Applying Machine Learning to Detect Subclinical Mastitis in Cows Using Readily Available Milk Flow Profile Data

Managing udder health as part of maintaining broader animal health protocols, and to ensure high-quality milk supply to dairies, is a significant challenge facing dairy farms today. Bovine mastitis, an inflammatory response of the mammary gland to infection, is well known to decrease milk yield and quality, with considerable adverse economic effects. Depending on the manifestation of the disease, mastitis in dairy cows can be divided into clinical mastitis and subclinical mastitis. Clinical mastitis is characterised by the presence of visible abnormalities, such as flakes and clots in milk. Subclinical mastitis, on the other hand, is not accompanied by any visible signs of infection. This makes subclinical mastitis more difficult to detect than clinical mastitis. The presence of subclinical mastitis is indicated by an elevated somatic cell count (SCC), which represents the internationally accepted standard for assessing udder health. Although SCC is the common indicator in mastitis detection studies, it is typically measured at most once a month on commercial dairy farms. Modern precision dairy farming technology is capable of measuring high frequency features, enabling the collection of detailed data on a range of variables. The most common features, including total milk yield, peak milk flow-rate, milking time, and electrical conductivity, are routinely recorded by almost all milking systems and provide a basic set of data for monitoring cow productivity and health, which we refer to as readily available features. These readily available large amounts of data make it possible to build mastitis detectors based on machine learning techniques. The research described in this thesis explores the use of milk flow profiles and machine learning for subclinical mastitis detection. The work presented introduces milk flow profiles, a novel and rich data source for cow health monitoring that is readily available on modern dairy farms, and shows that they can be used as a reliable and accurate indicator of cow identity. We also demonstrate that models for subclinical mastitis detection can be built based on readily available data (e.g. milk flow profiles) that are as accurate as models built using expensive, inaccessible data sources (e.g. milk characteristic data). The models built include a set of novel features designed to specifically capture the variation of milking within the herd and across the milking season for a cow that are shown to be accurate predictors of subclinical mastitis and are especially suited to this type of scenario. Finally, however, we illustrate that an explanation deficit arises because of the use of easily accessible data sources (and the features that must be derived from them) instead of richer, more expensive data sources and this is a price paid for using models based on these.