy Sci., vol. 101, no. 6, pp. 5466–5473, Jun. 2018.

VOLUME 11, 2023 71177

 E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

[12] C.-A. Duthie, J. M. Bowen, D. J. Bell, G. A. Miller, C. Mason, and [32] A. Virtala, ‘‘Morbidity from nonrespiratory diseases and mortality in dairy
M. J. Haskell, ‘‘Feeding behaviour and activity as early indicators of heifers during the first three months of life,’’ J. Amer. Veterinary Med.
disease in pre-weaned dairy calves,’’ Animal, vol. 15, no. 3, Mar. 2021, Assoc., vol. 208, no. 12, pp. 2043–2046, 1996.
Art. no. 100150. [33] E. K. Miller-Cushon, R. Bergeron, K. E. Leslie, and T. J. DeVries, ‘‘Effect
[13] T. R. Dunn, T. L. Ollivett, D. L. Renaud, K. E. Leslie, S. J. LeBlanc, of milk feeding level on development of feeding behavior in dairy calves,’’
T. F. Duffield, and D. F. Kelton, ‘‘The effect of lung consolidation, J. Dairy Sci., vol. 96, no. 1, pp. 551–564, Jan. 2013.
as determined by ultrasonography, on first-lactation milk production in [34] (Jan. 2022). Ontario Veterinary Medical Association. [Online]. Available:
Holstein dairy calves,’’ J. Dairy Sci., vol. 101, no. 6, pp. 5404–5410, https://www.ovma.org
Jun. 2018. [35] M. C. Cantor and J. H. C. Costa, ‘‘Daily behavioral measures recorded
[14] S. P. Dhoubhadel, ‘‘Precision agriculture technologies and farm profitabil- by precision technology devices may indicate bovine respiratory dis-
ity,’’ Tech. Rep., 2020. ease status in preweaned dairy calves,’’ J. Dairy Sci., vol. 105, no. 7,
[15] C. F. N.-A. Studies, ‘‘A national survey of hiring, compensation and pp. 6070–6082, Jul. 2022.
employee treatment practices on u.s. dairy farms,’’ Texas A&M Univ., [36] B. Pardon, K. De Bleecker, M. Hostens, J. Callens, J. Dewulf, and
College Station, TX, USA, Tech. Rep., 2020. P. Deprez, ‘‘Longitudinal study on morbidity and mortality in white veal
[16] A. Hawkins, K. Burdine, D. Amaral-Phillips, and J. H. C. Costa, ‘‘An eco- calves in Belgium,’’ BMC Veterinary Res., vol. 8, pp. 1–15, Dec. 2012.
nomic analysis of the costs associated with pre-weaning management [37] D. Sun, L. Webb, P. P. J. van der Tol, and K. van Reenen, ‘‘A systematic
strategies for dairy heifers,’’ Animals, vol. 9, no. 7, p. 471, Jul. 2019. review of automatic health monitoring in calves: Glimpsing the future from
[17] S. Zhang, Q. Su, and Q. Chen, ‘‘Application of machine learning in current practice,’’ Frontiers Veterinary Sci., vol. 8, pp. 55–71, Nov. 2021.
animal disease analysis and prediction,’’ Current Bioinf., vol. 16, no. 7, [38] D. Lovarelli, J. Bacenetti, and M. Guarino, ‘‘A review on dairy cattle farm-
pp. 972–982, Oct. 2021. ing: Is precision livestock farming the compromise for an environmental,
[18] R. Bennett, ‘‘The ‘direct costs’ of livestock disease: The development economic and social sustainable production?’’ J. Cleaner Prod., vol. 262,
of a system of models for the analysis of 30 endemic livestock dis- Jul. 2020, Art. no. 121409.
eases in great Britain,’’ J. Agricult. Econ., vol. 54, no. 1, pp. 55–71, [39] B. Remeseiro and V. Bolon-Canedo, ‘‘A review of feature selection meth-
Mar. 2003. ods in medical applications,’’ Comput. Biol. Med., vol. 112, Sep. 2019,
[19] M. Abdulla and M. T. Khasawneh, ‘‘G-forest: An ensemble method for Art. no. 103375.
cost-sensitive feature selection in gene expression microarrays,’’ Artif. [40] S. Benítez-Peña, R. Blanquero, E. Carrizosa, and P. Ramírez-Cobo, ‘‘Cost-
Intell. Med., vol. 108, Aug. 2020, Art. no. 101941. sensitive feature selection for support vector machines,’’ Comput. Oper.
[20] C. An and Q. Zhou, ‘‘A cost-sensitive feature selection method for high- Res., vol. 106, pp. 169–178, Jun. 2019.
dimensional data,’’ in Proc. 14th Int. Conf. Comput. Sci. Educ. (ICCSE),
Aug. 2019, pp. 1089–1094.
[21] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction.
Cambridge, MA, USA: MIT Press, 2018.
[22] E. Casella, ‘‘Precision livestock technology and health exams for bovine
respiratory disease,’’ Tech. Rep., 2023, doi: 10.21227/335v-mc16. ENRICO CASELLA (Member, IEEE) received the
