Can We Detect Mastitis earlier than Farmers?
Cathal Ryan, Christophe Guéret, Donagh Berry, Brian Mac Namee
CCan We Detect Mastitis earlier than Farmers?
Cathal Ryan, Christophe Guéret, Donagh Berry, Brian Mac NameeNovember 9, 2020
Abstract
The aim of this study was to build a modelling framework that would allowus to be able to detect mastitis infections before they would normally be foundby farmers through the introduction of machine learning techniques. In themaking of this we created two different modelling framework’s, one that workson the premise of detecting Sub Clinical mastitis infections at one Somatic CellCount recording in advance called SMA and the other tries to detect both SubClinical mastitis infections aswell as Clinical mastitis infections at any time thecow is milked called AMA. We also introduce the idea of two different featuresets for our study, these represent different characteristics that should be takeninto account when detecting infections, these were the idea of a cow differingto a farm mean and also trends in the lactation. We reported that the results forSMA are better than those created by AMA for Sub Clinical infections yet it hasthe significant disadvantage of only being able to classify Sub Clinical infectionsdue to how we recorded Sub Clinical infections as being any time a Somatic CellCount measurement went above a certain threshold where as CM could appearat any stage of lactation. Thus in some cases the lower accuracy values for AMAmight in fact be more beneficial to farmers.
Mastitis is one of the most common infections that harm dairy farms around theworld today with approximately 20-30% of a herd being affected around the world[9] with it resulting in many different effects such as reduced milk yield, veterinarycosts, and increased risk of early culling [2]. Mastitis infections can be further splitinto two seperate categories mainly Sub Clinical Mastitis Infections (SCM) and alsoClinical Infections (CM). They are mainly split on how they can be found by farmersor veterinarians, CM infections are mastitis infections which create swelling orclotting in the udder while SCM are seen as any mastitis infection that is only foundthrough the measuring of Somatic Cell Count (SCC) or the use of California MastitisTest.Many new decision support systems have been reported to help try to reduce theeffect of mastitis incidence or to detect these infections using different approaches.This can be broken into the idea of detecting infected cows compared to healthycows [11] or the detection of individual infections [7], [6], [15], etc. Many of thesestudies have certain downfalls such as them only including a small sample size of1 a r X i v : . [ q - b i o . Q M ] N ov ows [14], [15], using the same herd [6], [3], [16], [19] or the same year [13], [18] forthe entire dataset. These three examples are ways in which a study might be able toachieve over exagerated results due to the model learning a pattern that might notbe present in more general data. This was reduced in our study through the use ofboth multiple farms and years in our dataset while also including a large quantity ofboth clinical and sub clinical infections seen in Table 1.Within this study we use a wide range of features that are widely available tofarmers to help produce a modelling framework that could help reduce the impactof mastitis each year. We also try to improve on these base features through theintroduction of two sets of different feature ideas, the first being the calculation ofhow different a cows feature is to the mean value for that farm while the secondfeature is the addition of using time series data preparation approaches to describethe changes of features over time during a lactation. With these two sets of featureswe set up two modelling frameworks that could be used to determinemastitis infections before they occurred. The first modeling framework(named Somatic Cell Count Milking Alert) contained days at which Somatic CellCount (SCC) was recorded only and then we created a single alert that woulddetermine if there was a high enough probability that the cow would become infectedwith a Sub-Clinical Mastitis (SCM) infection the next time SCC was recorded, whilethe second modelling framework (named All Milking Alert) contained all days wherea cow was milked and a new alert was created for each of these which determined ifthere was a high probability that the cow would becomeinfected with either a SCM infection or a Clinical Mastitis (CM) infection within thenext 7 days.The contributions of this paper include the two different types of modellingframeworks namely AMA and SMA which show that using data widely available onIrish Dairy Farms. The rest of the paper is as followed, in Section 2 we introducethe different aspects included in this study, in Section 3 we introduce the differentresults associated with our study, in Section 4 we interpret the outcome of our resultsand then we conclude with Section 5 with a conclusion.Clinical Sub Clinical2791 18175Table 1: Total Amount of Mastitis Infections The data used within this study comprised of milking and health data provided fromCows situated in 7 different research farms around Ireland which can be seen inFigure 2.1 where each of the pointers points to a group of farms, this is due to the fact2hat some farms are closer in distance to others. The data contained information oncows born after 1st January 2010 and contained many different aspects of the cowshistory such as the milking characteristics for the cow, treatment records for the cow,the dry off and calving dates for each lactation, the farm each cow was located on,information on the genetics of the cow where available, clinical infection records forthe cow and also information on the weight of the individual cow.Due to the nature of how treatment administration could in some way affect themilking characteristic of a cow we decided to remove any milking observation thatwas recorded less than 7 days after an infection. To make sure we didn’t use thesame information when training and testing our model we decided to separate ourcombined data into two separate data sets, these being calving periods up to andincluding 2017 and calving periods within 2018.
