Determining pig holding type from British movement data using analytical and machine learning approaches

Determining the size, location and structure of a livestock population is an essential aspect of surveillance and research as it provides understanding of the representativeness and coverage of any project or scheme. It is an important input for a variety of epidemiological analyses, for example, allowing generation of more accurate sample size calculations for estimating prevalence or freedom from disease, cost-benefit analyses for control measures to reduce or eradicate livestock disease, or development of between-herd network models to estimate the impact of movement of animals between farms on the spread of livestock diseases. The work described here provides information on how British pig movement data was compared against other datasets related to the British pig population to define its appropriateness for defining pig holding demographics. The data were then used to identify the location of pig holdings and the estimated herd size (split into five categories). Two methods are described that were used to classify the holding type of the identified pig holdings. The first method was an epidemiological method that used expert opinion to determine a set of rules based on movement characteristics to classify each holding. The second method was a machine learning approach that used k means cluster analysis to automatically estimate the holding type based on a set of proxy indicators. Each method had a good accuracy rate, when compared to matched holdings present in data provided by the Annual June Agricultural Survey, but all misclassified some holdings. While both of the methods on their own provided a reasonable estimate, it was concluded that a consensus model, considering the results of both models and the Survey, provided the most accurate result. However, the machine learning approach was beneficial, as although some technical expertise was needed to set up the model, it was considerably faster to implement than the other method, as well as being quicker and easier to adapt and re-run with updated information.