Mastitis Detection  in Dairy Cows using Artificial Intelligence ( AI )  Sensors

Using Sensor Data to Detect Mastitis Treatment in Dairy Cows

The identification of the important  diseases like Mastitis in  Dairy cattle by a  Vet or a normal person is a challenging task. Every disease has comprised of its symptoms, and the characterization of diseases by the symptoms requires a trained and skilled person. However, there is a requirement for a real-time monitoring system, that detects the disease based on input data, i.e., sensor and visual data. Currently, there are different technologies including IoT, AI, DL, ML, blockchain, and robotics that can deliver a digital solution with real-time analytics and also with advanced security. IoT sensors embedded in the cattle obtain sensory data, and the vision-based device obtains visionary data. Pre-trained AI/DL/ML model can be applied to this data, based on input attributes the model will detect the diseases and send the disease detected to the vision device. Robotics also assists to collect vision data during the feeding process. Drones are used for tracking and capturing the visuals of the cattle in an outdoor environment. All these data can log into the cloud and also for better visualization on the digital platform. In addition to this, blockchain empowers to enhance the security of the data and also distributed in peer-2-peer network for better treatment and diagnosis to the cattle from any location.

Effective animal health observation techniques need consistent, high-quality, and sensible data for decision-making. Various Technologies are used to monitor Animal health Integrating information from several sources can assist in early disease identification and response through animals as well as early outbreak control. In this section, we will discuss the different digital technologies that empower the delivery of the digital network for real-time monitoring and prediction.

(a) Internet of Things (IoT) IoT is a modern concept that allows electronic appliances and sensors to connect via the internet to improve our lives . Overall, IoT is an innovation that brings altogether a broad variety of sensors, intelligent devices, and smart systems. Additionally, it makes use of quantum and nanotechnology to achieve previously unimaginable levels of storage, sensing, and computing speed.

Early malady detection and treatment in lactating cows are essential in animal husbandry. The healthiness of dairy cattle is improved by early diagnosis and therapy that enables concentrated application as soon as is practical. Mastitis or the mammary gland is inflamed by udder mastitis, which is one of the most prevalent diseases that results in financial losses since it reduces milk supply and results in animal deaths and earlier culling . One or more udder quarters may be impacted by mastitis, which(mastitis) is broken down into two types: subclinical mastitis and clinical mastitis, the latter of which manifests symptoms. Clinical mastitis infections of the mammary gland are those that have visible symptoms such as hardness, pain, abdominal swelling, or redness. On the other hand, subclinical mastitis infection of the mammary gland does not result in any obvious changes in milk, udder, or rough appearance, making it challenging to find and treat. In such a case, the milk is tested in a laboratory to determine the diagnosis. Another way to classify mastitis is by the duration of the disease, which is referred to as acute or chronic mastitis. A timely diagnosis of clinical mastitis using highly precise, rapid techniques can help to maintain stock health while also reducing quantitative and qualitative milk production losses. The diagnosis of disease in dairy cows is fairly comparable to that in humans. However, because animals cannot express their emotions verbally, the initial diagnosis is fairly difficult. The development of mastitis detection technology, sanitary conditions in the herd, animal genetics are being put to use to promote mastitis resilience, and other techniques are now accessible to dairy cows to assist them attain and perpetuate a healthy udder medical issue. The best techniques for detecting intra-mammary infections in cows are bacterial analysis and PCR assay. These techniques are costly and time-consuming, thus they cannot be used for normal population-level data collection .

Machine learning (ML) has proven to be effective at diagnosing human diseases as of late. The disease screening model based on clinical medical information is gaining importance in initial diagnosis. Maini et al.  used 5 Machine Learning (ML) techniques, including k-nearest neighbours (KNN), Logistic Regression (LR), Random Forest (RF), Naïve Bayes (NB), and Adaptive Boosting (AdaBoost), leveraging South Indian medical information to anticipate early cardiovascular diseases. Zhao et al.  used an RF model to predict chronic renal disorders. These techniques have been shown to be effective however they rely on manually extracted features to function properly.

Deep learning (DL) has become more prevalent recently for use in malady diagnosis. In order to determine if a patient may develop Alzheimer’s disease, Ljubic et al.  combined recurrent neural networks (RNN) and Long short-term memory (LSTM) to construct a DL model. To be able to forecast healthcare trajectories from medical records, Pham et al. developed an LSTM model. These DL approaches work well for illness diagnosis since they automatically extract features from large amounts of labelled data. Unfortunately, the professionals’ lack of diagnosis knowledge prevents the DL approaches from successfully uncovering the unrecorded relationships between diseases and symptoms.

