Using machine learning improves predictions of herd-level bovine tuberculosis breakdowns in Great Britain

Abstract

In the United Kingdom, despite decades of control efforts, bovine tuberculosis (bTB) has not been controlled and currently costs ~ £100 m annually. Critical in the failure of control efforts has been the lack of a sufficiently sensitive diagnostic test. Here we use machine learning (ML) to predict herd-level bTB breakdowns in Great Britain (GB) with the aim of improving herd-level diagnostic sensitivity. The results of routinely-collected herd-level tests were correlated with risk factor data. Four ML methods were independently trained with data from 2012-2014 including ~ 4700 positive herd-level test results annually. The best model's performance was compared to the observed sensitivity and specificity of the herd-level test calculated on the 2015 data resulting in an increased herd-level sensitivity from 61.3 to 67.6% (95% confidence interval (CI): 66.4-68.8%) and herd-level specificity from 90.5 to 92.3% (95% CI: 91.6-93.1%). This approach can improve predictive capability for herd-level bTB and support disease control.

Figure 1

Receiver operating characteristic curves of Gradient Boosted Trees performing on the development and testing datasets. Dots represent the HSe and HSp of the herd-level SICCT test observed in the development and testing sets. Crosses show the HSe and HSp of GBT predictions dichotomized with a threshold of 0.695.

Figure 2

(a) Map of all 3171 bTB infected farms identified by the GBT model in the testing set of 2015 data. (b) Map of 297 infected farms which were undetected by the herd-level SICCT test, but identified by the GBT model in the testing set of 2015 data. The maps were drawn using the Matplotlib v2.1.0 library of Python.

Figure 3

Comparison between iterative variable elimination performed with GBT and RF evaluated with the development set of 2014 data. Plot (a) shows AUC measured over varying number of input variables, where for every number of variables the model was retrained and evaluated separately. Plot (b) shows correlation between rankings obtained by GBT and RF.