The features used within this study comprised of both basic milking characteristicsat a cow level, the amount of clinical and subclinical infections for that cow, generalinformation about each milking the cow was recorded to have and the genetic meritfor that cow expressed as PTA.
The data used for this study included milk recordings for each cow. Cows in thisstudy were milked twice each day with different features being recorded at differentintervals. Total yield and maximum flow rate were individually milked at eachmilking every day that the cow was milked. The next most frequent recorded featureswere Fat %, Protein % and Lactose % which were recorded every 7 days with theirAM and PM recordings being measured at different days which is shown in Table 2.The least frequently recorded feature was Somatic Cell count (SCC) which was onlymeasured during AM milkings at times were Fat percentage, Protein percentage andLactose were recorded. Due to the fact that not all the features were measured each3ime the cow was milked we chose to keep the last observed value within a particularlactation as the non recorded feature.Table 2: AM & PM recordingsDays in Milk 10 11 12 13 14 15 16 17 18 19 20AM recordings Yes No No No No No No Yes No No NoAM fat 4 4 4 4 4 4 4 4.5 4.5 4.5 4.5PM recordings No No No Yes No No No No No No YesPM fat 3.5 3.5 3.5 4 4 4 4 4 4 4 4.5
One of the contributing aspects of this paper is the idea of if we can use how differenta cow is to the farm average to determine if the cow is set to change from healthy tosick in the coming days. This idea was constructed on each of the basic featureswhich led to 11 new features in our input features.
These features used the concept of how a cow’s milking characteristic changedduring its lactation period as a way of determining if the cow was set to changefrom healthy to sick in the coming days. For each milking feature we created itsmaximum, minimum, standard deviation, average and median value over the last15 and 30 days.
The data used for this study also included the dates at which a cow was said to havebecome infected with a CM or SCM infection. This information was used to createdummy features that could be seen as an illustration of how healthy the cow hasbeen since it first started milking. This was done by counting up the total amount ofClinical or sub clinical infections for each cow for its entire milking history and alsofor each lactation separately.
The data used for this study also included more basic information such as ’Daysin Milk’, ’Parity’, ’Month’. Days in Milk corresponded to the amount of days fromcalving to that milking record, while Parity corresponded to that particular lactationvalue for the cow and finally Month corresponded to the Month each milking recordwas recorded in.
The data used also contained information regarding the genetic aspect of the cowshistory through the use of predictive transmitting ability (PTA) for each cow. Three4ifferent values for each cow were given. These were a value for the first lactation,second lactation and then one for any lactation after the second lactation.
