Assessing the potential impact of applying a higher sensitivity test to selected cattle populations for the control of bovine tuberculosis in England.

Bovine tuberculosis (bTB) continues to be the costliest, most complex animal health problem in England. The effectiveness of the test-and-slaughter policy is hampered by the imperfect sensitivity of the surveillance tests. Up to half of recurrent incidents within 24 months of a previous one could have been due to undetected infected cattle not being removed. Improving diagnostic testing with more sensitive tests, like the interferon (IFN)-gamma test, is one of the government's top priorities. However, blanket deployment of such tests could result in more false positive results (due to imperfect specificity), together with logistical and cost-efficiency challenges. A targeted application of such tests in higher prevalence scenarios, such as a subpopulation of high-risk herds, could mitigate against these challenges. We developed classification machine learning algorithms (using 80% of 2012-2019 bTB surveillance data as the training set) to evaluate the deployment of IFN-gamma testing in high-risk herds (i.e. those at risk of an incident in England) in two testing data sets: i) the remaining 20% of 2012-19 data, and ii) 2020 bTB surveillance data. The resulting model, classification tree analysis, with an area under a receiver operating characteristic (ROC) curve (AUC) > 95, showed a 73% sensitivity and a 97% specificity in the 2012-2019 test dataset. Used on 2020 data, it predicted eight percent (3 510 of 41 493) of eligible active herds as at-risk of a bTB incident, the majority of them (66% or 2 328 herds) experiencing at least one. Whilst all predicted at-risk herds could have preventive measures applied, the additional application of IFN-gamma test in parallel interpretation to the statutory skin test, if the risk materialises, would have resulted in 8 585 additional IFN-gamma reactors detected (a 217% increase over the 2 710 IFN-gamma reactors already detected by tests carried out). Only 18% (330 of 1 819) of incidents in predicted high-risk herds had the IFN-gamma test applied in 2020. We therefore conclude that this methodology provides a better way of directing the application of the IFN-gamma test towards the high-risk subgroup of herds. Classification tree analysis ensured the systematic identification of high-risk herds to consistently apply additional measures in a targeted way. This could increase the detection of infected cattle more efficiently, preventing recurrence and accelerating efforts to achieve eradication by 2038. This methodology has wider application, like targeting improved biosecurity measures in avian influenza at-risk farms to limit damage to the industry in future outbreaks.

Machine learning classification methods informing the management of inconclusive reactors at bovine tuberculosis surveillance tests in England.

Bovine tuberculosis (bTB) remains one of the most complex, challenging, and costly animal health problems in England. Identifying and promptly removing all infected cattle from affected herds is key to its eradication strategy; the imperfect sensitivity of the diagnostic testing regime remaining a serious obstacle. The main diagnostic test for bTB in cattle in England, the Single Intradermal Comparative Cervical Tuberculin Test (SICCT: also known as the skin test), can produce inconclusive results below the reactor threshold. The immediate isolation of inconclusive reactor (IR) animals followed by a 60-day retest may not prevent lateral spread within the herd (if it is substandard, allowing transmission) or transmission to wildlife. Over half of IR-only herds that went on to have a positive skin test result (a bTB herd 'incident') in 2020, had it triggered by at least one IR not clearing their 60-day retest, instead of by another test within the previous 15 months. Machine learning classification algorithms (classification tree analysis and random forest), applied to England's 2012-2020 IR-only surveillance herd tests, identified at-risk tests for an incident at the IRs' 60-day retest. In this period, 4 739 out of 22 946 (21 %) IR-only surveillance tests disclosing 6 296 out of 42 685 total IRs, had an incident at retest (2 716 IRs became reactors and 3 580 IRs became two-time IRs). Both models showed an AUC above 80 % in the 2012-2019 dataset. Classification tree analysis was preferred due to its easy-to-interpret outputs, 70 % sensitivity, and 93 % specificity in the 20 % of 2019-2020 testing dataset. The paper aimed to identify IR-only surveillance tests at-risk of an incident at the 60-day retest to target them with appropriate measures to mitigate the IRs' risk. Sixteen percent (341 out of 2 177) of IR-only herd tests were identified as high-risk in the 2020 dataset, with 265 (78 %) of these having at least one reactor or IR at retest. Severe-level reinterpretation of the high-risk IR-only disclosing tests identified in this dataset would turn 68 out of the 590 (12 %) IRs into reactors, generating 23 incidents, the majority (19 or 83 %) part of the 265 incidents that would have been declared at the retest. Classification tree analysis used to identify IR-only high-risk tests in herds eligible for severe interpretation would enhance the sensitivity of the test-and-slaughter regime, cornerstone of the bTB eradication programme in England, further mitigating the risk of disease spread posed by IRs.

Decision tree machine learning applied to bovine tuberculosis risk factors to aid disease control decision making.