Artificial intelligence (AI) is intelligence manifested by machines, as opposed to the natural intelligence demonstrated by humans and animals. Machines mimic cognitive function associated with the human mind viz learning and problem assessment (Muthukrishnan et al., 2020). Artificial intelligence was invented as an academic discipline in 1959 and has tried many approaches during its lifetime. It includes brain stimulation, modeling human problem solving, logics, and databases of knowledge and imitation of human behaviour (Moran, 2007). Even after the establishment of AI as an academic discipline, it remains as non-significant scientific approach with limited practical application (Haenlein and Kaplan, 2019). The term Artificial Intelligence was stamped by John McCarthy (Crevier, 1993). The founders of AI were Allen Newell (CMU), Herbert Simon (CMU), John McCarthy (MIT), Marvin Minsky (MIT) and Arthur Samuel (IBM). Their students produced programmes that were describes as astonishing computers. They reportedly played better than average human, solved word problems in algebra and provided with logical theorems (Russell and Norvig, 2003). The founders of AI were optimistic and one of them; Herbert Simon believed that “machines will be capable, within twenty years of doing any work a man can do” (Simon, 1965). But, they failed to recognize the difficulty of some of the remaining tasks. The coming next years would be called as “AI winter” where functioning of research related AI was difficult due to difficulty in obtaining funds (Russell and Norvig, 2003). In 1980’s, research on AI was revived by commercial success of expert systems in which stimulation of knowledge and analytical power of human experts was done. It gradually restored its authentication in late 90’s and early 21st century by improving in functions like specific solution to problems related to logistics, data mining and medical diagnostics (Luger and Stubblefiled, 2004). Access to large number of data, algorithm improvements and faster programming of computers enabled advancement in machine learning. Machine learning and deep methods started to make a benchmark around 2010 (McKinsey & Company, 2017). The year 2015 was considered as a landmark year for artificial intelligence with increased number of software projects within Google that uses application of AI. There is increased demand of AI from sporadic usage in 2012 to current times based on its application (McKinsey & Company, 2017).

Mandates of Artificial intelligence

Information representation and reasoning: The Field of AI dedicated to represent information about the world in a language that can be utilised by computers to solve complex tasks (HayesRoth et al., 1983). ˆ Planning and automated scheduling: Unlike classical problems and their solutions, complex data must be discovered and optimised in multidimensional manner. For that AI planning/ scheduling is important and is concerned with the realization of strategies, sequences of action taken, execution by intelligent agents and autonomous robots (Malik, 2004). ˆ

Natural language processing (NLP): A subdivision of computer science, linguistics and AI which is concerned with interaction between human language and computer programming. They function in particular way to programme computers and processing of large amount of natural language data. Primary goal is making computer capable of understanding the contents of document presented. Then, it can accurately extract information from documents as well as categorization of the content themselves (Nadkarni et al., 2011).

Machine learning (ML): Process of study of computer algorithms that improve automatically through a large number of data and experience. It builds a model that utilises sample data, commonly referred as training data in order to make predictions without being explicitly biased to do so. Machine learning algorithms are used in a variety of applications such as in speech recognition, computer vision, email filtering and in medical field also, where it is unfeasible to develop conventional algorithm to do the needed task (El-Naqa et al., 2015).

Machine perception: Ability to use input information from detection sensors (such as microphones, cameras, wireless signals, sonar, radar and spectrum cameras) to deduce aspects of information in the world.

Application of AI in veterinary sciences

The National Animal Disease Referral Expert System (NADRES) of ICAR-NIVEDI is a system that works on combining and coordinating the alert and response mechanisms for the stake holders in prediction, prevention and control of animal disease threats (zoonotic ones also) through sharing of data, epidemiological studies and filed missions to asses and prevent outbreak, whenever needed. Combining livestock disease data and Artificial Intelligence techniques provide new opportunities to prevent outbreak and maintenance in the animal healthcare sector. From all the 31 AICRP centres, disease outbreak data is collected and is maintained in NADRES v2 (National Animal Disease Referral Expert System version 2). Two regression models, Generalized Linear Models (GLM) and Generalized Additive Models (GAM) and six machine learning algorithms, i.e. Random Forest (RF), Boosted Regression Tree (BRT), Artificial Neural Network (ANN), Multiple Adaptive Regression Spline (MARS), Flexible Discriminant Analysis (FDA) and Classification Tree Analysis (CTA) are used for disease modelling. The outcome of best fitted model/s were categorised into 6 risk levels-No risk (NR), Very low risk (VLR), Low risk (LR), Moderate risk (MR), High risk (HR) and Very high risk (VHR). Application of artificial intelligence for detecting left atrial enlargement on canine thoracic radiology (Li et al., 2021).

 Digital Technologies

Effective animal health observation techniques need consistent, high-quality, and sensible data for decision-making. Various Technologies are used to monitor Animal health. Integrating information from several sources can assist in early disease identification and response through animals as well as early outbreak control. In this section, we will discuss the different digital technologies that empower the delivery of the digital network for real-time monitoring and prediction.

