Data source

Abattoir data were accessed through the databases of the two pig abattoir lesion scoring health schemes which exist in Great Britain: Wholesome Pigs Scotland (WPS) (covering Scotland) and British Pig Health Scheme (BPHS) (covering England and Wales) [6]. The health schemes provide services in 17 pig abattoirs. Both schemes obtain a sample from each batch of pigs by assessing every second pig on the slaughter line. The scoring was carried out by swine veterinarians trained in this method of testing on the abattoir inspection line. The data were from a three year period (July 2005 to June 2008).

Dataset

For the purpose of this investigation, a batch is defined as a group of pigs from a single farm submitted to the abattoir on a particular date. A total of 6485 batches were included, submitted from 1138 farms, with a median of 4 batches assessed per farm (first quartile 2, third quartile 8). All the batches consisted of exactly 50 pigs assessed.

Scoring for the different pathologies

Consistency in the scoring of the pathologies

Both health schemes carried out exercises to standardise the definition of each lesion across the inspectors. One WPS assessor was involved in the training of all the other inspectors that carried out WPS and BPHS assessments during the three year period included in this study. Once a year, all the inspectors underwent a refresher/training day where the same pigs and pathologies were assessed by all the assessors and feed-back was provided by the trainer. These assessment exercises aimed to maintain the consistency in the scoring criteria across assessors by identifying and correcting potential misclassifications. Furthermore, the schemes aimed to include at least two assessors per abattoir and to place each assessor in at least two different abattoirs, thereby minimising the potential of operator bias.

Definition of pathology batch-status variables

Multi-dimensional machine learning methodology

The process explained below aims to identify an optimal multi-dimensional model, i.e. a graphical model displayed as a network of connections, where each connection (arc) describes a statistically significant association between the different lesions in the data. Figure 1 schematically represents the machine leaning structure discovery process utilised, which is initiated with numerous series of searches followed by steps to summarise the results of each search. This methodology consists of fitting models which are network structures technically referred to as directed acyclic graphs (a graph with no loops), in which nodes correspond to the pathology batch-status variables and arcs between nodes (represented by arrows) indicate that a direct probabilistic dependency (e.g. an association) exist between nodes.

Direction of the arrows

The direction of the arcs connecting nodes is informed by the data, reflecting the dependency structure which generated the data [26]. The direction only implies association and says nothing of causality. Arc direction is as a result of the underlying mathematics used to construct the models (technically the graph denotes a factorisation of the joint probability distribution of the data). Models with particular arc directions may be better fit to the data than with the reversed directions, and therefore preferred, however, it would be incorrect with this information alone to infer that the biological dependence between two nodes is supported more in one direction over another, all that can be inferred is that association exists between nodes.

Searching for locally optimal structures

The machine learning structure discovery process was performed through series of local heuristic searches using a standard approach proposed by Heckerman et al. [24]. Locally optimal models are identified by random-restart local hill climbing searches, also known as a “greedy search” [26], which seek to maximise the goodness of fit metric (network score) for each model. This network score is given by the (log) marginal likelihood of the data given the model; equivalent to the Bayes factor when using equal prior on each model structure. This search process can be thought of as roughly analogous to stepwise regression in linear modelling but conducted in multiple dimensions where the initial model from which the search commences is randomly chosen. The interrelationships within different batch-status categories for the same pathology are inversely related – i.e. when one batch-status is present the others are not. Therefore arcs connecting the different batch-status categories for the same pathology (e.g. EP zero with EP high, or EP high with EP mild/moderate, and so on) were banned from the search.

Summarising the results from the local searches

Identifying the final network

Out of the three structures described above, the pruned network is the model that provides the most robust and conservative approach and is therefore considered in this paper as the principal result. The strength of the association between two nodes (pathology batch-status variables) present in the pruned network was estimated by calculating the relative risk (RR) (also known as risk ratio) [32]. RR is calculated as the proportion of batches affected with condition A given condition B is present in the batch, divided by proportion of batches with the condition A but with condition B not present. The 95% confidence intervals (CI) for the RR were estimated using Monte-Carlo simulation.