Identifying and understanding the risk factors for endemic bovine tuberculosis (TB) in cattle herds is critical for the control of this disease. Exploratory machine learning techniques can uncover complex non-linear relationships and interactions within disease causation webs, and enhance our knowledge of TB risk factors and how they are interrelated. Classification tree analysis was used to reveal associations between predictors of TB in England and each of the three surveillance risk areas (High Risk, Edge, and Low Risk) in 2016, identifying the highest risk herds. The main classifying predictor for farms in England overall related to the TB prevalence in the 100 nearest cattle herds. In the High Risk and Edge areas it was the number of slaughterhouse destinations and in the Low Risk area it was the number of cattle tested in surveillance tests. How long ago the last confirmed incident was resolved was the most frequent classifier in trees; if within two years, leading to the highest risk group of herds in the High Risk and Low Risk areas. At least two different slaughterhouse destinations led to the highest risk group of herds in England, whereas in the Edge area it was a combination of no contiguous low-risk neighbours (i.e. in a 1 km radius) and a minimum proportion of 6-23 month-old cattle in November. A threshold value of prevalence in 100 nearest neighbours increased the risk in all areas, although the value was specific to each area. Having low-risk contiguous neighbours reduced the risk in the Edge and High Risk areas, whereas high-risk ones increased the risk in England overall and in the Edge area specifically. The best classification tree models informed multivariable binomial logistic regression models in each area, adding statistical inference outputs. These two approaches showed similar predictive performance although there were some disparities regarding what constituted high-risk predictors. Decision tree machine learning approaches can identify risk factors from webs of causation: information which may then be used to inform decision making for disease control purposes.

Assessing effects from four years of industry-led badger culling in England on the incidence of bovine tuberculosis in cattle, 2013-2017.

The objective was to measure the association between badger culling and bovine tuberculosis (TB) incidents in cattle herds in three areas of England between 2013-2017 (Gloucestershire and Somerset) and 2015-2017 (Dorset). Farming industry-selected licensed culling areas were matched to comparison areas. A TB incident was detection of new Mycobacterium bovis infection (post-mortem confirmed) in at least one animal in a herd. Intervention and comparison area incidence rates were compared in central zones where culling was conducted and surrounding buffer zones, through multivariable Poisson regression analyses. Central zone incidence rates in Gloucestershire (Incidence rate ratio (IRR) 0.34 (95% CI 0.29 to 0.39, p < 0.001) and Somerset (IRR 0.63 (95% CI 0.58 to 0.69, p < 0.001) were lower and no different in Dorset (IRR 1.10, 95% CI 0.96 to 1.27, p = 0.168) than comparison central zone rates. The buffer zone incidence rate was lower for Gloucestershire (IRR 0.64, 95% CI 0.58 to 0.70, p < 0.001), no different for Somerset (IRR 0.97, 95% CI 0.80 to 1.16, p = 0.767) and lower for Dorset (IRR 0.45, 95% CI 0.37 to 0.54, p < 0.001) than comparison buffer zone rates. Industry-led culling was associated with reductions in cattle TB incidence rates after four years but there were variations in effects between areas.

Methodology and preliminary results of a systematic literature review of ante-mortem and post-mortem diagnostic tests for bovine tuberculosis.

A systematic review was conducted to identify studies with data for statistical meta-analyses of sensitivity (Se) and specificity (Sp) of ante-mortem and post-mortem diagnostic tests for bovine tuberculosis (bTB) in cattle. Members of a working group (WG) developed and tested search criteria and developed a standardised two-stage review process, to identify primary studies with numerator and denominator data for test performance and an agreed range of covariate data. No limits were applied to year, language, region or type of test in initial searches of electronic databases. In stage 1, titles and available abstracts were reviewed. References that complied with stage 1 selection criteria were reviewed in entirety and agreed data were extracted from references that complied with stage 2 selection criteria. At stage 1, 9782 references were reviewed and 261 (2.6%) passed through to stage 2 where 215 English language references were each randomly allocated to two of 18 WG reviewers and 46 references in other languages were allocated to native speakers. Agreement regarding eligibility between reviewers of the same reference at stage 2 was moderate (Kappa statistic = 0.51) and a resolution procedure was conducted. Only 119 references (published 1934-2009) were identified with eligible performance estimates for one or more of 14 different diagnostic test types; despite a comprehensive search strategy and the global impact of bTB. Searches of electronic databases for diagnostic test performance data were found to be nonspecific with regard to identifying references with diagnostic test Se or Sp data. Guidelines for the content of abstracts to research papers reporting diagnostic test performance are presented. The results of meta-analyses of the sensitivity and specificity of the tests, and of an evaluation of the methodological quality of the source references, are presented in accompanying papers (Nuñez-Garcia et al., 2017; Downs et al., 2017).

Meta-analyses of the sensitivity and specificity of ante-mortem and post-mortem diagnostic tests for bovine tuberculosis in the UK and Ireland.

