Implementation of multivariate cumulative sum control charts in mastitis and lameness monitoring

Invited review: Toward a common language in data-driven mastitis detection research

Sensor technologies for mastitis detection have resulted in the collection and availability of a large amount of data. As a result, scientific publications reporting mastitis detection research have become less driven by approaches based on biological assumptions and more by data-driven modeling. Most of these approaches try to predict mastitis events from (combinations of) raw sensor data to which a wide variety of methods are applied originating from machine learning and classical statistical approaches. However, an even wider variety in terminologies is used by researchers for methods that are similar in nature. This makes it difficult for readers from other disciplines to understand the specific methods that are used and how these differ from each other. The aim of this paper was to provide a framework (filtering, transformation, and classification) for describing the different methods applied in sensor data-based clinical mastitis detection research and use this framework to review and categorize the approaches and underlying methods described in the scientific literature on mastitis detection. We identified 40 scientific publications between 1992 and 2020 that applied methods to detect clinical mastitis from sensor data. Based on these publications, we developed and used the framework and categorized these scientific publications into the 2 data processing techniques of filtering and transformation. These data processing techniques make raw data more amendable to be used for the third step in our framework, that of classification, which is used to distinguish between healthy and nonhealthy (mastitis) cows. Most publications (n = 34) used filtering or transformation, or a combination of these 2, for data processing before classification, whereas the remaining publications (n = 6) classified the observations directly from raw data. Concerning classification, applying a simple threshold was the most used method (n = 19 publications). Our work identified that within approaches several different methods and terminologies for similar methods were used. Not all publications provided a clear description of the method used, and therefore it seemed that different methods were used between publications, whereas in fact just a different terminology was used, or the other way around. This paper is intended to serve as a reference for people from various research disciplines who need to collaborate and communicate efficiently about the topic of sensor-based mastitis detection and the methods used in this context. The framework used in this paper can support future research to correctly classify approaches and methods, which can improve the understanding of scientific publication. We encourage future research on sensor-based animal disease detection, including that of mastitis detection, to use a more coherent terminology for methods, and clearly state which technique (e.g., filtering) and approach (e.g., moving average) are used. This paper, therefore, can serve as a starting point and further stimulates the interdisciplinary cooperation in sensor-based mastitis research.

Do we automatically detect health- or general welfare-related issues? A framework

The early detection of health disorders is a central goal in livestock production. Thus, a great demand for technologies enabling the automated detection of such issues exists. However, despite decades of research, precision livestock farming (PLF) technologies with sufficient accuracy and ready for implementation on commercial farms are rare. A central factor impeding technological development is likely the use of non-specific indicators for various issues. On commercial farms, where animals are exposed to changing environmental conditions, where they undergo different internal states and, most importantly, where they can be challenged by more than one issue at a time, such an approach leads inevitably to errors. To improve the accuracy of PLF technologies, the presented framework proposes a categorization of the aim of detection of issues related to general welfare, disease and distress and defined disease. Each decision level provides a different degree of information and therefore requires indicators varying in specificity. Based on these considerations, it becomes apparent that while most technologies aim to detect a defined health issue, they facilitate only the identification of issues related to general welfare. To achieve detection of specific issues, new indicators such as rhythmicity patterns of behaviour or physiological processes should be examined.

Variable selection for monitoring sickness behavior in lactating dairy cattle with the application of control charts

The present observational study investigated the application of multivariate cumulative sum (MCUSUM) control charts by including variables selected by principal component analysis and partial least squares (PLS) regression to identify sickness behavior in dairy cattle. Therefore, sensor information (24 variables) was collected from 480 milking cows on a German dairy farm between September 2018 and December 2019. These variables were gathered in potentially different scenarios on farm. In total, data from 749 animals were available for evaluation. Variables were chosen based on the information of 499 cows (62 healthy; 437 sick) with 93,598 observations. The available diagnoses were collected together to form 1,025 sickness events. Hence, the different numbers of selected variables were included into the MCUSUM control charts. The performance of the MCUSUM control charts was evaluated by a 10-fold cross validation; hence, 90% of the original data set (749 cows) represented the training data, and the remaining 10% was used to test the training results. On average, the 10 training data sets included 124,871 observations with 1,392 sickness events, and the 10 testing data sets included, on average, 13,704 observations with 153 sickness events. The MCUSUM generated from the variables selected by principal component analysis showed comparable results in training and testing in all scenarios; therefore, 70.0 to 97.4% of the sickness events were detected. The false-positive rates ranged from 8.5 to 29.6%, and thus they created at least 2.6 false-positive alerts per day in testing. The variables selected by the PLS regression approach showed comparable sickness detection rates (70.0-99.9%) as well as false-positive rates (8.2-62.8%) in most scenarios. The best performing scenario produced 2.5 false-positive alerts in testing. Summarizing, both approaches showed potential for practical implementation; however, the PLS variable selection approach showed fewer false positives. Therefore, the PLS regression approach could generate a more reliable sickness detection algorithm, if combined with MCUSUM control charts, and considered for practical implementation.

