EPI-AI: Automated Understanding and Alerting of Disease Outbreaks from Global News Media

Disease outbreaks, such as Zika, Ebola and SARS epidemics, are of the greatest importance to the international community and the UK/Canadian governments. Public health organisations need data as early as possible in an outbreak to respond rapidly and prevent human suffering. Traditional bio-surveillance relies on human laboratory networks, but these data are often unavailable in real-time, patchy in geographic coverage, and tuned to specific diseases. Digital disease surveillance (DDS) using Web-based news data overcomes some of these limitations, providing a critical supplement to traditional networks. However, current DDS systems rely to a large extent on manual screening of Web data for events of interest: a skilled and labour-intensive process given the volume, multilingualism, velocity and potential bias of news sources.
Research has shown that there is significant potential to automate DDS. Natural Language Processing (NLP) has been in use since the early 2000s to efficiently detect and track health threats from outbreak news reports. For example, the Canadian GPHIN system, which detected the first evidence of SARS, uses a combination of NLP and human experts to sift through over 20K online news reports each day in nine languages. However, traditional automated approaches are insensitive to context that can help experts to interpret risk factors and fail to take account of possible data biases.
Our goal in the EPI-AI project is to achieve a step-change in real-time automated DDS. Previous work has tended to take a siloed approach, focusing on Natural Language Processing methods or spatial analysis with little consideration of equality considerations that arise from biases in the data. We will use an interdisciplinary approach, combining expertise from three disciplines - computer science, epidemiology, and bioethics - to develop novel machine learning and statistical models adapted to the complex data and objectives of global epidemic surveillance.
Benefits that we see include: (i) improved geographic precision and coverage; (ii) improved ability to understand the topical focus of a report; (iii) automated normalisation of risk factors to a standard terminology for integration of evidence across systems; (iv) automated spatio-temporal analysis of reports to update global risk maps and trigger alerts; and (v) provision of contextual information on potential media bias to support interpretation of alerts.
This fundamentally interdisciplinary research will be closely aligned with key Canadian, UK and global public health stakeholders.

The development of novel neural and statistical machine learning (ML) methods for the extraction of structured event data from news media, the understanding of bias in news data, and the integration of event data with baseline data for the purpose of timely risk assessment will have an important scientific and technological impact. This investigation is highly relevant to a several domains such as public health, computer science, life science and medicine.
1. Public health experts performing infectious disease alerting, situation awareness and risk assessment will benefit from being more efficient and access to earlier warnings and greater coverage about health threats, e.g. pandemic influenza, Zika, Ebola and Marburg. In addition to using modern ML techniques to dramatically improve Natural Language Processing (NLP), the proposed technology will also integrate these methods into an advanced spatial analysis framework to support public health analysts in early alerting, tracking and risk assessment. The techniques supplement scarce human expertise (by replacing manual search and de-duplication), bring in evidence beyond national boundaries, cover segments of the population who may not interact with traditional clinical surveillance networks (e.g. patients who may not visit a GP), and incorporate media bias. With respect to risk assessment, the project will provide valuable data to calibrate the parameters of disease transmission models which are often hampered by insufficient data.
2. Life scientists and clinicians involved in translational studies will benefit from having a novel database of epidemiological evidence about infectious diseases and their associations to risk factors such as symptoms, locations and population descriptors with links to existing scientific data infrastructure through standard ontological codes. The database will include region-specific information about potential sources of media bias that influence reporting.
3. Industry involved in AI technologies and e-science will benefit from open source software tools and publications explaining state-of-the-art ML techniques for text mining and risk alerting that takes account of bias reduction. AI has huge potential in healthcare as a means to support patients and clinicians in making decisions and to reduce administrative costs. The techniques pioneered in this proposal for bias reduction, NLP/Machine Translation and risk analysis are highly relevant to a wide community involved in the development and use of AI in clinical practice, e.g. technology companies involved in R&D on electronic patient records, the pharmaceutical industry looking for online evidence to repurpose drugs, and online patient support networks.
4. Decision-makers and the public will benefit from having improved AI technologies for early detection of health threats and improved understanding of their benefits and limitations. The main benefit is the potential for the research results to make the world safer from the threat of emerging and re-emerging epidemics by strengthening our global capacity to detect and control such threats rapidly, before they cause extensive human suffering. Public understanding of the project will be aided through the openly available EPI-AI portal, conference publications and demonstrations.