Data science and AI in the age of COVID-19 - Reflections on the response of the UK's data science and AI community to the COVID-19 pandemic - The ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Data science and AI in the age of COVID-19 Reflections on the response of the UK’s data science and AI community to the COVID-19 pandemic
Executive summary Contents
This report summarises the findings of a science activity during the pandemic. The Foreword 4
series of workshops carried out by The Alan message was clear: better data would
Turing Institute, the UK’s national institute enable a better response. Preface 5
for data science and artificial intelligence
(AI), in late 2020 following the 'AI and data Third, issues of inequality and exclusion
related to data science and AI arose during 1. Spotlighting contributions from the data science 7
science in the age of COVID-19' conference. and AI community
The aim of the workshops was to capture the pandemic. These included concerns
the successes and challenges experienced about inadequate representation of minority
+ The Turing's response to COVID-19 10
by the UK’s data science and AI community groups in data, and low engagement with
during the COVID-19 pandemic, and the these groups, which could bias research 2. Data access and standardisation 12
community’s suggestions for how these and policy decisions. Workshop participants
challenges might be tackled. also raised concerns about inequalities 3. Inequality and exclusion 14
in the ease with which researchers could
Four key themes emerged from the access data, and about a lack of diversity 4. Communication 17
workshops. within the research community (and the
potential impacts of this). 5. Conclusions 20
First, the community made many
contributions to the UK’s response to Fourth, communication difficulties
Appendix A. Workshop participants and Organising 22
the pandemic, via national organisations, surfaced. While there have been excellent
Committee
research institutes and the healthcare examples of science communication
sector. Researchers responded to the crisis throughout the pandemic, participants
Appendix B. Workshop themes and reports 24
with ingenuity and determination, and the highlighted the challenges of
result was a range of new projects and communicating research findings and
collaborations that informed the pandemic uncertainties to policy makers and the
response and opened up new areas for public in a timely, accurate and clear
future study. manner.
Second, the single most consistent In this report, we outline the workshop
message across the workshops was the participants’ reflections and suggestions
importance – and at times lack – of robust relating to each of these themes, with the
and timely data. Problems around data aim of enabling the data science and AI
availability, access and standardisation community to respond better to the ongoing
spanned the entire spectrum of data pandemic, and future emergencies.
Edited by
Inken von Borzyskowski Anjali Mazumder
Assistant Professor of AI and Justice and Human
Political Science, University Rights Theme Lead, The Alan
College London Turing Institute
Bilal Mateen Michael Wooldridge
Clinical Data Science Fellow, Turing Fellow and Programme
The Alan Turing Institute; Co-Director for Artificial
Clinical Technology Lead, Intelligence, The Alan Turing
Wellcome Trust Institute
Workshop theme leads
Mark Briers (The Alan Turing Institute), Tao Cheng (UCL), Spiros Denaxas (UCL), John Dennis (University
of Exeter), Sabina Leonelli (University of Exeter), David Leslie (The Alan Turing Institute), Marion
Mafham (University of Oxford), Ed Manley (University of Leeds), Karyn Morrissey (University of Exeter),
Deepak Parashar (University of Warwick), Jim Weatherall (AstraZeneca)Foreword Preface
The COVID-19 pandemic has seen scientific and non-experts – contain valuable reflections At the end of 2019, a new highly infectious pharmaceutical interventions. The response
research move into public discourse in and suggestions for how the data science virus, now known as severe acute respiratory was remarkable for its breadth of engagement
unparalleled ways. Across the scientific and AI community might prepare for future syndrome coronavirus 2 (SARS-CoV-2), across disciplines, as demonstrated by
spectrum, researchers have stepped up, the emergencies. Indeed, the Turing has already was identified as the underlying pathogen the range of backgrounds of our workshop
data science and AI community included, to begun a large-scale project1 which aims to for a series of unexplained pneumonias participants and the diverse set of insights that
work alongside clinicians, policy makers and boost societal, governmental and economic (subsequently termed coronavirus disease they generated.
the government at the heart of the response, resilience to shocks such as this pandemic. 2019, or COVID-19), clustered in Wuhan,
directly impacting on our daily lives. China. By 30 January 2020, COVID-19 had Goals, origins and structure of the report
Ahead of the G7 summit in the UK in June become so prevalent that the World Health
Data science and AI is an inherently 2021, the leading scientific bodies of the G7 Organisation (WHO) declared it a Public Health This report was commissioned by The Alan
interdisciplinary community, and our activities nations (the ‘Science 7’) recently published a Emergency of International Concern – the Turing Institute with the aim of reflecting on
at The Alan Turing Institute in response to the call for more ‘data readiness’ in preparation WHO’s highest level of alarm. As we finish the UK's data science and AI response to the
pandemic reflect this. Our researchers have for future health emergencies.2 This is a this report in spring 2021, the disease itself pandemic. The Turing is the UK’s national
developed algorithms to monitor pedestrian timely amplification of the message in the has claimed over three million lives globally, institute for data science and AI, and partners
density and ensure social distancing on the Turing’s report about the need for increased with more than 170 million confirmed cases, with many of the UK’s leading universities and
streets of London; combined NHS datasets to data access and sharing, at a time when the and many more affected by the impacts of research centres to advance the country’s
help answer clinical questions about the effects pandemic continues to have catastrophic lockdowns and the unprecedented disruption capacity and competitiveness in these areas,
of COVID-19; explored what makes people impacts around the world. to the global economy. The UK has now been with the overall mission of changing the world
vulnerable to health-related misinformation; through two waves of the virus, with infections, for the better.
and improved the accuracy of the NHS My thanks to the editors for initiating and
delivering this report, and to all the workshop hospitalisations and fatalities in the second The Turing’s goals in undertaking this work
COVID-19 app. (See page 10 for more details of wave exceeding those in the first.
our response.) theme leads and participants. We hope it were twofold:
will be a valued contribution to the ongoing While pandemics appear to have occurred
This report has been edited by four researchers discussions about the data science and AI 1. To capture the initiatives and resources that
throughout human history, the COVID-19 have been developed by the data science and
with backgrounds spanning AI, data science, response, alongside notable reports from the pandemic is unique in one important respect.
public policy, human rights and medicine. They Centre for Data Ethics and Innovation,3 the Ada AI community during the pandemic.