There are some aspects of both models that are the same which includes removingmilkings if they were less than 10 days since the cow had calved. This is talked aboutin many studies due to the idea that SCC is seen to be very unreliable for farmersuntil a certain period after calving, yet the amount of days at which is seen safe touse milkings is again not widely agreed upon. All other milkings were kept within thedata sets for model 1 while for model 2 a further reduction was considered. This wasin regards to any milking that occurred to soon after a previous infection of that type,the period chosen for this consideration was set to 7 days such that we could assumethat the pattern formed by the infection was removed from the data and thus couldbe assumed to be again within a healthy state. We also created each model withusing the features regarding infection history using both SCM and CM infectionsfor both model types. The reason behind this is the idea that while CM infectionsare much less common than SCM and thus wont change frequently enough duringa lactation or throughout a cows lifespan compared to SCM and thus SCM couldpossibly allow for larger subgroups in the data sets we are using.
Within this modelling framework we are trying to answer the question of if we canpredict if a cow will be classed as having a SCM infection the next time they getrecorded for SCC from the current SCC measurement and other relevant features.This as a result can only create alerts when SCC is being measured and thus wouldonly give a farmer an alert for a small subset of days that their cows get milked andalso is unable to check for CM infections using our current data preparation.
This model uses all the data that is available instead of only milkings for which SCCwas recorded. As a result the outcome of this model is quite significantly differentto that of the first model. The first difference is that unlike the first model which canonly classify SCM infections the second model can also classify CM due to the modelcreating a new alert at each time a cow was milked.The structure of our data for this modelling approach is illustrated in Table 3.From this table we can see this model type generates four new columns within ourdata. These being ‘Time Till Infection’, ‘Time Healthy’, ‘Early Detection’ and also‘Unsure Day’. These are calculated separately for clinical and sub clinical infections.‘Time Till Infection’ counts up the amount of days till the next infection of that typewhich if at a value between 1-7 is passed to the column ‘Early Detect’ as a 1 andis seen as the positive class label when creating the model. While ‘Time Healthy’counts up the amount of days since the last infection which if at a value between 0-7is passed to the column ‘Unsure Day’ as a 1. Any milking day that has a value of 15or ‘Unsure Day’ is removed from the data set for that particular infection type. Theresults of this exclusion can be seen in Table 4.Clinical Sub ClinicalMilk Recordings 11836 97430Table 3: Amount of Milk Recordings Removed for Clinical and Sub Clinical DatasetsThe reasons for selecting 7 days for the period of ‘Early Detect’ was due to the factthat in our data SCC was measured on average once a week and thus similarities canbe kept between this model and model 1 for SCM infections. While the reason forchoosing the period for the ‘Unsure Days’ was closely linked to the same idea but wedecided to include milking days that were seen as the infected milking day such thateven if we were only able to generate an alert on the last possible day it would stillbe atleast the day before the infection occured such that some precautions could beachieved within a farm. The addition of removing ‘Unsure Days’ led us the possibilityto have to remove a subset of a cows lactation period for longer than 7 days if thecows infection kept occuring which allowed us to make sure that any milkingsincluded in this model were either healthy milkings or milkings that could beassumed to be not related to past infections.
The algorithm used for this study was a Gradient Boosting Algorithm. This waschosen firstly due to it being in the family of Decision Tree algorithms which havealready shown to work within a wide range of areas in the literature such as [17],[4]. We then further decided to use a Gradient Boosting Algorithm due to the factthat it can discover many different types of patterns and results while still beingless complex than other modelling types such as Multi-Layered Perceptron whichallows in some way the end user to see which features were the most important inclassifying the problem at hand.
When there is a large class imbalance in a binary classification problem the algorithmbeing used may create results that are biased towards the class that is in the majority.There are many ways to overcome this with the two main contributors beingResampling and also Threshold Moving.
