Today’s AI provides a flexible toolkit that can be deployed in response to pressing policy issues. ELISE researchers have driven effective deployment of these tools by combining technical expertise with collaborations that engage policymakers and citizens, delivering trustworthy AI decision-making tools.
Access to up-to-date data and insights about the dynamics of an infectious disease is critical in preventing its spread. By 2019, research using machine learning and data science to analyse outbreaks of the H1N1 virus in Mexico and Ebola in Central Africa had already demonstrated how new modelling tools could help analyse the trajectory of an outbreak and identify potential public health responses. In 2020, as COVID-19 rapidly spread across the world, the challenge for data science and machine learning researchers was to develop analytical tools that could be reliably deployed to help make sense of the pandemic, and to bridge the gap between research and policymaking.
Responding to this challenge, Nuria Oliver, co-Director of the ELLIS/ELISE Human-centric Machine Learning Programme, and her team at the ELLIS Unit Alicante Foundation created and led a coalition of scientists from different research institutions in the Valencian region of Spain with the goal of providing data-driven insights to support decision-making in the regional government. This interdisciplinary team – the Data Science against COVID-19 Taskforce – combined expertise in machine learning, data science, engineering, statistics, epidemiology, and policy, delivering new research in four key areas:
Modelling human mobility:
Understanding how people are travelling and interacting is vital in predicting the spread of disease. To help extract insights from large datasets describing people’s movements across the Valencian Region of Spain, the team developed analytical tools and accessible interfaces for data visualisation. These tools supported policy discussions about the impact of lockdowns or social containment measures on citizens’ mobility, helping identify where such measures were more or less successful, and modelling the spread of COVID-19.
Building computational epidemiological models:
Three different types of model – an SEIR model; an agent-based model; and a deep learning model – helped make predictions about how the pandemic would evolve, according to different scenarios for the use of non-pharmaceutical interventions, vaccination, or other policy measures and at different levels of spatial granularity. The resulting simulations allowed policymakers to explore the implications and trade-offs associated with different policy interventions.
Developing predictive models of healthcare systems:
In addition to a suite of epidemiological models, the team also created predictive modelling tools that could translate insights about the pandemic to predictions about the number of cases, hospitalisations, and intensive care patients at a local or regional level. These insights allowed healthcare authorities to plan for waves of new COVID-19 cases, for example as the seasons changed or social confinement periods ended.
Creating a new citizen science initiative:
Many aspects of people’s behavioural or emotional responses to the pandemic were not well-represented in existing datasets. To help plug these gaps, the team launched a large-scale online survey that captured people’s experiences. The survey quickly received widespread attention, with over 140,000 responses in its first 40 hours after launch and over 700,000 answers today. This became an important tool to understand how patterns of social contact changed, the economic effects of policy changes, the prevalence of different symptoms, and more. Rapid analysis and visualisation of results has allowed researchers and policymakers to better understand the practical impact of policy measures.
Strong connections with regional and national governments helped translate these data-driven insights to policy development and implementation. An important focus of the team’s work was interpreting, prioritising, and translating the insights from machine learning systems to actionable advice, aided by collaborations with policymakers in the design and use of these AI tools.
The resulting programme demonstrates the value of collaboration across scientific communities, publics, and public administration, leveraging data science and AI to deliver better-informed public policy responses. This work has delivered cutting-edge science, demonstrated through a range of publications. It has also directly influenced policy, for example shaping decisions about the deployment of non-pharmaceutical interventions in Valencia, providing an exemplar of data science-informed policymaking that has been recognised nationally and internationally. The AI tools created by this team have also received international recognition: a deep learning tool to forecast infection numbers across the world and identify optimal strategies for containing the spread of disease won a $500,000 XPRIZE Pandemic Response Challenge.
Methods and tools developed through this work also provide a proof-of-concept for AI in policymaking. The team’s policy modelling, for example, analyses a range of different policy interventions and their impact on the spread of disease. By allowing policymakers to explore different options and trade-offs, this model connects insights from data to real-world decision-making, opening the door to more effective use of AI for public services in future. In developing these methods, it is vital to understand how to create AI systems that function effectively in deployment, accounting for the limitations of real-world data, designing for the needs of real-world users, and putting people at the centre of AI development. Building on the success of the Data Science against COVID-19 Taskforce, the ELLIS Unit Alicante Foundation continues to advance research that addresses these needs.