It is the first pandemic to occur in the age of
have synthesised the views of 96 attendees to Lovelace Institute,4 the Royal Society’s DELVE data science and AI: the first pandemic in a 2. To gather the experiences and insights of
a series of workshops held at the end of 2020, initiative5 and the Royal Statistical Society,6 world of deep learning, ubiquitous computing, this community during the pandemic – what
which aimed to provide a snapshot of the uses among others. smartphones, wearable technology and social worked well, what didn’t, and how we as
of data science and AI during the pandemic, media. It is thus unsurprising that governments a community could respond better to this
and what we as a community can learn from The Turing looks forward to continuing its
role to convene and deliver activities and across the globe, including the UK’s, looked pandemic and future emergencies.
the experience. to data to inform their responses and help
reflections in response to this pandemic, and in The work has its origins in a one-day
While this represents a small part of what preparation for other crises. navigate challenges. The goal was to limit the
spread of the disease and its medical, social conference 'AI and data science in the age
will surely be a larger reflection exercise to of COVID-19,'7 which was held virtually by the
come, the central findings – issues of data and economic consequences. As such, the
Adrian Smith UK government stated that its policies were Turing on 24 November 2020 and featured talks
access and quality; inequality among both the from some of the leading voices in the UK’s
research community and wider society; and Institute Director and Chief Executive “guided by the science”, and later that ending
The Alan Turing Institute lockdowns depended on “data, not dates”. response to COVID-19. The free event attracted
communication difficulties between experts over 1,700 registrants from 35 countries, from
In response, many members of the UK’s data academia, industry, the public sector and the
science and AI community stepped forward, general public.
spearheading initiatives that they hoped would
assist the domestic and international response. A series of themed, virtual workshops followed
These initiatives came from individual in November and December 2020. The
academics, university research groups, the invitation to participate in these workshops
healthcare sector, national institutes and was widely circulated within the UK academic
others. They involved not just experts on community, via social media and the Turing’s
virology and epidemiological modelling, but network of partners and affiliates. The Turing
1
also researchers studying, for example, the used a lightweight reviewing process to select
social and economic consequences of non- the participants, who are listed in Appendix A.
1
1 https://www.turing.ac.uk/research/research-projects/shocks-and-resilience
2 https://royalsociety.org/-/media/about-us/international/g-science-statements/G7-data-for-international-health-emer-
gencies-31-03-2021.pdf
3 https://www.gov.uk/government/publications/covid-19-repository-and-public-attitudes-retrospective
4 https://www.adalovelaceinstitute.org/summary/learning-data-lessons 7 https://www.turing.ac.uk/events/ai-and-data-science-age-covid-19
5 https://rs-delve.github.io/reports/2020/11/24/data-readiness-lessons-from-an-emergency.html
6 https://rss.org.uk/statistics-data-and-covid
4 51. Spotlighting contributions from the
We took the view that, if the report was to and analysis were far more widespread. One
have any credibility, it would be essential to possible reason for this, suggested by the
include the views of as diverse a community CDEI report, is that COVID-19 was such a new
as possible. The scope of ‘data science and
AI’ for the purposes of this report is therefore
phenomenon that there was a lack of data with
which to train AI algorithms.
data science and AI community
deliberately broad, with participants including
ethicists, clinicians, mathematicians, policy We also note that while we as editors have
A key goal of the workshops was to document several healthcare technology companies.
advisors, and many more. done our best to report the suggestions
the contributions from the UK’s data science OpenSAFELY provides secure access to
of workshop participants as accurately as
and AI community in responding to the a database of over 58 million NHS patient
The eight workshop themes are listed in possible, in some cases we have applied
pandemic. In this chapter, we summarise the records, allowing researchers to answer urgent
Appendix B. These themes are themselves a light editing. This especially applies to the
contributions highlighted by the workshop clinical and public health questions related
reflection of input from over 20 experts, who suggestions around data access in Chapter
participants, which demonstrate some positive to COVID-19. Moreover, for the first time it
formed the Organising Committee for the 2, where we have taken care to put the
aspects of the data science and AI response: separates the development of the analysis
conference and workshops. All workshops participants’ aspirations for more accessible
increased data sharing, collaboration across code from the actual data (which never leave
were led by the Centre for Facilitation8 together and open data in the context of the complex
disciplines, and the development of new the servers of the electronic health record
with one or two workshop-specific theme legal and ethical obligations surrounding
datasets, repositories and initiatives that service provider or NHS Digital), and in doing
leads who were selected by the Organising access to sensitive national health, economic
mobilised COVID-19 research. so guarantees fully open and reproducible
Committee to supplement the facilitators’ and social information. Data access is
code. This is a noteworthy development that
expertise with domain knowledge. For each clearly an important issue which needs to We emphasise that these spotlights are occurred largely due to the permissive nature
workshop, the facilitators and theme leads be addressed, and there are many steps to nothing more than that: spotlights. The of the regulatory environment necessitated by
summarised the discussions in a report, enabling responsible data flows. workshops provided a snapshot of the the pandemic.