Resampling works on the idea of creating a more balanced data set of the differentclasses that are in your original data set. It can be split between Undersamplingand Oversampling, within both areas there a wide range of different algorithimsthat have been constructed to reduce the effect of having unequal class sizes [5],6 aysin Milk InfectionRecorded Time TillInfection TimeHealthy EarlyDetection UnsureDay
10 0 10 NA 0 011 0 9 NA 0 012 0 8 NA 0 013 0 7 NA 1 014 0 6 NA 1 015 0 5 NA 1 016 0 4 NA 1 017 0 3 NA 1 018 0 2 NA 1 019 0 1 NA 1 020 1 0 0 0 121 0 NA 1 0 122 0 NA 2 0 123 0 NA 3 0 124 0 NA 4 0 125 0 NA 5 0 126 0 NA 6 0 127 0 NA 7 0 1Table 4: AMA Structure[20], [1], [8] are just some examples. For this study we chose to use an Oversamplingapproach, this is due to the widely considered point that Undersampling has thecapability to lose a large quantity of information from the majority class while alsoreducing the training set size to a much smaller size and thus over fitting to theminority class can occur.Within the area of Oversampling there is a large amount of different algorithmsthat have been created to alter the class distribution. The main algorithim that isused in the literature is SMOTE [5] which creates new observations between twoobservations that are seen to be similar to each other. Its been shown that thisalgorithim has many drawbacks and as a result many new algorithims have comeout that try to handle these problems, one such algorithim is ADASYN [8] whichtries to focus on creating new observations that are both similar to its own smallerclass but also similar to the larger class. The general idea of oversampling can beseen in Figure 1 which shows a simulated data set before and after ADASYN hasbeen conducted.
Another way that the issue of Class Imbalance can be resolved is through the use ofThreshold Moving. This method involves the use of moving the default 0.5 probability7igure 1: Example of OverSamplingthreshold to a more valid value. There are two main methods in which this canbe accomplished which include the use of F1-score and Youden’s j-Statistic [21].Moving the threshold by using a F1-score relies on calculating the final results forRecall/Sensitivity and Precision through the following equation:2 ∗ Pr eci si on ∗ Recal lPr eci si on + Recal l (1)Moving the threshold by using Youden’s j-Statistic is computed by the followingequation: sensi ti vi t y + speci f i ci t y − In this section we will look at how well both models do at their respective goal andhow we measured the success of each model as this is different for both methods.8s the data we are working with in this study was seen as imbalanced whenwe are reporting result metrics from our models we decided against the use of themore general metrics such as Accuracy, in favour of other metrics such as Specificity,Sensitivity, Balanced Accuracy and also Geometric Mean which are all describedbelow.
Accur ac y = T P + T NT P + T N + F P + F N (3)
Sensi ti vi t y = T PT P + F N , Speci f i ci t y = T NT N + F P (4)
B al anced Accur ac y = Sensi ti vi t y + Speci f i ci t y
Geometr i c Mean = (cid:113) Sensi ti vi t y ∗ Speci f i ci t y (6)The reason for this is due to the fact that in the creation of Accuracy the overallimbalance of a data set is not taken into account thus with a dataset that has 90% ofthe observations belonging to the larger class we could achieve an overall accuracyvalue of 90% by simply predicting all observations as the larger class, this whilegiving us a high value for accuracy is pointless in the modeling of the smaller classas that is normally the class of interest.
The output from Model 1 is the easiest to interpret as we can simply look at theconfusion matrix which is seen in Table 5. The 4 different outputs of Table 5 can beseen as the class label that will be given the next time the cow gets tested for SCC. Wecan see the final results in Tables 6,7 and then also the Statistical Measures in Table 8.We can see from these three tables that this model structure is able to classify wellfor both cows that are going to be healthy or going to have a high SCC the next timeit gets recorded.