[23] E. Casella, M. C. Cantor, S. Silvestri, D. L. Renaud, and J. H. C. Costa, B.S. degree in computer and telecommunication
‘‘Cost-aware inference of bovine respiratory disease in calves using preci- engineering and the M.S. degree in computer engi-
sion livestock technology,’’ in Proc. 18th Int. Conf. Distrib. Comput. Sensor neering from the University of Palermo, Italy, in
Syst. (DCOSS), May 2022, pp. 109–116. 2015 and 2018, respectively. He is currently pur-
[24] J. M. Bowen, M. J. Haskell, G. A. Miller, C. S. Mason, D. J. Bell, and suing the Ph.D. degree in computer science with
C.-A. Duthie, ‘‘Early prediction of respiratory disease in preweaning dairy the University of Kentucky, USA.
calves using feeding and activity behaviors,’’ J. Dairy Sci., vol. 104, no. 11, His research lies at the intersection of artificial
pp. 12009–12018, Nov. 2021. intelligence and cyber–physical systems in a vari-
[25] M. C. Cantor, E. Casella, S. Silvestri, D. L. Renaud, and J. H. C. Costa, ety of multidisciplinary areas, such as precision
‘‘Using machine learning and behavioral patterns observed by automated farming and smart grids. His research interests include leveraging machine
feeders and accelerometers for the early indication of clinical bovine
learning to exploit the patterns inherently present in these domains, applying
respiratory disease status in preweaned dairy calves,’’ Frontiers Animal
optimization techniques to carry out domain-specific objectives, designing
Sci., vol. 3, Jul. 2022, Art. no. 852359.
algorithm to evaluate trade-offs in resource-constrained environments, and
[26] E. Osuna and F. Girosi, ‘‘Reducing the run-time complexity of support
vector machines,’’ in Proc. Int. Conf. Pattern Recognit. Princeton, NJ,
developing data analysis pipelines. He has coauthored seven peer-reviewed
USA: Citeseer, 1998, pp. 1–10. scientific articles and two scientific abstracts.
[27] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu,
‘‘LightGBM: A highly efficient gradient boosting decision tree,’’ in Proc.
Adv. Neural Inf. Process. Syst., vol. 30, 2017.
[28] R. Dantzer and K. W. Kelley, ‘‘Twenty years of research on cytokine-
induced sickness behavior,’’ Brain, Behav., Immunity, vol. 21, no. 2,
MELISSA C. CANTOR received the B.S. degree in
pp. 153–160, Feb. 2007.
animal science from the University of Kentucky,
[29] M. C. Cantor, D. L. Renaud, and J. H. C. Costa, ‘‘Nutraceutical intervention
USA, in 2011, the M.S. degree in dairy science
with colostrum replacer: Can we reduce disease hazard, ameliorate disease
severity, and improve performance in preweaned dairy calves?’’ J. Dairy
from the University of Wisconsin–Madison, USA,
Sci., vol. 104, no. 6, pp. 7168–7176, Jun. 2021. in 2015, and the Ph.D. degree in animal science.
[30] V. Rhodes, E. G. Ryan, C. J. Hayes, C. McAloon, L. O’Grady, S. Hoey,
She is currently an Assistant Professor in
J. F. Mee, B. Pardon, B. Earley, and C. G. McAloon, ‘‘Diagnosis of precision dairy science with Pennsylvania State
respiratory disease in preweaned dairy calves using sequential thoracic University, USA. Her research interests include
ultrasonography and clinical respiratory scoring: Temporal transitions interaction of diseases; precision nutrition strate-
and association with growth rates,’’ J. Dairy Sci., vol. 104, no. 10, gies; and the environment with dairy cattle health,
pp. 11165–11175, Oct. 2021. performance, and welfare. During her Postdoctoral Fellowship at the Uni-
[31] W. J. Love, T. W. Lehenbauer, A. L. Van Eenennaam, C. M. Drake, versity of Guelph, which she completed in 2023, she focused on temporal
P. H. Kass, T. B. Farver, and S. S. Aly, ‘‘Sensitivity and specificity of on- changes in precision-based behaviors and changes in the microbial commu-
farm scoring systems and nasal culture to detect bovine respiratory disease nity in the lower gut surrounding diarrhea bouts in calves. Her extension
complex in preweaned dairy calves,’’ J. Veterinary Diagnostic Invest., program focuses on improving calf management on farms and improving
vol. 28, no. 2, pp. 119–128, Mar. 2016. the utility of precision technology data for the dairy.