which have been lightly edited and linked from community’s response, but there were many
Appendix B. These workshop reports provided One final and important word. Our report
other contributions that were not covered in Another new initiative was the COVID-19
the ‘raw data’ for this report. Given the range is preliminary in the sense that we hope
the workshops. An overview of The Alan Turing Genomics UK Consortium (COG-UK),13
and complexity of the workshops, it has not it will serve as a starting point for a more
Institute’s key COVID-19 projects can be found which has sequenced more than 490,000
been possible to include all of the participants’ comprehensive and systematic review of
on page 10 of this report, and further examples SARS-CoV-2 virus genomes to date, providing
views and suggestions, but we as editors have the uses of data science and AI during the
of the community’s response can be found important information on viral transmission and
tried to draw out the main themes. pandemic. The report was originally conceived
in resources such as the CDEI’s COVID-19 the emergence of new variants. COG-UK was
in summer 2020, when daily reported COVID-19
respository.10 Finally, we acknowledge that achieved through the coordination of several
Although our remit includes both data science fatalities in the UK were down to single figures,
there may be some bias in this chapter, with organisations (NHS, the UK’s public health
and AI – reflecting the Turing’s role as the and there was some hope that the virus might
participants more likely to mention projects agencies, the Wellcome Sanger Institute and
national institute for these two subjects – the be largely contained in a single wave. Work
that they are affiliated with. various academic institutions). Having those
majority of the feedback we received from on this report continued during the November
sequences available from the earliest stages of
workshop participants related specifically 2020 to spring 2021 lockdowns when the
Noteworthy datasets and databases the pandemic was key to pushing forward the
to data science, and our report reflects this. disease was again prevalent. In this sense,
development of vaccines.
This chimes with the 'COVID-19 repository many (but not all) of the observations made The UK government’s action to issue statutory
and public attitudes' report9 published by the here relate to the first wave of the pandemic in regulators such as NHS Digital with a Control Existing systems and databases from ISARIC
Centre for Data Ethics and Innovation (CDEI) the UK, from January to summer 2020, prior to of Patient Information (COPI) notice11 to (International Severe Acute Respiratory and
in March 2021, which found that “conventional the onset of the UK’s second wave. We hope make confidential patient data available to emerging Infection Consortium)14 and ICNARC
data analysis has been at the heart of the that the report will be of value to healthcare researchers and policy makers for COVID- (Intensive Care National Audit & Research
COVID-19 response, not AI”. There were professionals, data science and AI researchers, 19-specific purposes had far-reaching Centre)15 were repurposed and expanded for
certainly innovative applications of AI during and policy/decision makers as we continue to consequences, and participants noted the the pandemic, which kept costs down and
the pandemic, but the evidence suggests that manage the ongoing pandemic. positive effect this had on data sharing. enabled the data science community to act
more traditional methods of data collection more quickly. Similarly, linked data from GPs
For example, it paved the way for a new and NHS trusts were available to researchers
statistical analytics platform called through the pre-existing Discovery East
OpenSAFELY,12 a collaboration between the London16 platform. Participants also
University of Oxford’s DataLab, the London highlighted the utility of aggregate data on
School of Hygiene & Tropical Medicine, and critical care patients collected by CHESS
1
10 https://www.gov.uk/government/publications/covid-19-repository-and-public-attitudes-retrospective
1
11 https://digital.nhs.uk/coronavirus/coronavirus-covid-19-response-information-governance-hub/control-of-patient-in-
formation-copi-notice
12 https://opensafely.org
13 https://www.cogconsortium.uk
8 https://www.centreforfacilitation.co.uk 14 https://isaric.org
9 https://www.gov.uk/government/publications/covid-19-repository-and-public-attitudes-retrospective 15 https://www.icnarc.org
16 https://www.eastlondonhcp.nhs.uk/ourplans/discovery-east-london-case-study.htm
6 7(COVID-19 Hospitalisation in England Participants also cited the DECOVID31 project, Other initiatives Workshop participants also noted positive
Surveillance System),17 which was adapted which aimed to create a detailed, frequently examples of public engagement from the
from the UK Severe Influenza Surveillance updated database of anonymised patient The Royal Society was highly active during data science and AI community, such as the
System by Public Health England. health data during the pandemic. The project the pandemic, using its convening power contributions of Professor Christina Pagel and
was initiated during the early stages of the to support the UK’s pandemic response Professor Devi Sridhar, in helping to improve
Participants noted that a major breakthrough pandemic, and involved researchers from The through two initiatives. First, RAMP (Rapid understanding of the pandemic and policy
in the data science community’s response to Alan Turing Institute and four other founding Assistance in Modelling the Pandemic)34 interventions. Participants also highlighted
the pandemic was the opening up (and sharing) institutes. The initial funding, diverted from an brought together individuals with modelling the work by the Ada Lovelace Institute as a
of various mobility datasets18 by commercial existing Turing grant from the Engineering and expertise from a diverse range of disciplines good example of how to effectively engage the
providers, such as Cuebiq,19 Google,20 and Physical Sciences Research Council (EPSRC), to support the COVID-19 pandemic modelling public during lockdown. In May and June 2020
Facebook’s Data for Good.21 These enabled covered the transfer and combination of data effort. For example, workshop participants (during the first UK lockdown) the Ada Lovelace
the detailed spatial analysis and large-scale from two NHS trusts, plus the data analysis highlighted a collaboration between RAMP’s Institute worked with Traverse,39 Involve,40
simulation of viral transmission that was planning. The work continues with in-kind Urban Analytics team and Improbable (a UK and Bang the Table41 to conduct a rapid online
important for generating evidence for the contributions from researchers around the technology company), which developed a deliberation with 28 members of the public to
government and the public on the potential UK, who are analysing the data (now covering ‘micro-simulation’ model of the spread of explore attitudes to the use of COVID-19-related
and actual impacts of lockdowns. Public 185,000 patients) to shed light on four key COVID-19 based on highly realistic synthetic technologies, such as contact tracing apps, in
transport data from Transport for London and questions, including when to put critical data of people’s daily activity – i.e. where they exiting lockdown.42 The aim of the project was
the Department for Transport were cited as COVID-19 patients onto a ventilator, and how go (home, shops, school, work) and for how to test this new engagement methodology,
central to arguments about the effectiveness of patients with long-term health conditions are long. The project’s main outcome was a model which could be used when face-to-face public
the first lockdown. The ability to monitor data affected by COVID-19. Ethicists in the Turing’s of Devon’s entire 800,000-person population, deliberations are not possible, and it resulted
in real time through dashboards (e.g. i-sense public policy programme have been embedded which allowed researchers to compare the in a report on how to build trust in technologies
COVID RED,22 Evergreen,23 GOV.UK24 and within the research teams from the start of the impact of different intervention scenarios at a developed in response to the pandemic.43
Public Health Scotland25) also helped the project to ensure algorithms are implemented local level.35
data science community to respond and report to the highest standards of transparency and Lastly, although not mentioned by workshop
trends to the public. Second, DELVE (Data Evaluation and Learning participants, the editors would like to highlight
bias mitigation. The first results from DECOVID for Viral Epidemics)36 is a multidisciplinary
are expected later in 2021. The Trinity Challenge44 as a particularly
Workshop participants praised the work of group of researchers convened by the Royal exciting initiative to come out of the pandemic.