Predicted
Negative Positive
Actual
Negative Healthy Milking Found Healthy Milking not FoundPositive Infected Milking not Found Infected Milking FoundTable 5: SMA Classification Matrix
Predicted
Healthy Infected
Actual
Healthy 20462 2608Infected 671 2092Table 6: Model 1 Classification Matrix using SCM records9 redicted
Healthy Infected
Actual
Healthy 20031 3039Infected 602 2161Table 7: Model 1 Classification Matrix using CM records
Results
SCM records CM recordsBalanced Accuracy 82.21 82.52Geometric Mean 81.95 82.41Specificity 88.7 86.83Sensitivity 75.71 78.21Table 8: Model 1 Accuracy Metrics
It is possible to look at the results for this approach in the same way that is givenfor the first modelling approach which is given in Tables 9,10,11 and 12. Yet whenwe look at these tables more closely we can see that the results given by these aren’texactly what we want to showcase, to accomplish this we must alter the final resultsin a small way.To interpret the results outputted from AMA requires more preparation as eachmilking that is predicted doesn’t necessarily need to be used within the final outputof results. The idea behind this is due to the fact that we are creating alerts thatwould inform the farmer whether or not the particular cow is at risk of a mastitisinfection within the next 7 days. Thus if an alert was given when the cow is 20 Daysin Milk that said the cow is at risk and another alert was given at 21 Days in Milk thefarmer wouldn’t need the results from the later milking due to the farmer alreadybeing able to take the neccessary precautions. Thus we altered the idea of
Specificity to take this into account. This was done by first calculating the total amount ofinfections that were included in the final test dataset and then after the individualalerts were created we illustrated a infection being found if it had atleast 1 alertcreated within the 7 days before hand. This lead us to further be able to calculatethe earliest time at which each infection was found and as a result use that as a wayto track how the progress of infections were found as we got closer to the days of theinfections.Further improvements were also made to accomodate the alerts created forhealthy milkings (or those milkings that were predicted not to have any infectionswithin the next 7 days) throughout the year instead of just one calculated result.This allowed us to be able to see how our model predicted healthy outcomes as thelactation period continued.As with this model we were able to predict for both SCM and CM infections10e were left with two seperate results for both the modified
Sensitivity and also
Specificity . Predicted
Healthy Infected
Actual
Healthy 156602 30048Infected 2965 6194Table 9: Model 2 Classification Matrix for SCM model using CM records
Predicted
Healthy Infected
Actual
Healthy 193074 19575Infected 1030 906Table 10: Model 2 Classification Matrix for CM model using CM records
Predicted
Healthy Infected
Actual
Healthy 151097 35553Infected 2598 6561Table 11: Model 2 Classification Matrix for SCM model using SCM records
The final results for predicting SCM infections is shown in Table 13, 14 andalso Figure 2 which illustrates the progress of picking up infections as the timelineprogresses and then also the average proportion of healthy milkings correctlypredicted each day of the year using both SCM and CM infection history separately.
One addition was included within the models when trying to predict CM, this is to dowith the idea of how many days for each milking since SCC was last recorded. Thiswas not included within the SCM predicting models due to the idea that SCM canonly be found when SCC is measured and due to it being measured on average 7 daysapart we thought it might lead to overfitting which is not present when measuringfor CM due to them being found whenever the farmer notices them instead.The final results for predicting CM infections is shown in Table 16, 15 and alsoFigure 3.2.2 which illustrates the progress of picking up infections as the timeline11 redicted
Healthy Infected
Actual
Healthy 189893 22756Infected 1023 913Table 12: Model 2 Classification Matrix for CM model using SCM recordsDays till Infection 7 6 5 4 3 2 1Using Sub Clinical Records 62.90 74.63 77.21 79.18 79.72 80.08 80.72Using Clinical Records 59.37 70.86 73.07 75.69 76.66 77.41 78.08Table 13: Sub-Clinical Infectionsprogresses and then also the average proportion of healthy milkings correctlypredicted each day of the year using both SCM and CM infection history separately.