Multivariate syndromic surveillance for cattle diseases: Epidemic simulation and algorithm performance evaluation

Multivariate Syndromic Surveillance (SyS) systems that simultaneously assess and combine information from different data sources are especially useful for strengthening surveillance systems for early detection of infectious disease epidemics. Despite the strong motivation for implementing multivariate SyS and there being numerous methods reported, the number of operational multivariate SyS systems in veterinary medicine is still very small. One possible reason is that assessing the performance of such surveillance systems remains challenging because field epidemic data are often unavailable. The objective of this study is to demonstrate a practical multivariate event detection method (directionally sensitive multivariate control charts) that can be easily applied in livestock disease SyS, using syndrome time series data from the Swiss cattle population as an example. We present a standardized method for simulating multivariate epidemics of different diseases using four diseases as examples: Bovine Virus Diarrhea (BVD), Infectious Bovine Rhinotracheitis (IBR), Bluetongue virus (BTV) and Schmallenberg virus (SV). Two directional multivariate control chart algorithms, Multivariate Exponentially Weighted Moving Average (MEWMA) and Multivariate Cumulative Sum (MCUSUM) were compared. The two algorithms were evaluated using 12 syndrome time series extracted from two Swiss national databases. The two algorithms were able to detect all simulated epidemics around 4.5 months after the start of the epidemic, with a specificity of 95%. However, the results varied depending on the algorithm and the disease. The MEWMA algorithm always detected epidemics earlier than the MCUSUM, and epidemics of IBR and SV were detected earlier than epidemics of BVD and BTV. Our results show that the two directional multivariate control charts are promising methods for combining information from multiple time series for early detection of subtle changes in time series from a population without producing an unreasonable amount of false alarms. The approach that we used for simulating multivariate epidemics is relatively easy to implement and could be used in other situations where real epidemic data are unavailable. We believe that our study results can support the implementation and assessment of multivariate SyS systems in animal health.

Methodological challenges to multivariate syndromic surveillance: a case study using Swiss animal health data

BACKGROUND

In an era of ubiquitous electronic collection of animal health data, multivariate surveillance systems (which concurrently monitor several data streams) should have a greater probability of detecting disease events than univariate systems. However, despite their limitations, univariate aberration detection algorithms are used in most active syndromic surveillance (SyS) systems because of their ease of application and interpretation. On the other hand, a stochastic modelling-based approach to multivariate surveillance offers more flexibility, allowing for the retention of historical outbreaks, for overdispersion and for non-stationarity. While such methods are not new, they are yet to be applied to animal health surveillance data. We applied an example of such stochastic model, Held and colleagues' two-component model, to two multivariate animal health datasets from Switzerland.

RESULTS

In our first application, multivariate time series of the number of laboratories test requests were derived from Swiss animal diagnostic laboratories. We compare the performance of the two-component model to parallel monitoring using an improved Farrington algorithm and found both methods yield a satisfactorily low false alarm rate. However, the calibration test of the two-component model on the one-step ahead predictions proved satisfactory, making such an approach suitable for outbreak prediction. In our second application, the two-component model was applied to the multivariate time series of the number of cattle abortions and the number of test requests for bovine viral diarrhea (a disease that often results in abortions). We found that there is a two days lagged effect from the number of abortions to the number of test requests. We further compared the joint modelling and univariate modelling of the number of laboratory test requests time series. The joint modelling approach showed evidence of superiority in terms of forecasting abilities.

