Poultry farm vehicle movements as risk factors for the 2014/15 highly pathogenic avian influenza epidemics in the Republic of Korea

Since the first HPAI epidemics in 2003, there has been little epidemiological research on the association between HPAI epidemics and vehicle movements around poultry farms. This study examined the relationship between vehicle movements around poultry farms and the 2014/15 HPAI epidemics in the Republic of Korea using two methods: a boosted regression trees (BRT) model and logistic regression of a generalized linear model (GLM). The BRT model considers the non-linearity association between the frequency of vehicle movements around poultry farms and the HPAI outbreak status per province using a machine learning technique. In contrast, a GLM assesses the relationship based on the traditional frequentist method. Among the three types of vehicle movements (outbound, inbound, and within), only the outbound was found to be a risk factor of the 2014/15 HPAI epidemics according to both the BRT model and multivariate logistic regression of GLM. In the BRT model results, the median relative contribution of the log-transformed outbound variable was 53.68 (range: 39.99 – 67.58) in the 2014 epidemics and 49.79 (range: 33.90 – 56.38) in the 2015 epidemics. In the GLM results, the odds ratio of the log-transformed outbound variable was 1.22 for the 2014 HPAI epidemics (p < 0.001) and 2.48 for the 2015 HPAI epidemics (p < 0.001), respectively. The results indicated that intensive disinfection measures on outbound movement were needed to reduce the risk of HPAI spread. The current BRT models are suitable for risk analysis because the median area under the receiver operating characteristic curve was 0.83 (range: 0.74 – 0.91) and 0.85 (range: 0.73 – 0.87) for the 2014 and 2015 epidemics models, respectively. The Akaike information criterion scores for the multivariate logistic regression of GLM were 150.27 and 78.21 for the 2014 and 2015 epidemics models, respectively. These scores were relatively lower than those from the univariate logistic regression of GLM.