• Internet of Things (IoT) IoT is a modern concept that allows electronic appliances and sensors to connect via the internet to improve our lives . Overall, IoT is an innovation that brings altogether a broad variety of sensors, intelligent devices, and smart systems. Additionally, it makes use of quantum and nonotechnology.
• Artificial Intelligence The general term “AI” describes the use of a computer to simulate intelligent behavior with the least amount of human involvement. It is a modern method of utilizing machines to perform muscle operations and illustrate complex problems in a “cognitive” manner. AI is very important in exhibiting intelligent behavior, learning, demonstrating, and advising the user. The combination of training, perception, problem solving, and tailoring new solutions to the system is a larger definition of artificial intelligence .
• Machine learning ——-Machine learning (ML) is a modern AI application that encourages the existence of giving machines entrance to data for more to improve the human design and simply learning them for themselves . When big data, data science, and analysis are mentioned, intelligence and ML are frequently combined. ML is a very efficient solution for dealing with such large amounts of data in multinational industries. One of the most important technological approaches to AI is ML, which has served as the foundation for many recent developments and commercial applications .
• Cloud Computing ———-Cloud computing has arisen from the distributed software architecture intending to provide hosted services over the internet, and it can be an efficient alternative to owning and managing computer resources and applications for many enterprises . The four major cloud models are infrastructure as a service (IaaS), container as a service (CaaS), software as a service (SaaS), and platform as a service (PaaS), It also improves data security, data and application access efficiency, Quality of Service (QoS), lowers operating costs.
• (e) Edge Computing ———Edge computing provides restricted and decentralized infrastructure, creating essential resources edge to data sources and avoiding the need to transport data to a centralized cloud . Edge Computing allows data created by various IoT items to be processed at the network’s edge rather than being transmitted to a remote centralized or distributed Cloud platform. Edge computing decreases latency, and load on a network core and enhances protection through storage data in infrastructure.
• Fog Computing Fog Computing (FC) facilitates filtering, analysis, and data handling, at the edge of a network, pushing resources down to IoT devices. Depending on the desired QoS, FC enables computation and storage resources to disperse along the pathway between cloud computing and IoT devices . Fog computing architecture encompasses of cloud in the upper, devices with extremely minimal intensity at the bottom, and fog nodes servers in the center.
• (g) Robotics The utilization of mobile robots has grown dramatically in recent years. They can now be found in a variety of locations, including industrial, domestic, educational, and healthcare setting . The robot should be capable of creating an environment model, evaluating its present location and orientation inside the environment using this model, and navigating throughout the environment to reach the goal spots.
• (h) Drones Drones are the airborne vehicle that encompasses unmanned aerial vehicles (UAVs) able to commute thousands of kilometers and small drones . Drones are aerial vehicles that do not fly autonomously, by human operators, and can deliver harmful or safe payloads. In recent years, the most interest has been focused on flying robots for planetary exploration, military surveillance, and search-and-rescue missions. Drones benefit from the ability to operate a broad range of missions such as patrolling, reconnaissance, protection, and aerology.
• (i) Bigdata Big data refers to big data collections with large, diversified, and complicated structures that are challenging to store, analyze, and visualize for subsequent processes or results . Big data analytics refers to the technique of researching huge amounts of data in order to uncover hidden patterns and hidden relationships. This important information for businesses or organizations can assist them to obtain richer and deeper insights and gain a competitive advantage.
• (j) Blockchain Blockchain is a novel technology that is built on a distributed data structure that is shared over a decentralized network . Blockchain provides excellent protection against information manipulation. The blockchain serves as a distributed ledger system. It works in tandem with IoT to allow machine-to-machine interactions. It employs a database containing a collection of transactions. These transactions are verified and recorded in a centralized ledger .
• (k) Robotic Process Automation
• Robotic Process Automation (RPA), such as AI and ML, is a technology that automates jobs. The use of software to automate corporate operations such as application processing, transaction processing, data management, and even email response is known as robotic process automation (RPA). RPA automates repetitive tasks that were previously performed by humans .

Machine learning and its functioning

Machine learning (ML) is the process of utilising mathematical models of data to make computer learn without direct instruction given. With the application of algorithms, it helps in identification of patterns within the data and thus creates a model to make future predictions and decisions. With experience and provision of large number of data, accuracy in machine learning achieved. It requires human engineering and domain expertise to design information extractor that can transform raw data into suitable representation from which algorithm model can learn and detect patterns .

Supervised machine learning: Algorithms make predictions based on a set of labelled examples from the past experience. It is useful when it’s known what outcome should be. It helps learner to identify the similarities and differences when object to be classified have many variable qualities within their own domain and still have fundamental property which identifies them (Cherkassky and Mulier, 2007).

Unsupervised machine learning: Algorithm allows machine to work on unlabelled data and discover pattern on its own that was previously undetected. It is useful when no information related to outcome is known (Cherkassky and Mulier, 2007).

Reinforcement learning: Algorithms that learn from outcomes and decide which action to take next. Algorithm receives feedback that helps it to determine whether the choices it made were correct, indecisive or incorrect. Here a part of the data is labelled while other parts are unlabeled. In such case, the labelled one is used to provide aid in learning of unlabelled part. This kind of learning closely emulates with the development of skills by human and most of the processes in nature (Sutton and Barto, 1998).

Steps in machine learning: The steps involved in machine learning; objective definition, collection of data, preparation of data, selection of algorithm, training of model, testing of model, prediction and then deployment of data (Kotsiantis et al., 2007). Selection of algorithms in machine learning

1. Linear regression algorithm: It comes under Supervised Learning technique, used for solving regression problems only. Helps in prediction of continuous dependent variable with the aid of independent variables. Primary goal is to find the best possible line that can predict accurately the outcome of continuous dependent variable. If single independent variable is utilised for prediction then it is termed as single linear regression whereas use of multiple variables form multiple linear regression. The algorithm establishes the best possible relationship between dependent and independent variables to make linear regression line (Bishop and Nasrabadi, 2006).