71178 VOLUME 11, 2023

E. Casella et al.: Machine Learning and Optimization Framework for the Early Diagnosis of BRD

MEGAN M. WOODRUM SETSER received the JOAO H. C. COSTA received the B.S. degree
B.S. degree in animal science and the M.S. degree in agroecology and sustainable agriculture, and
in animal and food science from the University of the M.S. degree in agroecosystems with a focus
Kentucky, USA, in 2018 and 2021, respectively, on animal science from the Federal University of
where she is currently pursuing the Ph.D. degree, Santa Catarina, Brazil, in 2009 and 2012, respec-
with a focus on the relationship between calf per- tively, and the Ph.D. degree in animal science from
sonality traits and their relationship with health, The University of British Columbia, Canada.
performance, and behavioral measures from pre- He is currently an Associate Professor with the
cision technologies in the home environment. Her Department of Animal and Veterinary Sciences,
M.Sc. program focused on the development of an University of Vermont, USA. His expertise is in
on-farm personality assessment for dairy calves. precision dairy, and dairy cattle management related to behavior, nutrition,
She is an Early Career Scientist, who has coauthored three peer-reviewed one health, welfare, and sustainability. He has a very active research program
scientific articles and 26 scientific abstracts. Her research interests include supported by significant extramural funding, has received numerous faculty
dairy calf behavior, calf management, and precision dairy technology. and graduate student research awards, and has close to 100 refereed pub-
lications, besides being an active internationally and nationally recognized
speaker.

SIMONE SILVESTRI (Senior Member, IEEE)

received the M.S. and Ph.D. degrees in computer
science from the Sapienza University of Rome,
Italy, in 2006 and 2010, respectively.
He is currently an Associate Professor with the
Department of Computer Science, University of
Kentucky. He has published more than 80 papers in
international journals and conferences, including
IEEE TRANSACTIONS ON MOBILE COMPUTING, IEEE
TRANSACTIONS ON SMART GRID, IEEE INFOCOM,
and IEEE ICDCS. His research interests include cyber–physical–human
systems, the Internet of Things, energy management systems, terrestrial and
aerial mobile networks, and network management. He has served on the
organizing committee of several international conferences, and the technical
program committee of over 100 conferences.

VOLUME 11, 2023 71179