Parameters in the search algorithm

The optimal number of local searches required to identify a robust machine learning structure is problem specific. In this study, the number was determined empirically by running two parallel sets of searches, differing in the set.seed value. The number of local searches was increased until both sets reached the same majority consensus network, thereby suggesting that a sufficient numbers of searches had been run to provide robust outputs. The results from both sets of searches were pooled to identify the best overall single network which, combined with the majority consensus network, led to the pruned network.

The analyses were performed in R [33] using a library written by FIL (freely available upon request) to perform the structure search. Other broadly similar libraries are available for use within R from CRAN (Comprehensive R Archive Network) website, and similar toolboxes are available for use with MATLAB.

Graphical outputs

The “majority consensus network” is presented in Figure 2 and provides complementary information to the main output from this investigation, the “pruned network”, which is presented in Figure 3 completed with estimated RRs. The arcs presented in the “pruned network” could be identified in the “majority consensus network” to determine the percentage of local searches in which the particular arc appears, informing about the robustness. Thus, the “majority consensus network” (Figure 2) shows that the connections leading to milk spots from hepatic scarring and papular dermatitis, are the most robust – present in more than 90% of searches. In the pruned network (Figure 3) these arcs are retained, and it is observed that those batches with hepatic scarring had a moderate risk of milk spots compare to those batches not presenting hepatic scaring; likewise batches with papular dermatitis had a milder risk of milk spots compare to those with papular dermatitis absent. Figure 3 also shows that batches with mild or moderate levels of pleurisy were more likely to be enzootic pneumonia-like free than those with other category levels of pleurisy. The pneumonia free batches were more likely to be also free of papular dermatitis than those with pneumonia. Batches with a high level of enzootic pneumonia-like lesions had a moderate risk (i.e. RRs between 1.5 and 2.5) of having a high level of pleurisy compare to those batches with pneumonia absent or with moderate/low level. Having abscesses is associated with batches with a higher level of pleurisy compare to those with no abscesses. Batches with a moderate/low level of pleurisy had a negative risk of abscess and pericarditis compare to other batches with other levels of pleurisy. There is stronger risk (i.e. RRs over 2.5) of having peritonitis if pericarditis is present in the batch than if it is absent; conversely those batches with zero level of pleurisy are more likely to be peritonitis free than those with pleurisy present. Batches with peritonitis also had a milder risk (i.e. RRs between 1 and 1.5) of hepatic scarring compare to batches without peritonitis. Batches with pyaemia had a mild risk of hepatic scarring and a strong risk of having tail damage compare to those batches with pyaemia absent.

The interrelationship between the pathologies

This study has provided information on the nature of hepatic scarring which is thought to be a post healing stage of milk spots; but for which other aetiologies can not be discarded. The results suggest that both Ascaris suum and systemic bacterial infections are independently interrelated with the presence of the liver capsule scarring. The former is reflected in the moderate association with milk spots, which seems to be a highly robust interrelationship as it was recruited in 92% of the searches. Different stages of A. suum parasitism within the same batch may take place; leading to coexistence of active milk spot lesions with those already healed, i.e. hepatic scarring. The potential bacterial aetiology of hepatic scarring is suggested as its risk increases with the presence of pyaemia – suggesting that both pathologies may be associated with Arcanobacterium pyogenes. In addition, hepatic scarring appears interrelated with peritonitis, which is typically present in systemic infections by Haemophilus parasuis (responsible for Glasser’s disease) or Streptococcus suis[34, 35]. These latter infectious agents would also explain the positive association between peritonitis and pericarditis. Likewise it was observed that when pleurisy was absent the chance of being peritonitis free increased.

Severe pneumonic pasteurellosis is typically manifested by abscessation and thoracic wall adherences [36] which explains the associations detected between pleurisy and abscess. Conversely, absence of pneumonia is associated with lower levels of abscesses. High batch prevalence of enzootic pneumonia-like lesions is interrelated to high levels of pleurisy, which is an expected finding as both respiratory conditions share common husbandry risk factors [6] and Mycoplasma hyopneumoniae (main pathogen for enzootic pneumonia) contributes to the occurrence of pleurisy [4]. This latter association may reflect the presence of poor health levels, particularly in the control of respiratory diseases. Alternatively, batches with mild or moderate levels of pleurisy appear more likely to be free of enzootic pneumonia, pericarditis and abscess, reflecting perhaps high health batches. Papular dermatitis is associated with the presence of milk spots, both being parasitic conditions. This association could reflect poor parasitic control strategies for some producers and highlights the fact that, even with current systems of production, parasitism is still neglected by some sectors of the industry. These results could be used to optimise abattoir inspection strategies. For example, when papular dermatitis is detected in the pigs (e.g. during the ante mortem inspections) the meat inspectors should place more emphasis in the liver inspections of those batches. This would be a proxy for the implementation of risk based surveillance abattoir polices that could optimise the use of industry and government resources [37].