2. Logistic analysis: Logistic analysis is used to predict the dependent variable which is categorical in nature with the help of independent variables. Output of logistic regression lies between 0 to 1 only. Use of activation function (sigmoid function) to possibly weigh the sum of inputs and map the values between 0 and 1. Curve obtained in such function is known as sigmoid curve. This analysis is based on the Maximum likelihood estimation and according to it, the observed data should be most probable.

3. Decision tree algorithm: Decision Tree algorithm is a kind of supervised learning technique used for classification and regression problems, preferably used for classification. It is represented as tree-based classifier in which internal nodes represent features of dataset, branches depict the decision rule and each leaf node presents the outcome. The decision node used to make decision and have multiple branches whereas leaf nodes are outputs of those decisions and have no further branches. It represents all the possible outcomes of a problem graphically, based on the given conditions (Charbuty and Abdulazeez, 2021).

4. Random forest algorithm: One of the popular machine learning algorithm which belongs supervised learning technique. Represents both the classification and regression problems in machine learning, based on concept of ensemble learning in which multiple classifiers combines to solve complex set of data and meanwhile improves the performance of the model. It contains number of decision tree of various subsets of given complex data and takes the average to enhance the accuracy of predictive model. Instead of single decision tree, it relies on multiple trees and takes prediction from each based on majority outcomes therefore predicts the final output (Ren et al., 2017).

5. Support vector algorithm: Popular algorithm of supervised learning technique used for regression and classification problems. Main goal is to create a decision boundary that can separate n-dimensional space into subsets that can be easily put by new data point in future in correct category. This decision boundary is called as hyperplane and vector machine selects the extreme vectors which creates these hyperplane. The extreme vectors are known as support vectors and therefore, the algorithm is coined as Support Vector machine (Somvanshi et al., 2016).

6. K-nearest algorithm: Is an Unsupervised Learning algorithm, which groups the unlabeled dataset into numerable different clusters. Each cluster is associated with a centroid. The main aim is to minimise the sum of distances between data point and their corresponding clusters. It performs mainly two functions; determination of best values for K points/centroids by an intrusive/repeated process and assigning each data points to its nearest centroid (those points nearest to particular k-centre creates a cluster) (Zhang, 2016).

Convolutional neural network:

The Convolutional Neural Networks (CNN) is one of the most notable approaches in deep learning in which multiple layers are trained in a manner that has been found very (LeCun et al., 1998). It’s a type of deep-learning algorithm which is designed to process information that doesn’t exhibits natural spatial variance (e.g., visuals; whose meaning doesn’t change during translation). It doesn’t require human supervision, automatically detects the significant features within given data (Alzubaidi et al., 2021). In ImageNet Large Scale Visual Recognition Challenge (ILSVRC), the most popular CNN architectures were proposed by top competitors. ImageNet is a project that aims to develop a widespread database of visual information that can be used by researchers in the field of object identification and recognition. AlexNet was the first known model to improve image classification programme considerably, with an error rate of 16.4% using the ImageNet dataset. Whereas, VGG16 considered as preferred choice for extracting features from images by the majority of researchers in visual recognition pattern community (Monshi et al., 2020).

Understanding of Deep learning and its principle of functioning

DL (Deep learning) is a promising subfield of machine learning, composed of multiple layers, uses raw data as input, and improves the representations of data. DL in which a machine is fed with raw data and develops its own representations needed for pattern recognition. Composed of multiple layers of representation and are typically arranged sequentially. They are composed of a large number of nonlinear, primitive operations in which raw data is put into representation layer in the beginning and later is fed onto next layers, transforming into abstract like representation (Bordes et al., 2012). Deep learning algorithms are classified into categories; Convolutional neural network, Restricted Boltzmann Machines, Auto encoder and Sparse Coding.

Convolutional neural network: The Convolutional Neural Networks (CNN) is one of the most notable approaches in deep learning in which multiple layers are trained in a manner that has been found very (LeCun et al., 1998). It’s a type of deep-learning algorithm which is designed to process information that doesn’t exhibits natural spatial variance (e.g., visuals; whose meaning doesn’t change during translation). It doesn’t require human supervision, automatically detects the significant features within given data (Alzubaidi et al., 2021). In ImageNet Large Scale Visual Recognition Challenge (ILSVRC), the most popular CNN architectures were proposed by top competitors. ImageNet is a project that aims to develop a widespread database of visual information that can be used by researchers in the field of object identification and recognition. AlexNet was the first known model to improve image classification programme considerably, with an error rate of 16.4% using the ImageNet dataset. Whereas, VGG16 considered as preferred choice for extracting features from images by the majority of researchers in visual recognition pattern community (Monshi et al., 2020).