the Office for National Statistics (ONS) in Finally, participants noted regional data Society to contribute data-driven analysis to This coalition of organisations from business,
supplying data during the pandemic, including management systems that had proved the UK’s pandemic response, providing input academia and the social sector is offering
its work on deaths stratified by ethnicity,26 valuable. The SAIL Databank32 data linkage through SAGE, the government’s Scientific an award pool of £10m for ground-breaking,
the social impacts of COVID-19,27 and how service was instrumental in redeploying Advisory Group for Emergencies. Outputs from data-driven solutions that will help countries
people spent their time during lockdown.28 existing datasets to support pandemic efforts the group include reports on the efficacy of to better identify, respond to, and recover from
There was also praise for the work of Health in Wales. And DataLoch33 is a repository of face masks in tackling COVID-1937 and the risks outbreaks of disease.
Data Research UK (HDR UK), specifically all routine health and social care data for the associated with pupils returning to schools in
its Innovation Gateway,29 which provides Edinburgh and South East Scotland Region, September 2020.38
information about the datasets and tools held which provides data to a range of researchers
by members of the UK Health Data Research to address COVID-19-related questions.
Alliance,30 facilitating the navigation of multiple
resources relevant to COVID-19. ...some positive aspects of the data
1 science and AI response: increased data
1
sharing, collaboration across disciplines,
and the development of new datasets,
17 https://www.england.nhs.uk/coronavirus/wp-content/uploads/sites/52/2020/03/phe-letter-to-trusts-re-daily-covid-
19-hospital-surveillance-11-march-2020.pdf
18 http://theodi.org/wp-content/uploads/2021/04/Data4COVID19_0329_v3.pdf
19 https://www.cuebiq.com
20 https://www.google.com/covid19/mobility
repositories and initiatives that mobilised
21 https://dataforgood.fb.com/docs/covid19
22 https://covid.i-sense.org.uk COVID-19 research.
23 https://www.evergreen-life.co.uk/covid-19-heat-map
24 https://coronavirus.data.gov.uk
25 https://public.tableau.com/profile/phs.covid.19#!/vizhome/COVID-19DailyDashboard_15960160643010/Overview
26 https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/articles/coronavirus-
covid19relateddeathsbyethnicgroupenglandandwales/2march2020to15may2020 34 https://royalsociety.org/topics-policy/Health%20and%20wellbeing/ramp
27 https://www.ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/healthandwellbeing/bulletins/coro- 35 https://www.improbable.io/blog/improbable-synthetic-environment-technology-accelerates-uk-pandemic-modelling
navirusandthesocialimpactsongreatbritain/4december2020 36 https://rs-delve.github.io
28 https://www.ons.gov.uk/economy/nationalaccounts/satelliteaccounts/bulletins/coronavirusandhowpeoplespent- 37 https://royalsociety.org/news/2020/05/delve-group-publishes-evidence-paper-on-use-of-face-masks
theirtimeunderrestrictions/28marchto26april2020 38 https://royalsociety.org/news/2020/07/delve-opening-schools-should-be-prioritised-report
29 https://www.healthdatagateway.org 39 http://traverse.org.uk
30 https://ukhealthdata.org 40 https://www.involve.org.uk
31 https://www.decovid.org 41 https://www.bangthetable.com
32 https://saildatabank.com 42 https://www.adalovelaceinstitute.org/project/rapid-online-deliberation-on-covid-19-technologies
33 https://www.ed.ac.uk/usher/dataloch 43 https://www.adalovelaceinstitute.org/report/confidence-in-crisis-building-public-trust-contact-tracing-app
44 https://thetrinitychallenge.org
8 9The Turing’s response to COVID-19 Understanding vulnerability to health-related
misinformation
Since the beginning of the pandemic, The Alan Turing Institute has been
In response to the growing problem of misinformation around COVID-19, the Turing’s
working to tackle the spread and effects of COVID-19. Here are some of public policy programme launched a project, funded by The Health Foundation, to
our key projects – visit our dedicated webpage for more info. understand who is most vulnerable. They found that people with lower numerical,
health and digital literacy tend to fare worse at assessing health-related statements,
which suggests that developing people’s literacies has the potential to make a big
difference to their ability to identify misinformation. The researchers are now hoping
to feed into government policy-making around measures to counter the problem.