When we look at the results we can clearly see the main difference between SMAand AMA with regards to its capability of determining SCM infections is the fact thatSMA achieves a much lower rate of alerts that are said to be Positive when in factthe cow would be healthy which is seen as the total amount of FN. Yet we can alsosee that the Sensitivity value for SMA is higher than that for AMA in the majority ofcases. This leads us to the opinion that the inclusion of many other days reduces theeffect of classifying an infection further out from its onset. This might be as a resultof SCC only being recorded every 7 days alongside the majority of other featuresand thus with AMA the inclusion of the Time Series adapted feature set leads to theintroduction of many rows that apart from the yield and maximum flow rate valuesare exactly the same for 3 days in a row. This could potentially be a reason for whythere is not an increase when there is more information available for our model towork on. Thus it might be appropriate to think that our model could work betterif each milking feature was recorded more frequently than the average of 7 days inthis study. This would also allow us to use a smaller period to define the Time Seriesfeatures on as the values of 15 and 30 days were chosen to make sure that thesefeatures were being constructed on more than just one unique value.We can also see that the act of correctly finding CM is much harder than thatof finding SCM, in part this could come down to the fact that there was a muchlarger proportion of SCM infections within our data set in the first place but also CMinfections can come in many different forms mainly being Contagious orEnvironmental infections [10]. It has been show that these two subsets of mastitisinfections while still being mastitis occur in different ways but also are illustrated bydifferent results which could have resulted in our CM infections grouping themselves12CM History CM HistoryProportion of Healthy Correctly Classified 81.19 84.51Table 14: Healthy Milkings for SCM modelFigure 2: Timeline of SCM infectionsinto so called ’small disjoints’ [12] which is a major issue within imbalanced data.This occurs when the minority class is actually made up of a multitude of smallerclasses which might belong in some way to the same class but will still have largedifferences to other observations in that class. This problem seems to only comeabout when you deal with a small sample size and in most instances will disappearas the positive samples get bigger.
Within this study we tried to answer two separate questions, the first being if wecould predict SCM infections one SCC recording in advance and the second was tosee if we could predict SCM and CM infections at least 7 days before they would13CM History CM HistoryProportion of Healthy Correctly Classified 88.02 89.98Table 15: Healthy Milkings for CM modelDays till Infection 7 6 5 4 3 2 1Using Sub Clinical Records 50.35 55.87 59.78 61.05 62.45 63.79 66.55Using Clinical Records 49.28 55.87 57.65 59.29 61.05 62.41 63.48Table 16: Clinical InfectionsFigure 3: Timeline of CM infectionsbe detectable. From our results we can say that this has been accomplished with arelatively high accuracy for both questions. Yet even though the results were relativelyhigh there still lies areas that that the results could be improved upon. A main aspectof this is with regards to the large amounts of FN created. This was reduced whenwe go to the SMA model which only looked at days where SCC was recorded. Yet14he advantages of AMA which include the ability to classify CM infections mightoutweigh the disadvantage of having high false alerts created.
Acknowledgements
This publication has emanated from research conducted with the financial supportof Science Foundation Ireland (SFI) and the Department of Agriculture, Food andMarine on behalf of the Government of Ireland under Grant Number [16/RC/3835].
References [1] Sukarna Barua, Md. Monirul Islam, Xin Yao, and Kazuyuki Murase. Mwmote–majority weighted minority oversampling technique for imbalanced dataset learning.
IEEE Transactions on Knowledge and Data Engineering ,26(2):405–425, 2014.[2] D. Cavero, K.-H. Tölle, G. Rave, C. Buxadé, and J. Krieter. Analysing serial datafor mastitis detection by means of local regression.
Livestock Science , 110(1-2):101–110, 2007.[3] M.g.g. Chagunda, N.c. Friggens, M.d. Rasmussen, and T. Larsen. A model fordetection of individual cow mastitis based on an indicator measured in milk.
Journal of Dairy Science , 89(8):2980–2998, 2006.[4] Yung-Chia Chang, Kuei-Hu Chang, and Guan-Jhih Wu. Application of extremegradient boosting trees in the construction of credit risk assessment models forfinancial institutions.
Applied Soft Computing , 73:914 – 920, 2018.[5] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. Smote: Syntheticminority over-sampling technique.