In general, mastitis is regarded as being the most prominent production disease in developed dairyproducing countries, both in terms of incidence as well as in economic consequences (Hogeveen et al., 2019). Mastitis affects 2 milk quality measures used broadly in dairy-producing countries: SCC and, if mastitis is clinical, visibly abnormal milk. On dairy farms working without an automatic milking system (AMS), a wellestablished method to detect clinical mastitis (CM) is to strip milk before unit attachment and to check the foremilk for abnormalities (Rasmussen, 2005; Rasmussen and Bjerring, 2005). The development and use of automated mastitis detection systems has received much attention since the proliferation of AMS during the mid-1990s (Frost et al., 1993), and currently AMS are installed on both smaller farms (Jacobs and Siegford, 2012) and an increasing number of larger farms. Automated detection of mastitis is an integral part of automatic milking to ensure milk quality and maintain animal welfare through the prompt attention to cows with clinical (painful) mastitis. The demands for CM detection models have been defined to be 80% sensitivity and 99% specificity (ISO, 2007). There is a large variation in the scope and quality of the scientific literature on the use of sensors and algorithms of CM detection models. As a consequence of the suboptimal performance of current CM detection systems, farmers either have to check many cows per day (false-positives, which are a consequence of less than perfect specificity) or have to accept that a lower proportion of cows with CM will be detected in time (false-negatives, which are a consequence of relatively low sensitivity). Regarding the detection of CM, it seems that farmers prefer a lower sensitivity over a lower specificity (Mollenhorst et al., 2012). Until now, research on automated mastitis detection systems has been aimed at detection of CM, using methods from the visual mastitis detection era as the gold standard. Also, in the advice provided to farmers, the importance of visual confirmation of CM takes a central position. Through the technological progress that is being made, improvements in the development of sensors and algorithms are to be expected (FadulPacheco et al., 2021; Slob et al., 2021). However, due to the nature of mastitis as a complex of different types of infections interacting with the cow and resulting in variable subclinical and clinical effects, the improvement in the overall performance of sensor systems in the detection of CM in general will be challenging. In a sensor-based detection system aimed at CM, a detection alert that is a false-positive may very well be a case of subclinical mastitis (SCM), sometimes even with a considerable increase in SCC but without clinical signs. Given these observations, we can argue that there is a need for a novel approach to the detection and management of mastitis supported by sensor systems. Mastitis situations should be defined from a temporal urgency to intervene rather than to be solely based on their comparability to nonsensor udder health indicators (the old paradigms). In the context of the preceding argument and reasoning from a farm management perspective, we can define 4 different situations where management action is needed and where sensor systems may support the farmers’ management: 1. Cows with severe CM needing immediate attention; 2. Cows with SCM, mild, or moderate CM not needing immediate attention; 3. Cows needing attention at drying off; and 4. Monitoring of udder health at the herd level. It is our hypothesis that the sensor systems with algorithms developed for specific mastitis situations may be more beneficial for the improvement of udder health than the use of novel sensors or algorithms that indiscriminately try to detect CM, as is the current approach. The redefinition of mastitis situations based on the timing and urgency of potential intervention has consequences for the evaluation and demands on the performance of sensor systems.

SENSOR SYSTEMS TO SUPPORT MASTITIS MANAGEMENT

Intramammary infections result in decreased milk production and compositional changes in the milk that vary with the intensity and duration of the infection. These changes are associated with the inflammatory response as a result of bacteria entering and multiplying in the udder (Harmon, 1994), leading to compositional changes linked to the production of numerous mediators of inflammation (Wellnitz and Bruckmaier, 2012) and to the change in blood capillary permeability due to mastitis (Kitchen, 1981). An important compositional change consists of the influx of polymorphonuclear neutrophil leukocytes into the mammary tissue, making up a large proportion of the SCC (Kelly et al., 2000; Wickstrom et al., 2009), forming the basis of the use of SCC as a test to monitor udder health (Harmon, 1994; Dufour and Dohoo, 2013; Damm et al., 2017). More recently, these changes in milk composition were used to develop sensors aimed at automated inline or online detection of CM (Hogeveen et al., 2010; Martins et al., 2019).

Commercially Available Sensors

A large number of sensors are currently commercially available. Electrical conductivity (EC) is the measure of the resistance of a material to an electric current and is linked to a change in blood capillary permeability due to mastitis. For decades, this change in EC has been used as indicator for CM (e.g., Nielen et al., 1992; Hamann and Zecconi, 1998). Following mastitis enzymatic reactions, l-Lactate dehydrogenase will appear as part of the cow’s innate immune response against infection (presence of udder pathogens) and changes in cellular membrane (Chagunda et al., 2006; Friggens et al., 2007), reflecting the host response to an IMI rather than the IMI itself (Jorgensen et al., 2016). A sensor system based on inline measurements of l-Lactate dehydrogenase is commercially available. Sensor systems are also available to measure a change in color, which is a visible aspect of abnormal milk, mostly due to CM. The principle of the color measurement sensor is based on the reflection of light generated by a light-emitting diode (Ouweltjes and Hogeveen, 2001) or light transmission instead of light reflection (Whyte et al., 2004). More recently, 3 sensor systems measuring SCC became commercially available. Two systems are measuring SCC indirectly, either based on gel formation of the milk (comparable to the California Mastitis Test; Deng et al., 2020) or physical measurements in the milk flow. The third SCC sensor is based on staining of a milk sample and optical counting of the number of cells by fluorescence (Dalen et al., 2019a). In addition to sensor systems directly aimed at measuring mastitis indicators, many other sensor systems are on the market that measure one or more aspects that may serve as support to more specific mastitis sensors (Caja et al., 2016), activity monitors or behavior sensors (e.g., Van Hertem et al., 2016), milk production and constituent sensors, such as the commonly available electronic milk meters and fat and protein sensors, location sensors (e.g., Barker et al., 2018), rumination measurement (e.g., Grinter et al., 2019; Hamilton et al., 2019), temperature sensors (e.g., Kim et al., 2019), rumen pH sensors (e.g., Doroodmand et al., 2016), automated body condition scoring systems (e.g., Spoliansky et al., 2016; Mullins et al., 2019), and BW measuring systems (e.g., Maltz et al., 1997).