Helping London to navigate lockdown safely
Estimating positive COVID-19 test counts
As London locked down in spring 2020, a team in the Turing’s data-centric
engineering programme began Project Odysseus to monitor levels of The Turing has partnered with the Royal Statistical Society to
activity on the city streets. Working with the Greater London Authority provide modelling and machine learning expertise to the UK
(GLA) and Transport for London (TfL), the researchers fed their algorithms government’s Joint Biosecurity Centre. One of the key outputs so
with data from traffic cameras and sensors to provide anonymised, near- far is a statistical model that uses incoming COVID-19 test data
real time estimates of pedestrian densities and distances. TfL used this tool to estimate (‘nowcast’) the total number of positive tests in local
during the pandemic’s first wave to make numerous interventions to keep authorities. It can take up to five days for PCR tests to be processed
people socially distanced, such as moving bus stops, widening pavements and reported, so these nowcasts will give authorities an earlier
and closing parking bays. picture of the disease’s spread and aid decision-making.
Combining data from NHS trusts Improving the accuracy of the NHS
Initiated by the Turing and four other partners, the
COVID-19 app
DECOVID project has created a detailed database of
anonymised patient health data. A major breakthrough Turing researchers have played an integral role in the
has been the transfer and combination of data from development of the NHS COVID-19 app, providing technical
two NHS trusts, covering 185,000 patients. The work advice to the Department of Health and Social Care. The
continues with in-kind contributions from researchers researchers improved the algorithm behind the app’s Bluetooth
around the UK, who are analysing the data to shed light Low Energy contract tracing technology, so that it more
on four key clinical questions, including when to put accurately calculates the risk that the phone’s user has been
critical COVID-19 patients onto a ventilator, and how in contact with a COVID-positive person. A statistical analysis
patients with long-term health conditions are affected by published in Nature estimated that the app prevented around
COVID-19. The first results are expected later in 2021. 600,000 COVID-19 cases in October-December 2020 alone, due
to people self-isolating following contact with an infected person.
Modelling the spread of COVID-19 in urban areas
Building resilience against future crises
The Turing led the urban analytics workstream of the Royal Society’s Rapid
Assistance in Modelling the Pandemic (RAMP) initiative. Using highly realistic data A new two-year, multidisciplinary project at the Turing is seeking to better
of people’s daily activities in towns and cities, the workstream developed a model understand how interconnected health, social and economic systems are
that simulates COVID-19 transmission at an individual level. A demonstration affected by shocks such as the pandemic. The ‘Shocks and resilience’
model of Devon’s entire population allowed researchers to compare the impact of project will develop a coupled epidemiologic and socio-economic model
different lockdown strategies. The team is now scaling up its model to a national of the spread and societal effects of COVID-19, as well as more generalised
level, and is in dialogue with policy makers about using it to inform decision- models for other complex, socio-economic systems. The overall aim is to
making in this pandemic and future health emergencies. produce data, methodologies and tools to enable policy makers to make
better informed decisions, boosting the resilience of local and national
governments against future shocks.
102. Data access and standardisation Suggestions techniques work robustly, and at scale, with
real-world data, while also considering data
Reflecting on these data challenges, governance principles and practices.
participants made the following suggestions
Despite the contributions highlighted in expertise to the national effort. However, there for the data science and AI community: – Automate data collection systems.
Chapter 1, workshop participants articulated was a strong perception in the workshops that Data collection is resource-intensive, so
a number of challenges that hindered the providing researchers with better data access – Aspire to a research culture in which automated data collection could reduce
efforts of the data science and AI community and generating a more level playing field data are shared as openly as legal and the reliance on frontline staff to record
to contribute to the national effort. The most regardless of affiliation and connections (see ethical obligations permit, with central data using manual systems. This aspiration
prominent of these were around data access also Chapter 3) would likely have helped better repositories, or ‘data lakes’, for cleaned and needs to be set against the obvious need
and standardisation – not having access to address the knowledge and policy challenges anonymised data (or weblinks to data) ready to ensure the rights of individuals and
the desired data, or having data that were not during the pandemic. for analysis. Document which data exist (to communities to privacy and agency.
suitably formatted or documented. avoid duplication of effort) and make them
– Encourage more data sharing
widely available.
Data standardisation agreements. Develop more general data
Data access – Data repositories should signpost to sharing agreements as opposed to those for
Many participants also noted a lack of data
open datasets and, for non-open data, narrow purposes/groups of users to reduce
While some members of the data science and standardisation as a significant hurdle for data
to the details of data sharing agreements barriers and delays.
AI community had easy and rapid access to scientists during the pandemic. Different data
and protocols for securely accessing
relevant data, several participants commented standards and codification of metadata, and – Develop protocols for minimum
the datasets. This can ensure the wider
that many researchers were limited in their lack of dataset documentation, meant that data standardisation of data fields (including
provision and availability of shared datasets
contributions because access to data was were difficult to find, link and assess in terms of some baseline data to always be included)
and equitable data access.
often restricted, inconsistent or slow. missingness and biases, limiting the scope of to ease linking of datasets and comparable
and confidence in analyses. – Investigate ways in which access to metadata. As part of this, consider following
For example, some hospitalisation data sensitive data (e.g. from the NHS) may the ‘FAIR Guiding Principles for scientific
were not available early in the pandemic, For example, ideally it would have been be enabled while respecting professional, data management and stewardship’:47
and geographically disaggregated data for possible to link datasets from different studies, ethical and legal obligations surrounding the Findability, Accessibility, Interoperability and
local analysis and crafting of solutions (e.g. but many studies and consortia funded data. This will require developing new ways Reusability.
local lockdowns) were also not available by UK Research and Innovation (UKRI, e.g. to allow researchers to securely access
to all relevant academic groups. Further, ISARIC4C45 and PHOSP-COVID46) are relatively – Conduct a stress test of the new
personal data, perhaps using differential
local-level economic data on employment standalone. It would also have been beneficial standardisation specifications by using
privacy or federated learning techniques.
and production by industry sector were to standardise codes across datasets, including regular events (such as the annual influenza
This in turn presents new research
not accessible or existent (and possibly variables for vulnerable and underserved season) to prepare for the next pandemic.
