Analysis of Heterogeneous Data Sources for Veterinary Syndromic Surveillance to Improve Public Health Response and Aid Decision Making

The standard technique of implementing veterinary syndromic surveillance (VSyS) is the detection of temporal or spatial anomalies in the occurrence of health incidents above a set threshold in an observed population using the Frequentist modelling approach. Most implementation of this technique also requires the removal of historical outbreaks from the datasets to construct baselines. Unfortunately, some challenges exist, such as data scarcity, delayed reporting of health incidents, and variable data availability from sources, which make the VSyS implementation and alarm interpretation difficult, particularly when quantifying surveillance risk with associated uncertainties. This problem indicates that alternate or improved techniques are required to interpret alarms when incorporating uncertainties and previous knowledge of health incidents into the model to inform decision-making. Such methods must be capable of retaining historical outbreaks to assess surveillance risk.
In this research work, the Stochastic Quantitative Risk Assessment (SQRA) model was proposed and developed for detecting and quantifying the risk of disease outbreaks with associated uncertainties using the Bayesian probabilistic approach in PyMC3. A systematic and comparative evaluation of the available techniques was used to select the most appropriate method and software packages based on flexibility, efficiency, usability, ability to retain historical outbreaks, and the ease of developing a model in Python. The social media datasets (Twitter) were first applied to infer a possible disease outbreak incident with associated uncertainties. Then, the inferences were subsequently updated using datasets from the clinical and other healthcare sources to reduce uncertainties in the model and validate the outbreak. Therefore, the proposed SQRA model demonstrates an approach that uses the successive refinement of analysis of different data streams to define a changepoint signalling a disease outbreak.
The SQRA model was tested and validated to show the method's effectiveness and reliability for differentiating and identifying risk regions with corresponding changepoints to interpret an ongoing disease outbreak incident. This demonstrates that a technique such as the SQRA method obtained through this research may aid in overcoming some of the difficulties identified in VSyS, such as data scarcity, delayed reporting, and variable availability of data from sources, ultimately contributing to science and practice.