CONCLUSIONS

Stochastic modelling approaches offer the potential to address more realistic surveillance scenarios through, for example, the inclusion of times series specific parameters, or of covariates known to have an impact on syndrome counts. Nevertheless, many methodological challenges to multivariate surveillance of animal SyS data still remain. Deciding on the amount of corroboration among data streams that is required to escalate into an alert is not a trivial task given the sparse data on the events under consideration (e.g. disease outbreaks).

Assessment of foot health and animal welfare: clinical findings in 229 dairy Mediterranean Buffaloes (Bubalus bubalis) affected by foot disorders

BACKGROUND

Lameness represents the third most important health-related cause of economic loss in the dairy industry after fertility and mastitis. Although, dairy Mediterranean Buffaloes (MB) and dairy cows share similar breeding systems predisposing to similar herd problems, published studies exploring its relevance and role in these ruminants are still rare and incomplete. The aims of this study were to describe the clinical findings of foot disorders (FDs) in dairy MB and their influence on animal welfare, determined by assessment of locomotion score (LS), body condition score (BCS) and cleanliness score (CS).

RESULTS

Of 1297 multiparous MB submitted to routine trimming procedures, 229 buffaloes showed at least one FD. The prevalence of buffaloes affected by FDs was 17.7 %, while motility and lameness indexes were 84.1 % (1091/1297) and 15.9 % (206/1297), respectively. Overgrowth was present in 17.0 % (220/1297), corkscrew claw in 15.8 % (205/1297), interdigital phlegmon in 0.9 % (12/1297), white line abscess in 0.8 % (11/1297), digital dermatitis in 0.1 % (1/1297) and interdigital hyperplasia in 0.1 % (1/1297). Simultaneous presence of FDs was recorded in 17.0 % of MB (221/1297): overgrowth and corkscrew claw occurred together in 15.8 % of cases (205/1297), overgrowth and interdigital phlegmon in 0.3 % (4/1297), overgrowth and white line abscess in 0.8 % (11/1297), digital dermatitis and interdigital hyperplasia in 0.1 % (1/1297). The presence of FDs was always associated with lameness (LS > 2), except from 23 MB with simultaneous overgrowth and interdigital phlegmon occurrence. The majority of MB within the under-conditioned group (95.5 %, 43/45) and all those with CS > 2 (122/122) had a locomotion score above the threshold of normality (LS > 2). Furthermore, foot diseases such as interdigital hyperplasia, white line abscess and digital dermatitis or interdigital hyperplasia seemed to occur more frequently associated with decreased BCS and increased CS scores.

CONCLUSIONS

This study describes for the first time the involvement of white line disease, interdigital phlegmona, digital dermatitis and interdigital hyperplasia in foot disorders of dairy Mediterranean buffalo and shows their association with an impairment of animal welfare.

Lameness Detection in Dairy Cows: Part 2. Use of Sensors to Automatically Register Changes in Locomotion or Behavior

Simple Summary

As lame cows produce less milk and tendto have other health problems, finding and treating lame cows is very importantfor farmers. Sensors that measure behaviors associated with lameness in cowscan help by alerting the farmer of those cows in need of treatment. This reviewgives an overview of sensors for automated lameness detection and discussessome practical considerations for investigating and applying such systems inpractice.

Abstract

Despite the research on opportunities toautomatically measure lameness in cattle, lameness detection systems are notwidely available commercially and are only used on a few dairy farms. However, farmers need to be aware of the lame cows in their herds in order treat themproperly and in a timely fashion. Many papers have focused on the automatedmeasurement of gait or behavioral cow characteristics related to lameness. Inorder for such automated measurements to be used in a detection system, algorithms to distinguish between non-lame and mildly or severely lame cowsneed to be developed and validated. Few studies have reached this latter stageof the development process. Also, comparison between the different approachesis impeded by the wide range of practical settings used to measure the gait or behavioralcharacteristic (e.g., measurements during normal farming routine or duringexperiments; cows guided or walking at their own speed) and by the differentdefinitions of lame cows. In the majority of the publications, mildly lame cowsare included in the non-lame cow group, which limits the possibility of alsodetecting early lameness cases. In this review, studies that used sensortechnology to measure changes in gait or behavior of cows related to lamenessare discussed together with practical considerations when conducting lamenessresearch. In addition, other prerequisites for any lameness detection system onfarms (e.g., need for early detection, real-time measurements) are discussed.