challenges around making these latter
often both), making the economic impact of groups, ethnicity and socio-economic status.
proposed lockdown measures on different Better data standardisation would have
industries, such as the hospitality sector, enabled data scientists to more easily find,
more difficult to assess. More broadly, understand, link and triangulate information
systematic data on non-pharmaceutical from different datasets in order to better
interventions (social distancing, mask wearing, answer policy-relevant questions and to allow
lockdowns), and particularly compliance testing for individual-level or context effects on
with such interventions, have not been made outcomes (e.g. healthcare, employment, public
appropriately available, making it difficult information). As editors, we acknowledge that
to measure the impact of these policies on this aspiration for linked and standardised
behaviour. datasets will be difficult to achieve in the near-
Consequently, insights from the data science
term: different research communities have
1
Aspire to a research culture in which
and AI community were not as informative and
different standards, formal or informal, different
modes of working, and often use similar data are shared as openly as legal
robust as they could have been. It is unclear
how much the quantity and quality of research
terminology with quite distinct meanings.
Nevertheless, the value of such datasets is
and ethical obligations permit, with
could have been increased if a larger share of
1the community had been able to contribute its
clear, and much more could clearly be done. central repositories, or ‘data lakes’, for
cleaned and anonymised data ready
for analysis.
45 https://isaric4c.net
46 https://www.phosp.org 47 https://www.go-fair.org/fair-principles
12 133. Inequality and exclusion
not have access to the internet or a computer, Suggestions
or who have not engaged with initiatives due to
historical distrust within their community. Reflecting on these challenges, participants
made the following suggestions for the data
The COVID-19 pandemic has brought societal level. For instance, participants cited problems Finally, participants commented that the science and AI community. We note that many
inequality into sharp focus, with the disease around the standardisation of capture of socio- data science and AI community had been of the suggestions in Chapter 2 regarding data
having a much greater impact on some groups economic status and ethnicity. insufficiently engaged with disadvantaged access and standardisation are also relevant
than others. In England and Wales during the groups on what COVID-19-related research to the data privilege issues highlighted in this
first wave of the pandemic, nearly all minority While participants acknowledged that a robust
would be useful for them, with the Ada chapter.
ethnic groups had higher mortality rates than and rigorous system exists for obtaining
Lovelace Institute cited as one of the few
the White ethnic population.48 There were epidemiological data via institutions such as – Prioritise understanding the impact of
organisations talking to these groups.
also marked socio-economic differences, with Public Health England, they noted that such COVID-19 on different ethnic and social
mortality rates in the most deprived areas a system does not exist for wider economic, groups, and ensure that these insights
mobility and socio-economic data, and that this Inequality and exclusion within the research
around double those in the least deprived are considered in conjunction with other
gap hampered socio-economic research in community
areas.49 key risk factors (e.g. age, pre-existing
the early weeks of the pandemic. Participants Many of the issues around data access health conditions, profession, care home
Alongside these broader societal issues, also noted a lack of sufficiently granular, local- highlighted in Chapter 2 are also issues of residency). This would help promote
problems of inequality and bias relating to data level data on transmission of COVID-19 across inequality, with some researchers having much inclusion of underserved groups by funders,
science and AI have also been highlighted communities, which hid some of the social easier access than others to sufficient quality researchers designing studies, reviewers
by the pandemic. There have been long-term inequalities associated with the initial phase of data. Workshop participants recognised that evaluating focus areas, and teams delivering
concerns about the potential of these research the pandemic. researchers with access to certain people modelling projects.
areas to marginalise certain groups, but the or institutes were more privileged, leading to
rise during the pandemic of new datasets, data – Address deficiencies in data sources. For
“data haves and data have-nots”. For instance,
example, the physical, mental health and
capture/analysis methods and data-driven "The COVID-19 pandemic has those who had access to members of SAGE
economic consequences of the pandemic
technologies such as contact tracing apps, or who had established relationships with
plus ongoing discussions around immunity brought societal inequality agencies in health and other relevant areas
are likely to be severe on some groups, and
and vaccine passports, has meant that these capturing any missing or incomplete data
issues are more pertinent than ever. We need into sharp focus." were able to contribute more easily to the
research response.
could be key to understanding the medium-
ongoing scrutiny to ensure that any tools and and long-term impacts on underrepresented
technologies developed by the data science There was also geographical inequality within groups.
Participants raised concerns that the scramble
and AI community respond to people’s health, the data science response, with some areas – Develop clear protocols for collecting
to use existing datasets, or to rapidly create
socio-economic and security needs in a fair of the UK having far greater access to local data on protected characteristics,
new ones, ran the risk of proceeding without
and equitable manner. data and success working with it than others. including, for example, ethnicity, sex,
due regard for sampling biases. These biases,
Participants commented that higher quality, age and socio-economic status. Better
Inequality and exclusion were common such as insufficient representation of minority
spatio-temporal data could have assisted the monitoring of this information in studies and
themes in the workshops, and the participants’ groups, could be ‘baked into’ datasets, and
support of policy at both a local and national clinical trials could help identify critical data
comments fell into two main categories: may be due to systemic discrimination,
level. gaps.
structural inequalities and/or data collection
– Inequality and exclusion within society constraints. This could in turn lead to biased Finally, participants noted a lack of diversity – Develop clear protocols for generating
research and policies that exacerbate pre- in the data science and AI community, anonymised and synthetic data, so that
– Inequality and exclusion within the research
existing inequalities. Data availability, detail which could bias research or policy in important demographic information can
community
and representativeness during clinical trials a number of ways, and emphasises the be included in open datasets (one example
also had shortcomings, which inhibited the need to make diversity more of a priority is the OpenPseudonymiser51 application
Inequality and exclusion within society
inclusion of patient subgroups, and may have in recruitment. This lack of diversity could developed at the University of Nottingham).