Presence of pyaemia in the batch is associated with presence of tail damage. This latter pathology is known to be involved in early stages of the pathogenesis of pyaemia, by facilitating an entry access for bacteria [10]. At pig level these two lesions might not coexist simultaneously due to the time gap between the tail damage and the development of the pyaemia [10], but a batch level investigation may have assisted to find such association. The approach used in this study, investigating batch level prevalence, not only maintains coherence with the nature of pig production, but would have also assisted in the identification of any association when two pathologies may be part of the same causal pathway (e.g. milk spots and hepatic scarring). In this scenario it is likely that the pathologies do not coexist in the same pig, therefore pig level investigation would be an inefficient way of exploring their association. Furthermore, pathologies presented in a mild form or during the healing process can be missed in the abattoir inspections; whereas if they are present in more than one pig, the chance of being detected by the abattoir assessors increases leading to a more adequate batch level classification.

Clustering in the structure of data

In this study the impact on the analyses of the potential clustering structure in the data has been mitigated by modelling the data at batch level. Batch is typically the lowest and likely the strongest level of clustering present in abattoir data [38], particularly in health scheme pig abattoir data [6, 8]. It is also arguable that for the type of analyses presented – particularly the multi-dimensional aspect– clustering is of far less concern than in other types of traditional statistical analyses. For other potential levels of clustering to be an issue, e.g. on-farm (or abattoir or season), this would require that on different farms the proportion of batches which have, for example, {lesion A present given that lesion B is present and lesion C is not present and lesion D is present and so on…} are substantially different, and similarly for all the other conditional probabilities in the model. This form of "group-effect" is unlikely to be sizable after having already jointly adjusted for all the other conditions present in a batch. Hence, intuitively it could be argued that the machine learning methodology is robust to clustering, whether this is at farm/abattoir/season level. In practical terms, this assertion cannot be rigorously tested with this methodology and it should therefore be acknowledged as a potential limitation in this study.

Constraints of abattoir gross pathology data

The different pathologies were presented in this paper with their most typical cause (Table 1) and although some of them, i.e. EP-like lesion, milk spots and papular dermatitis, can be considered good proxies for specific pathogens [9, 15, 39, 40], none of them are strictly pathognomonic. The data obtained from abattoir monitoring carried out by the health schemes offered here a unique opportunity to explore the associations between these relevant pig pathologies. The presence of operator bias across the assessors, affecting the gross pathology classification cannot be absolutely ruled out, but the definition of the lesions did not change during the period included in our study. Additionally, the health schemes organise training and refresher days for the veterinarians and conduct internal comparisons on the same pigs assessed by different veterinarians, aiming to maintain assessor consistency over time

The results from this study are applicable to the whole study population, i.e. those farms participating in the pig health schemes, and particularly to those units that submitted several batches of pigs over the time period included in the study. Additionally, the results could be extrapolated to the population of British pig commercial units, as the assessments carried by the health schemes are considered representative of the British commercial sector [41].

Further discussion on the structure discovery approach

The multi-dimensional machine learning methodology presented is well suited for investigating multiple associations between pathologies, generating hypotheses about potential interrelationships. Linear models and their generalizations, for example, would have required designating one variable as a response and modelling the rest as a set of independent predictors. Multivariate techniques like principal component and factor analyses utilise dimension reduction to facilitate the identification of uncorrelated subgroups of variables (i.e. principal components and factors). In contrast, machine learning structure discovery does not reduce the dimensions of the data and its graphical nature allows for ready interpretation of all associations present. In this study, a small variable domain – ten pathologies studied in 14 variables – is modelled with a substantial amount of data, providing the ideal scenario for structure discovery multivariate analyses [25].