Lameness detection in dairy cattle: single predictor v. multivariate analysis of image-based posture processing and behaviour and performance sensing

The objective of this study was to evaluate if a multi-sensor system (milk, activity, body posture) was a better classifier for lameness than the single-sensor-based detection models. Between September 2013 and August 2014, 3629 cow observations were collected on a commercial dairy farm in Belgium. Human locomotion scoring was used as reference for the model development and evaluation. Cow behaviour and performance was measured with existing sensors that were already present at the farm. A prototype of three-dimensional-based video recording system was used to quantify automatically the back posture of a cow. For the single predictor comparisons, a receiver operating characteristics curve was made. For the multivariate detection models, logistic regression and generalized linear mixed models (GLMM) were developed. The best lameness classification model was obtained by the multi-sensor analysis (area under the receiver operating characteristics curve (AUC)=0.757±0.029), containing a combination of milk and milking variables, activity and gait and posture variables from videos. Second, the multivariate video-based system (AUC=0.732±0.011) performed better than the multivariate milk sensors (AUC=0.604±0.026) and the multivariate behaviour sensors (AUC=0.633±0.018). The video-based system performed better than the combined behaviour and performance-based detection model (AUC=0.669±0.028), indicating that it is worthwhile to consider a video-based lameness detection system, regardless the presence of other existing sensors in the farm. The results suggest that Θ2, the feature variable for the back curvature around the hip joints, with an AUC of 0.719 is the best single predictor variable for lameness detection based on locomotion scoring. In general, this study showed that the video-based back posture monitoring system is outperforming the behaviour and performance sensing techniques for locomotion scoring-based lameness detection. A GLMM with seven specific variables (walking speed, back posture measurement, daytime activity, milk yield, lactation stage, milk peak flow rate and milk peak conductivity) is the best combination of variables for lameness classification. The accuracy on four-level lameness classification was 60.3%. The accuracy improved to 79.8% for binary lameness classification. The binary GLMM obtained a sensitivity of 68.5% and a specificity of 87.6%, which both exceed the sensitivity (52.1%±4.7%) and specificity (83.2%±2.3%) of the multi-sensor logistic regression model. This shows that the repeated measures analysis in the GLMM, taking into account the individual history of the animal, outperforms the classification when thresholds based on herd level (a statistical population) are used.

The effect of routine hoof trimming on locomotion score, ruminating time, activity, and milk yield of dairy cows

The objective of this study was to quantify the effect of hoof trimming on cow behavior (ruminating time, activity, and locomotion score) and performance (milk yield) over time. Data were gathered from a commercial dairy farm in Israel where routine hoof trimming is done by a trained hoof trimmer twice per year on the entire herd. In total, 288 cows spread over 6 groups with varying production levels were used for the analysis. Cow behavior was measured continuously with a commercial neck activity logger and a ruminating time logger (HR-Tag, SCR Engineers Ltd., Netanya, Israel). Milk yield was recorded during each milking session with a commercial milk flow sensor (Free Flow, SCR Engineers Ltd.). A trained observer assigned on the spot 5-point locomotion scores during 19 nighttime milking occasions between 22 October 2012 and 4 February 2013. Behavioral and performance data were gathered from 1wk before hoof trimming until 1wk after hoof trimming. A generalized linear mixed model was used to statistically test all main and interactive effects of hoof trimming, parity, lactation stage, and hoof lesion presence on ruminating time, neck activity, milk yield, and locomotion score. The results on locomotion scores show that the proportional distribution of cows in the different locomotion score classes changes significantly after trimming. The proportion of cows with a locomotion score ≥3 increases from 14% before to 34% directly after the hoof trimming. Two months after the trimming, the number of cows with a locomotion score ≥3 reduced to 20%, which was still higher than the baseline values 2wk before the trimming. The neck activity level was significantly reduced 1d after trimming (380±6 bits/d) compared with before trimming (389±6 bits/d). Each one-unit increase in locomotion score reduced cow activity level by 4.488 bits/d. The effect of hoof trimming on ruminating time was affected by an interaction effect with parity. The effect of hoof trimming on locomotion scores was affected by an interaction effect with lactation stage and tended to be affected by interaction effects with hoof lesion presence, indicating that cows with a lesion reacted different to the trimming than cows without a lesion did. The results show that the routine hoof trimming affected dairy cow behavior and performance in this farm.