Sensor Systems

The sensors can be seen as the basic element of sensor systems (Figure 1; Rutten et al., 2013), delivering data regarding the composition of milk or the physiological status of a cow. By using computational algorithms, these data can be connected to events that are of interest to herd managers and thus, can be converted to useful management information. Regarding mastitis management, algorithms are aimed at the identification of a deviation from normality that could be predictive of mastitis. As a further step, diagnosis can be carried out, defined as judgment about the presence of a particular illness after an appropriate diagnostic procedure has been performed. Algorithms to detect events of interest may use data from one or more sensors, potentially combined with data from other (farm) sources (Dominiak and Kristensen, 2017; Slob et al., 2021). The algorithms are expected to detect deviations from normality that could be predictive of a specific status that is of interest to the decision maker. For most sensor systems this prediction of an abnormal status (alert) needs to be confirmed by either clinical examination or secondary testing (e.g., bacteriological culture or PCR of one or more quarter milk samples). When an abnormal status is confirmed, this would constitute a diagnosis. The diagnosis can be combined with other knowledge (such as economic information), to produce advice for the farmer (decision support). Decision support can be very complex but can be compiled as well into standard operating procedures (SOP) on what to do with the diagnosis that has been made. Finally, a decision needs to be made and acted upon by the farmer. Based on preprocessing of data classification, alerts are generated (van der Voort et al., 2019) that always have a level of uncertainty. Sensor systems may or may not provide a probability of an event interest. The overall performance of a sensor system is dependent on the quality of the sensor(s) in combination with the quality of the algorithm. To build useful algorithms, it is of high importance to clearly define the events to be detected. The event should be defined in such a way that it can be linked to potential interventions. The developers of an algorithm need data to build it. Such data typically consist of the sensor and other data that are going to be used by the algorithm in combination with measurements of the events (the reference standard). The reference standard is usually measured by nonsensor technologies (e.g., by visual observation, clinical examination, and other diagnostic tests such as bacteriological culture).

Evaluating Sensor Systems

Sensors for detection of mastitis or abnormal milk are considered diagnostic tests, which can be characterized using epidemiological parameters. It is very important that the event of interest is clearly defined; especially because the demands for a test might differ for the event of interest (Kamphuis et al., 2013). For instance, the demands for the detection of visually abnormal milk to fulfill milk quality requirements (Rasmussen, 2005) may differ from the demands for the detection of CM. Basic evaluation measures of a sensor system are repeatability and reproducibility. Repeatability can be defined as the closeness of the agreement between the results of successive measurements of a sensor in the same sample. The repeated measurements should be done under circumstances as equal as possible. Reproducibility can be defined as the closeness of the agreement between the results of successive measures of the sensors under varying circumstances, using another sensor and measuring under different circumstances. Evaluating a sensor for its practical use can be done in an experimental setting or based on collection of routine data where the alerts given by the sensor are compared with the occurrence of an event as determined by the reference standard (gold standard; Kamphuis et al., 2016). Because both the start of the gold standard event and the alert are points in time, a time window has to be defined to connect these 2. The length of the time window has an effect on the classification of alerts, as well as on the usefulness of the alerts (Kamphuis et al., 2013). Given a certain predefined time window, the outcomes of an evaluation experiment can be classified as follows (Kamphuis et al., 2013): number of observations in which the event occurs with an alert (TruePosCount); number of observations in which the event occurs without an alert (FalseNegCount); number of observations in which the event does not occur with an alert (FalsePosCount); and number of observations in which the event does not occur without an alert (TrueNegCount). Using the basic classifications, the performance of a sensor for discrete events can be evaluated as follows. Because they are independent of the prevalence of the event (mastitis), the 2 most important parameters to evaluate sensor systems are sensitivity and specificity:

Sensitivity (%) = 100 × TruePosCount/ (TruePosCount + FalseNegCount), and

 Specificity (%) = 100 × TrueNegCount/ (FalsePosCount + TrueNegCount).

For those such as farmers relying on the sensor, however, the sensitivity and specificity are not notable. A farmer sees alerts and wants to know if the alert is a true positive case or a false-positive case. This interpretation is influenced by not only the sensitivity and specificity of the test system, but also the underlying prevalence of the event (mastitis). Hence, the following 2 definitions were proposed for a practical evaluation of sensor systems (Sherlock et al., 2008): Success Rate = TruePosCount/ (TruePosCount + FalsePosCount), and False Alert Rate = 1,000 × FalsePosCount/ Total Cow Milkings. Success rate (or positive predictive value) varies with the prevalence of the condition being monitored and is dependent on specific farm situations. Therefore, when defining demands for sensor systems that can be objectively evaluated, it is best to use sensitivity and specificity, while keeping the expected range of prevalence of the condition to be monitored in mind (Kamphuis et al., 2013).