Bovine Tuberculosis (bTB) in cattle is a global health problem and eradication of the disease requires accurate estimates of diagnostic test performance to optimize their efficiency. The objective of this study was, through statistical meta-analyses, to obtain estimates of sensitivity (Se) and specificity (Sp), for 14 different ante-mortem and post-mortem diagnostic tests for bTB in cattle. Using data from a systematic review of the scientific literature (published 1934-2009) diagnostic Se and Sp were estimated using Bayesian logistic regression models adjusting for confounding factors. Random effect terms were used to account for unexplained heterogeneity. Parameters in the models were implemented using Markov Chain Monte Carlo (MCMC), and posterior distributions for the diagnostic parameters with adjustment for covariates (confounding factors) were obtained using the inverse logit function. Estimates for Se and/or Sp of the tuberculin skin tests and the IFN-γ blood test were compared with estimates published 2010-2015. Median Se for the single intradermal comparative cervical tuberculin skin (SICCT) test (standard interpretation) was 0.50 and Bayesian credible intervals (CrI) were wide (95% CrI 0.26, 0.78). Median Sp for the SICCT test was 1.00 (95% CrI 0.99, 1.00). Estimates for the IFN-γ blood test Bovine Purified Protein Derivative (PPD)-Avian PPD and Early Secreted Antigen target 6 and Culture Filtrate Protein 10 (ESAT-6/CFP10) ESAT6/CFP10 were 0.67 (95% CrI 0.49, 0.82) and 0.78 (95% CrI 0.60, 0.90) respectively for Se, and 0.98 (95% CrI 0.96, 0.99) and 0.99 (95% CrI 0.99, 1.00) for Sp. The study provides an overview of the accuracy of a range of contemporary diagnostic tests for bTB in cattle. Better understanding of diagnostic test performance is essential for the design of effective control strategies and their evaluation.

Evaluation of the methodological quality of studies of the performance of diagnostic tests for bovine tuberculosis using QUADAS.

There has been little assessment of the methodological quality of studies measuring the performance (sensitivity and/or specificity) of diagnostic tests for animal diseases. In a systematic review, 190 studies of tests for bovine tuberculosis (bTB) in cattle (published 1934-2009) were assessed by at least one of 18 reviewers using the QUADAS (Quality Assessment of Diagnostic Accuracy Studies) checklist adapted for animal disease tests. VETQUADAS (VQ) included items measuring clarity in reporting (n = 3), internal validity (n = 9) and external validity (n = 2). A similar pattern for compliance was observed in studies of different diagnostic test types. Compliance significantly improved with year of publication for all items measuring clarity in reporting and external validity but only improved in four of the nine items measuring internal validity (p < 0.05). 107 references, of which 83 had performance data eligible for inclusion in a meta-analysis were reviewed by two reviewers. In these references, agreement between reviewers' responses was 71% for compliance, 32% for unsure and 29% for non-compliance. Mean compliance with reporting items was 2, 5.2 for internal validity and 1.5 for external validity. The index test result was described in sufficient detail in 80.1% of studies and was interpreted without knowledge of the reference standard test result in only 33.1%. Loss to follow-up was adequately explained in only 31.1% of studies. The prevalence of deficiencies observed may be due to inadequate reporting but may also reflect lack of attention to methodological issues that could bias the results of diagnostic test performance estimates. QUADAS was a useful tool for assessing and comparing the quality of studies measuring the performance of diagnostic tests but might be improved further by including explicit assessment of population sampling strategy.

Assessing the effects of the first 2 years of industry-led badger culling in England on the incidence of bovine tuberculosis in cattle in 2013-2015.

Culling badgers to control the transmission of bovine tuberculosis (TB) between this wildlife reservoir and cattle has been widely debated. Industry-led culling began in Somerset and Gloucestershire between August and November 2013 to reduce local badger populations. Industry-led culling is not designed to be a randomized and controlled trial of the impact of culling on cattle incidence. Nevertheless, it is important to monitor the effects of the culling and, taking the study limitations into account, perform a cautious evaluation of the impacts. A standardized method for selecting areas matched to culling areas in factors found to affect cattle TB risk has been developed to evaluate the impact of badger culling on cattle TB incidence. The association between cattle TB incidence and badger culling in the first 2 years has been assessed. Descriptive analyses without controlling for confounding showed no association between culling and TB incidence for Somerset, or for either of the buffer areas for the first 2 years since culling began. A weak association was observed in Gloucestershire for Year 1 only. Multivariable analysis adjusting for confounding factors showed that reductions in TB incidence were associated with culling in the first 2 years in both the Somerset and Gloucestershire intervention areas when compared to areas with no culling (incidence rate ratio (IRR): 0.79, 95% CI: 0.72-0.87, p < .001 and IRR: 0.42, 95% CI: 0.34-0.51, p < .001, respectively). An increase in incidence was associated with culling in the 2-km buffer surrounding the Somerset intervention area (IRR: 1.38, 95% CI: 1.09-1.75, p = .008), but not in Gloucestershire (IRR: 0.91, 95% CI: 0.77-1.07, p = .243). As only 2 intervention areas with 2 years of data are available for analysis, and the biological cause-effect relationship behind the statistical associations is difficult to determine, it would be unwise to use these findings to develop generalizable inferences about the effectiveness of the policy at present.