Bayesian Tools for Early Disease Detection

Abstract

LPAI is a (poultry) disease which comes with mild and non-specific clinical symptoms. It has the potential to mutate to HPAI, which is highly contagious and comes with serious clinical symptoms such as high mortality. In 2003, an HPAI outbreak in the Netherlands cost more than a billion Euro, and approximately 30 million animals had to be culled in order to control the outbreak. Such consequences make it important to detect LPAI as early as possible, before it can mutate to HPAI. We have developed and studied multiple Bayesian tools to assist the veterinarian in this task. The early detection of LPAI should start at the poultry farm, where daily measurements of the production parameters would ideally be registered. Important production parameters of a flock of laying hens are the feed intake, water intake, mortality rate, egg production and egg weight. Comparing the measured values against the expected ones is the basis for detecting anomalies in the production parameters. With living animals, such as laying hens, the expected production values differ among flocks due to natural variability. We have developed the pseudo point method to remove this natural variability from the measurements for a production parameter. Based on an expected trend, expressed by a mathematical function, a prediction function is derived. Each measurement is used to update this prediction function, causing it to slowly adapt to the measured data. For a veterinarian visiting a poultry farm, it is very hard to distinguish LPAI from other poultry diseases. We have developed a Bayesian network to assist the veterinarian in determining whether there is an increased probability of the flock being infected with an LPAI virus. This probability is mainly based on farm-specific details, clinical signs on flock level, clinical signs on individual bird level and post-mortem findings. Upon constructing Bayesian networks, the noisy-OR model and its generalisations are frequently used for reducing the number of parameter probabilities to be assessed. Empirical research has shown that using the noisy-OR model for nearly every causal mechanism in a network, does not seriously hamper the quality of its output. We have provided a formal underpinning of this finding and have identified under which conditions applying the noisy-OR model can potentially harm the network's output. Additionally, we have developed the intercausal cancellation model, of which the noisy-OR model is a special case, to provide network engineers with a tool to model cancellation effects between cause variables.