COWS WITH SEVERE MASTITIS NEEDING IMMEDIATE ATTENTION

The severity of CM can be classified as mild, moderate, or severe (IDF, 2011), also described as a clinical severity score of 1, 2, or 3 (Pinzon-Sanchez and Ruegg, 2011; Ruegg, 2012). Cows with mild CM have abnormal milk as the only clinical sign. Cows with moderate CM have abnormal milk and changes in the affected udder quarter, but the general condition of the cow is not affected. The affected udder quarter(s) typically show signs of inflammation, such as swelling, redness, pain, or warmth. In contrast to cows with mild or moderate CM, cows with severe CM have a compromised general condition with one or more systemic signs of illness. Typical signs are an abnormally increased or decreased rectal temperature, dehydration or marked depression (inappetence or recumbency). In most herds, a minority (5–14%) of the clinical cases are severe (Ruegg, 2012). Even though the incidence of severe CM within a herd is usually low, the severe consequences underline the necessity of accurate and early detection for cow welfare, as well as economic reasons. A cow with severe CM needs immediate attention and treatment. The aim of immediate treatment is to save the cow’s life, reduce clinical signs and thereby improve the cow’s welfare, and increase chances of cure and return to acceptable milk production. The sudden onset of severe clinical signs (within a few hours) and the severely compromised welfare, with risk of the affected cow being lost if not immediately identified and treated properly, poses high demands on an on-farm sensor system. On dairy farms with a conventional milking system, milkers can detect severe CM when fetching the cow or preparing the cow for milking. In such a system, it is relatively easy to detect severe CM, even with minimal procedures for mastitis detection (e.g., no prestripping). On farms with an AMS, detection of severe CM needs to be done using sensor systems. Sensor systems may also be helpful on largescale dairy farms with large capacity rotary milking parlors, because in these systems, there is a very short period of time that milkers interact with the cow.

Demands for Sensor Systems

To prevent further animal suffering and potentially the death of the cow, the purpose of a sensor system to detect cows with severe CM is to detect virtually all severe cases of CM within hours after, and definitely not more than 12 h after, the onset of the severe CM. A very high (close to 100%) sensitivity is needed to ensure the trust of farmers in the system to be able to detect cows that would definitely be detected visually. Because of the clear consequences of severe CM (e.g., toxic mastitis) on SCC, EC, and milk production, it is expected that a very high sensitivity is possible. Regarding the short time interval, there is, to our knowledge, no research available that provides insight into the effects of different times to treatment for severe CM. Demands to identify cows with severe CM are the following: (1) A sensitivity of >95% and preferably as close to 100% as possible to ensure that close to all cows with true cases of severe CM are detected; (2) a high specificity to detect a cow with compromised general health condition of >99%, which minimizes the number of false-positive alerts; and (3) a narrow time window (≤12 h). It should be recognized that with such a high demand in terms of performance (sensitivity >95% and specificity >99%), false-positive alerts will be an issue. Consider a herd of 100 milking cows with an annual incidence of CM of 30%, with 10% of those CM cases being severe CM. Cows are tested (assessed) twice a day for 365 d to identify the 3 severe CM cases. Although the sensitivity is high enough to detect all 3 cases, the success rate would be less than 0.5%. The sensor system would result in, on average, 733 positive alerts/yr, with all but 3 of those being false-positive alerts. This means that, on average 2 false-positive alerts/d are given. It is expected, that a large proportion of these false-positive alerts for severe CM are associated with mild or moderate CM. As a consequence, in practice, these alerts will not be regarded as a false-positive because there is value in examining those cows. When using an AMS, due to the sudden onset of severe CM, sensor information based on changes in milk and measurements collected solely during milking will not be sufficient for all cases. Due to systemic signs, cows with severe CM may not visit the AMS and sensor data from the previous milking may show little if any deviations. A combination of several sensor-based and AMS-based indicators may therefore have to be combined to reach the necessary demands. Sufficient performance of a sensor-based detection system may also be reached by the combination of sensor-based or automatic milking-based monitoring systems in combination with additional monitoring strategies including visual observations linked to certain events (e.g., a cow that does not visit the AMS spontaneously can be checked for disease symptoms when being fetched). To date, sensitivity and specificity to identify severe CM based solely on sensor data are unknown.

Associated Management

Cows identified as potentially having severe CM require immediate attention by the farm personnel to evaluate the presence and severity of clinical signs, discriminate between moderate and severe cases and to decide how the case should be managed. A clinical examination of milk, udder, and general condition of the cow must be carried out including the following aspects: (1) visual inspection of milk, (2) visual inspection and palpation of the udder, and (3) assessment of general condition (e.g., rectal temperature, dehydration, rumen functioning, attitude and diarrhea). A herd-specific protocol, developed together with the herd veterinarian, should be present on the identification and treatment of severe CM.

