Skip to main content

Research Repository

Advanced Search

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

Romero, M. Pilar; Chang, Yu-Mei; Brunton, Lucy A.; Parry, Jessica; Prosser, Alison; Upton, Paul; Drewe, Julian A.


M. Pilar Romero

Yu-Mei Chang

Lucy A. Brunton

Jessica Parry

Alison Prosser

Paul Upton

Julian A. Drewe


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.


Romero, M. P., Chang, Y., Brunton, L. A., Parry, J., Prosser, A., Upton, P., & Drewe, J. A. (2021). Machine learning classification methods informing the management of inconclusive reactors at bovine tuberculosis surveillance tests in England. Preventive Veterinary Medicine,

Journal Article Type Article
Acceptance Date Dec 20, 2021
Publication Date Dec 20, 2021
Deposit Date Jul 29, 2021
Publicly Available Date Dec 20, 2022
Print ISSN 0167-5877
Publisher Elsevier
Peer Reviewed Peer Reviewed
Public URL