Journal of Artificial Intelligence Research ,16:321–357, 2002.[6] Esmaeil Ebrahimie, Faezeh Ebrahimi, Mansour Ebrahimi, Sarah Tomlinson,and Kiro R Petrovski. A large-scale study of indicators of sub-clinical mastitisin dairy cattle by attribute weighting analysis of milk composition features:highlighting the predictive power of lactose and electrical conductivity.
Journalof Dairy Research , 85(2):193–200, 2018.[7] N.c. Friggens, M.g.g. Chagunda, M. Bjerring, C. Ridder, S. Hojsgaard, andT. Larsen. Estimating degree of mastitis from time-series measurements inmilk: A test of a model based on lactate dehydrogenase measurements.
Journalof Dairy Science , 90(12):5415–5427, 2007.[8] Haibo He, Yang Bai, Edwardo A. Garcia, and Shutao Li. Adasyn: Adaptivesynthetic sampling approach for imbalanced learning. , 2008. 159] Bjørg Heringstad, Gunnar Klemetsdal, and John Ruane. Selection for mastitisresistance in dairy cattle: a review with focus on the situation in the nordiccountries.
Livestock Production Science , 64(2-3):95–106, 2000.[10] Robert M Hyde and Peter M Down. Automated prediction of mastitis infectionpatterns in dairy herds using machine learning.
Scientific Reports , 10(1), 2020.[11] Robert M. Hyde, Peter M. Down, Andrew J. Bradley, James E. Breen, ChrisHudson, Katharine A. Leach, and Martin J. Green. Automated prediction ofmastitis infection patterns in dairy herds using machine learning.
ScientificReports , 10(1), 2020.[12] Taeho Jo and Nathalie Japkowicz. Class imbalances versus small disjuncts.
ACMSIGKDD Explorations Newsletter , 6(1):40–49, 2004.[13] C. Kamphuis, H. Mollenhorst, A. Feelders, D. Pietersma, and H. Hogeveen.Decision-tree induction to detect clinical mastitis with automatic milking.
Computers and Electronics in Agriculture , 70(1):60–68, 2010.[14] Heejin Kim, Younjeong Min, and Byoungju Choi. Real-time temperaturemonitoring for the early detection of mastitis in dairy cattle: Methods and caseresearches.
Computers and Electronics in Agriculture , 162:119–125, 2019.[15] R.m De Mol and W Ouweltjes. Detection model for mastitis in cows milked inan automatic milking system.
Preventive Veterinary Medicine , 49(1-2):71–82,2001.[16] I. Panchal, I.k. Sawhney, A.k. Sharma, and A.k. Dang. Classification of healthyand mastitis murrah buffaloes by application of neural network models usingyield and milk quality parameters.
Computers and Electronics in Agriculture ,127:242–248, 2016.[17] R. Punmiya and S. Choe. Energy theft detection using gradient boosting theftdetector with feature engineering-based preprocessing.
IEEE Transactions onSmart Grid , 10(2):2326–2329, 2019.[18] I.m.g.a. Santman-Berends, R.g.m. Olde Riekerink, O.c. Sampimon, G. VanSchaik, and T.j.g.m. Lam. Incidence of subclinical mastitis in dutch dairyheifers in the first 100 days in lactation and associated risk factors.
Journalof Dairy Science , 95(5):2476–2484, 2012.[19] Zhibin Sun, Sandhya Samarasinghe, and Jenny Jago. Detection of mastitis andits stage of progression by automatic milking systems using artificial neuralnetworks.
Journal of Dairy Research , 77(2):168–175, 2009.[20] José A. Sáez, Julián Luengo, Jerzy Stefanowski, and Francisco Herrera.Smote–ipf: Addressing the noisy and borderline examples problem inimbalanced classification by a re-sampling method with filtering.
InformationSciences , 291:184–203, 2015.[21] W. J. Youden. Index for rating diagnostic tests.