Participants felt that the potential of the data made it more difficult to assess, for example, lead to lower engagement with vulnerable Great strides have been made over the past
science and AI community to help understand the efficacy of COVID-19 treatments and and underrepresented groups, for instance, decade in developing realistic synthetic
the impact of COVID-19 on different ethnic and vaccines in minority ethnic groups. potentially resulting in data and policy biases simulation data (e.g. in ‘digital twins’
socio-economic groups was not fully realised due to insufficient representation. The low projects). Such work can be leveraged and
due to several challenges. Participants also noted significant inequality
take-up of vaccination in some minority extended to provide synthetic datasets for
and exclusion challenges around the use of
Many participants commented on a lack of ethnic groups demonstrates the need for modelling potential future pandemics and
mobility, mobile device and other digital data in
relevant or consistent data. While ONS and communication and engagement by trusted their impact on minority groups.
behavioural analysis. For instance, traditionally
other public data providers released up-to-date representatives50 – and these often differ from
underrepresented groups also tend to be
COVID-19 data on minority groups, many data the demographics of national representatives.
scarce in digital data – especially those on the
sources had gaps at both the national and local ‘invisible’ side of the digital divide, who might
1
43
50 https://www.theguardian.com/world/2021/mar/25/clinics-pop-up-in-london-to-help-low-vaccine-take-up
51 https://www.openpseudonymiser.org4. Communication
– Make more granular data available at the – Increase the diversity of representatives
local level, and expertise and resources in academia, the government and
easier to find and deploy, so that local the media. Increasing diversity and
organisations have the autonomy to work descriptive representation in organisations
with their data (as opposed to the data – among researchers and UK government Although collaboration within the data science – Make greater strides to bring
being held, and decisions made, centrally). representatives alike – can create smarter, and AI community increased during the transdisciplinary groups together. As
This can increase local resilience to health more innovative and more inclusive teams pandemic, as demonstrated by the examples well as research groups, this includes
emergencies. and solutions.52 For example, diverse in Chapter 1, issues of communication were local communities, health and social care
research teams will be key to addressing frequently raised by workshop participants, providers, local partners, analysts, third
– Increase engagement with minority falling into three main categories: sector organisations, private organisations,
the issue of algorithmic bias. Increasing
and underrepresented communities. and industry.
diversity also sets role models for future – Communication between experts
Improve data representativeness, minimise
generations and thus provides lasting – Speed up the research pipeline by
algorithmic bias and develop trust by – Communication between researchers and
benefits. One way to improve diversity is developing a robust research/analysis/
engaging and consulting with representative policy makers
to make it a core criterion in hiring and review framework to reduce delays in
groups, such as ‘citizen groups’, at regular
appointment decisions. – Communication between researchers and publishing ideas while maintaining high
intervals. Develop mechanisms for well-
funded community involvement in setting the public standards of quality and transparency.
research priorities and for the end-to-end – Promote global data sharing for health
participatory design of projects. Communication between experts emergencies. Data sharing agreements
Connecting expertise – the right people to between the UK and other countries take
the right data to the right problem – was time to organise, and many were previously
identified by workshop participants as an coordinated and/or financially supported by
important bottleneck in working to understand EU organisations, which will undoubtedly
COVID-19, constraining the community’s become a more complicated endeavour
ability to respond. More collaboration between post-Brexit. Having more flexible regulations
often disparate groups, such as data asset for future emergencies will help facilitate
holders, subject matter experts, researchers global learning as well as research and
and clinicians, could have helped to share policy solutions in the UK. The editors
expertise, avoid duplication, and improve note that facilitating safe and equitable
analyses. data sharing between countries is also
recommended by the 'Data for international
Participants also identified a need for more health emergencies' statement published by
international collaboration. While data the science academies of the G7 nations in
access and exchanges across national March 2021.54
borders happened to some degree during
the pandemic, such as through the EULAR Communication between researchers and
COVID-19 Database,53 many other efforts policy makers
remained decentralised, informal and ad hoc.
More could be learned by the UK from other The transfer of knowledge from academic
countries in similar situations, particularly research to real-world policy was a crucial
Increasing diversity and those with greater experience or capacity in
dealing with health emergencies.
aspect of the scientific response to the
pandemic. Within the data science and
descriptive representation AI community, participants noted that
communication with policy makers could be
Suggestions
in organisations – among Participants made the following suggestions
improved in order to better identify policy
needs and improve knowledge transfer.
researchers and UK government for the data science and AI community:
Participants also highlighted a lack of
representatives alike – can
1
– Develop collaborative working
relationships between data scientists
transparency in policy-making. It was difficult
to know which studies had ‘cut through’ and
create smarter, more innovative and clinicians. Data scientists can provide been considered by government and advisory
insights about the collection and storage groups when making policy interventions, and
and more inclusive teams and of data that clinicians may not be aware which data policy makers were using to inform
of, while clinicians can provide valuable their decisions. Increased transparency would
solutions. 1
insights about the multi-dimensional nature help researchers to understand what scientific
of health and social data collection. approaches and insights are most valuable to
53 https://www.eular.org/eular_covid19_database.cfm
52 https://hbr.org/2016/11/why-diverse-teams-are-smarter ; https://www.turing.ac.uk/news/where-are-women-map- 54 https://royalsociety.org/-/media/about-us/international/g-science-statements/G7-data-for-international-health-emer-
ping-gender-job-gap-ai gencies-31-03-2021.pdf
16 17policy makers, and would also give the public Workshop participants agreed that, although – Increase the understanding and – Use data visualisation: this can be
a clearer picture of the evidence behind policy there had been successful examples of public involvement of data scientists in media an effective means of communicating
decisions, potentially increasing public trust engagement from the community during the processes. Media outlets have increasingly complex scientific topics to non-specialists.
