Public Health Dashboard powered by Digital Disease Detection for Local Epidemiologists: A Gastrointestinal illness case study

A toolkit harnessing high-throughput machine learning classification on publicly available data has been developed. A dashboard powered by this toolkit has been designed to provide digital disease detection for public health action. The disease detection algorithms rely on guided machine learning to filter for statements of interest via a human-trained Natural Language ToolKit (NLTK) classifier. The initial case study uses data from both Twitter and online media reports to identify outbreaks of gastro-intestinal illness. Statements are assigned a level of relevance and then a clustering algorithm detects elevations in disease activity across geographic regions. Relevant statements are then visualized on the dashboard as timeseries, annotated maps, and word flow series. Contextual meta-data for each statement is also presented along with more traditional sources like health care surveillance data, CDC reports, and school absenteeism. Importantly, the dashboard is designed to crowdsource relevance of the statements and incorporate additional data that isn’t present in the original statements.  This allows humans to augment the automated methods and share information through the dashboard itself. For example, an Emergency Medical Technician may know that a mentioned restaurant is operated by the same owner as restaurants previously cited for health violations, with this knowledge the local epidemiologist decides to take action thus quickly halting the outbreak. To ensure usability feedback from an expert panel of epidemiologists, pharmacists, culture and linguistic specialists, and infectious disease physicians was integrated in the dashboard’s design reflecting geographical norms regarding illness and disease. Additionally an evaluation component of the project will capture the utility of identified events and will be used to iteratively refine the dashboard.  The toolkit was designed to be readily adaptable for detection of diverse disease risk factors, languages, and public data sources. This modularity ensures an effective and efficient social media surveillance dashboard tailored for local use.