COWS WITH SCM OR MILD OR MODERATE CM NOT NEEDING IMMEDIATE ATTENTION

In addition to severe CM, 2 other types of mastitis can be distinguished: SCM and mild or moderate CM. SCM is an inflammation of the mammary gland that is not visible and requires a diagnostic test for detection. The most commonly used diagnostic test to identify SCM is milk SCC. Subclinical mastitis is the most prevalent form of mastitis (IDF, 2011). Mild or moderate CM is an inflammation in the udder characterized by observable abnormalities in milk, such as clots of flakes, with no or moderate signs of swelling of the mammary gland or systemic illness (IDF, 2011). Research on treatment systems with delayed treatment of mild and moderate CM to allow further on-farm bacteriological diagnosis has shown that it is possible to be judicious with the use of antimicrobials without any negative effects on mastitis cure rate and production (Lago et al., 2011; Pinzon-Sanchez and Ruegg, 2011; McDougall et al., 2018; Vasquez et al., 2018; Bates et al., 2020). Moreover, from an animal welfare point of view, there is no need to treat cows with mild CM immediately. Therefore, reasons to treat mild CM are to prevent the case from developing into a severe CM, to prevent a reduction of milk quality (bulk milk SCC), and to reduce the transmission of mastitis to other cows (Osteras and Solverod, 2009). Different diagnostic tools and thresholds for diagnosing SCM have been discussed. The basis for such discussions should be repeated sampling and bacteriological culture of quarter milk samples in a herd (Dalen et al., 2019a). Because of the costs, such an approach can be used for research, but it is not rational for routine decision support. Because the definition of SCM is based on detecting signs of inflammation, detection is most commonly through the routine measurement of SCC at the cow level. An elevated SCC is frequently associated with IMI (Reksen et al., 2008). Therefore, traditionally, SCC from DHI systems have been used as a proxy for monitoring IMI status at the cow, herd, and population levels (Schukken et al., 2003). Similar arguments for treatment of mild CM can be used in the discussion about treatment of SCM. Although in general it is not advised to treat SCM, the arguments that are used in studies on the treatment of SCM are to prevent the progression of the SCM case into (severe) CM, to improve milk quality and to prevent transmission of IMI to other cows (e.g., van den Borne et al., 2010; Leitner et al., 2017; Gussmann et al., 2019; van den Borne et al., 2019). Due to the costs involved with routine measurement of the cow’s SCC, testing is stretched to a 3- to 6-week interval and combined with other milk recording measurements. As a result, the cow’s mastitis status is not known at all times and many infections are undetected in between the 2 SCC measurements. Sensor systems have the potential to provide routine monitoring of SCM and mild CM on a daily basis.

Demands for Sensor Systems

The purpose of a sensor system to detect mild or moderate mastitis not needing immediate attention is either to stimulate the application of further management procedures (i.e., diagnostic testing) or to monitor the cow’s status, which gives the cow the opportunity to cure spontaneously, reduces the risk that the mastitis case will become chronic, prevents it from developing into severe CM, and avoids unnecessary production losses. The mild and potentially chronic nature of SCM or mild CM, means that immediate intervention with these cases is not necessary. Sensor systems that identify cows with mild CM or SCM should fulfill the following demands: 1. A sensitivity of >80% is needed to ensure that an acceptable number of cows with mild CM or SCM are detected. Although a higher sensitivity is always better because of the relatively mild consequences of these cases, it is not important to have a very high sensitivity. 2. A very high specificity of >99.5% is needed to prevent false-positive alerts. 3. Because there is no need for a quick intervention, a wide time window of approximately 7 d can be allowed, although a shorter time window may also be used. For this mastitis situation, we consider the same 100 milking cow herd with 30% CM annual incidence and 50% SCM annual incidence. From the CM cases, 90% is mild or moderate. Because interventions do not necessarily have to be carried out immediately, we consider a once-weekly moment to confirm and follow up with cows with mastitis that do not need immediate attention. That means there will be 5,200 tests/yr. A sensitivity of 80% would mean that 62 of the 77 SCM or mild or moderate CM cases would be identified. A specificity of 99.5% would, for this specific herd, lead to a success rate of 70% with 25 false-positive alerts/yr

Limitations of artificial intelligence

Major limitations of application of artificial intelligence in diagnosis and treatment includes; lack of reliable reporting system requiring tens of millions of image/ text samples which are not readily available, samples are structured with scattered an non-uniform information that did not help in facilitation in the learning process of deep learning models, most of the models require labelled data for supervised learning and manual labelling of data is a task of difficulties (Anwar et al., 2018), at individual level, work is not done and coordination with engineers and skilled labour is required (Esteva et al., 2019). For this, major efforts are required from the information technology industry to achieve desired efficacy, accuracy and cost effective. Laboratories and health centres need to work in harmony to progress the implementation and utilisation of electronic health records (Esteva et al., 2019). Image-level diagnostics have been quite successful at employing CNN-based methods. Deep-learning models have achieved physician-level accuracy at a broad variety of diagnostic tasks. AI can assist physicians by providing up-to-date medical information from journals, textbooks and clinical practices to inform proper patient care. Machine learning also plays a role in phenotype prediction from genetic data, disease risk, forecast of epidemic and pandemic (Esteva et al., 2019). A neural network, known as COVNet, examine 4,300 CT scans and accurately distinguish between patients with COVID-19 and other community-acquired pneumonia and lung diseases (Singh et al., 2021). Sofie is the most advanced veterinary medical search tool available. Sofie uses IBM Watson Technology® to search its vast knowledge base of over 40,000 pages of peer-reviewed, evidence-based reference materials from the top veterinary textbooks, journals, and conference proceedings (Sandhya et al., 2020).

Compiled  & Shared by- Team, LITD (Livestock Institute of Training & Development)

Image-Courtesy-Google

Reference-On Request.