and compliance. pandemic, there were also shortcomings referenced preprints, which are of Participants cited Harry Stevens’s ‘corona
in communication, particularly around the varying quality given preprint protocols simulator’ article in The Washington Post as
limitations and uncertainties of research. and accelerated peer review processes. a particularly effective example.59
"The transfer of knowledge Helping the public to understand the findings Communicate to the public what is and what
– Support independent input and
and caveats of modelling studies, for example, is not (or not yet) quality, peer-reviewed
from academic research could enhance trust in the research and research.
communication: workshop participants
highlighted the benefits of an Independent
to real-world policy was a increase support for and compliance with
policies. Trust and acceptance of policies might
– Consider building on pre-existing SAGE,60 which provided a model for
mechanisms to support communication, how scientists can analyse, interpret
crucial aspect of the scientific also be increased by communicating how data
for example via the Science Media and comment on a situation from a more
science and modelling can promote positive
response to the pandemic." health outcomes. For example, targeted
Centre,58 which curates expert reactions
to noteworthy science stories, helping
politically neutral standpoint.
communication might have helped improve
journalists to access nuanced and unfiltered
compliance with the isolation notifications sent
information.
Suggestions through the NHS Test and Trace system.57
Participants made the following suggestions
for the data science and AI community: Suggestions
Participants made the following suggestions
– Build sustainable links between the data
for the data science and AI community:
science community and those who are
closer to policy-making. – Provide clear, simplified, accessible
– Provide training for policy makers to help information for non-specialists about
them understand the research process, studies’ findings and predictions, as well as
e.g. study design and limitations, and the the underlying research designs (data used,
uncertainty around estimates and findings. data quality, computational processes, etc.).
– Provide training for researchers and – Communicate transparently about the
public health professionals to help them limitations of studies, such as uncertainty
better communicate to policy makers the in the models/predictions and potential
findings, limitations and uncertainties of biases in the data.
their studies and models. – Build public trust by addressing concerns
related to data security, privacy and
Communication between researchers and confidentiality, and by communicating how
the public the data and models are informing policies.
Throughout the pandemic, data science and – Take a more proactive role in countering
AI have been in the public eye as never before. misinformation, for example about testing,
Every day presents a new swathe of statistics contact tracing and vaccines. More effective
about COVID-19, and data scientists have communication could have mitigated
been increasingly called upon to communicate against the many counter-messages, ‘fake
their research to non-specialists. Meanwhile, news' stories, and biased reporting that
the public has been able to directly input into emerged during the pandemic. Also, present
COVID-19 research through initiatives such as data in a way that is complete and not prone
the COVID Symptom Study app,55 and there to misinterpretation.
have been vigorous debates on the ethics of
– Train researchers in how to communicate
AI and algorithmic bias, especially around the
their findings to the public and media,
1UK GCSE and A-Level grading controversy in
summer 2020.56
especially around the use of models in Throughout the pandemic, data
making decisions and predictions.
science and AI have been in the
1 public eye as never before.
55 https://covid.joinzoe.com 58 https://www.sciencemediacentre.org
56 https://en.wikipedia.org/wiki/2020_UK_GCSE_and_A-Level_grading_controversy 59 https://www.washingtonpost.com/graphics/2020/world/corona-simulator
57 https://www.medrxiv.org/content/10.1101/2020.09.15.20191957v1 60 https://www.independentsage.org
18 195. Conclusions Acknowledgements
Since the first UK lockdown in March 2020, is key to reducing the chances of data and We thank all conference and workshop This work was supported by Wave 1 of the
when little was known about COVID-19, our research being misused or misinterpreted. participants who have shared their experiences UKRI Strategic Priorities Fund under the
knowledge of the disease and its effects have and insights with us, and have contributed to EPSRC Grant EP/T001569/1, particularly
much progressed. The UK’s data science and Workshop participants have made many this report. the ‘Health’ theme within that grant and The
AI community has played an integral role in suggestions for how the community might Alan Turing Institute. We also acknowledge
this scientific effort, and this report provides address these challenges. These include We thank Jessie Wand for her amazing efforts the contributions of the ‘Health and medical
a snapshot of the community’s contributions. the provision of accessible and centralised in organising the complex series of conference sciences’ programme at The Alan Turing
But equally, the pandemic has pulled into sharp ‘data lakes’; more equitable data access; meetings and workshops that led to this report, Institute.
focus a number of areas where we can and protocols for data standardisation; increased and James Lloyd, science writer, for his tireless
should do better. representation of, and engagement with, efforts in pulling the final draft of this report
minority groups; training for researchers in together. We are also grateful to the Centre
First, participants reported difficulties communicating their work to non-specialists; for Facilitation for holding the workshops and
accessing sufficiently timely, robust, granular, and initiatives to increase public understanding writing the workshop notes.
standardised and documented data. There of research findings and uncertainty, to better
were also issues of privilege, with some counter misinformation.
researchers able to access data much more
easily than others. If the community can make progress in these
areas, then when we are next faced with a
Second, the pandemic has highlighted pre- pandemic – and the historical record strongly
existing societal issues of inequality and suggests that this is a ‘when’ rather than an ‘if’
exclusion, and the role that data science – we should be better placed as a collective to
and AI can play in mitigating or exacerbating respond.
these. In our workshops, participants raised
concerns about a lack of data about, and Navigating our way through the pandemic
engagement with, minority ethnic and socio- without the knowledge and resources of
economic groups, and a lack of diversity within the data science and AI community would
the research community and decision-making have been markedly more difficult. These are
organisations. transformational times for the community as
its research becomes ever more embedded
Third, the pandemic has underlined both the in everyday life. We need to draw on our
importance and difficulty of communicating experiences during this pandemic to ensure
transparently with other researchers, that data science and AI continue to change
policy makers and the public, particularly lives for the better.
around issues of modelling and uncertainty.
Participants agreed that better communication
20 21You can also read
