AI Strategy

Accelerating
Artificial Intelligence
in health and care:
results from a state
of the nation survey
AUTUMN 2018
Accelerating Artificial Intelligence in health and care:
results from a state of the nation survey 3
CONTENTS
4: Foreword
6: Introduction
8: Executive summary
10: What do we mean by AI in health and care?
14: Results of the national survey about AI

technologies in health and care
22: Real world analysis: feasibility and

implementation
38: Summary and next steps
44: Appendix 1: Case studies
48: Appendix 2: Further reading
52: Appendix 3: Glossary
59: Acknowledgements
62: About The AHSN Network AI Initiative

4 Accelerating Artificial Intelligence in health and care:
results from a state of the nation survey
FOREWORD
New technologies that harness envisions what can be achieved As it stands, the NHS is primed to We have already made some industry, academia, innovators By working together we will
the power of data, like artificial when the vast potential of AI is use AI to improve its efficiency, important steps forward in and commissioners we will be able to explore all potential
intelligence (AI), present huge unlocked. The report is based deliver better outcomes and this area. These include the continuously iterate the principles avenues of opportunity, and risk,
opportunities to transform on a survey conducted by NHS prevent ill health. However, introduction of a new national and guidance contained in the and to make sure that none are
healthcare, improve the quality England and the AHSN Network we must be realistic about the data opt-out and by the Bill, Code. Together, we will work to missed. We hope that, based on
of people’s lives, and to make the AI Initiative, and it underlines the challenges. First and foremost, currently before Parliament, to ensure that the NHS gets the this reassurance and reflecting on
job of working within the health potential for AI to contribute to the public must have confidence put the National Data Guardian maximum possible benefits from the information presented in this
and care system more rewarding. improved care: 94% of the UK’s AI that AI (and the health data which on a statutory footing. At the end these partnerships, both for enlightening report, you are left
We are determined to harness thought-leaders cite AI as being fuels the development of new of this report we point to, for the existing use cases of AI and those feeling as optimistic as we are
this potential. extremely important or very algorithms) is being used safely, first time, a Code of Conduct for that appear in the future. And, of about the ability of technology,
important for diagnostics; 89% legally and ethically, and that Digital Health Innovations and course, these developments all and AI in particular, to transform
While these opportunities are
support this view for operational the benefits of the partnerships Intelligence Algorithms, which is take place in the context of our health and care.
available to every country, the
and administrative goals; and, between AI companies and the designed to provide a national review of the current regulatory
UK is well-placed to take a global
79% have this opinion in regard to NHS are being shared fairly. As set of ‘rules of engagement’ for framework and analysis of the
advantage in this field. By virtue
the benefits for health promotion a consequence, realising the any NHS organisation entering future needs of the health and
of our universal single-payer
and preventative health. The potential of AI in health and into a partnership with an AI care workforce.
system, the complete longitudinal
report cites many exciting care requires changes to data developer.
datasets the NHS holds on every
examples of pilot schemes and infrastructure, organisational A collaborative approach is
citizen’s health and care, and
more developed programmes structures, commercial Inevitably, there is still more to do important; no single partner in
our world-leading AI and tech
that are already delivering better arrangements, and models to seize the opportunities ahead. this endeavour has a monopoly
industries, our goal should be to
healthcare for British patients. of consent. By working collaboratively with on wisdom about what will work.
bring the transformative power of
AI to every corner of the NHS.
For that reason, we are delighted
to introduce this ‘state of the
nation’ report, which looks to the
future of health and care and
Matt Hancock Lord O’Shaughnessy

Secretary of State Parliamentary Under
Secretary of State
Department of Health
and Social Care Department of Health
and Social Care
Photo attribution: Chris McAndrew [CC BY 3.0 (https://

creativecommons.org/licenses/by/3.0)], via Wikimedia Commons
INTRODUCTION
The report is split into
In recent years there have been a
number of policy reports published
intelligent technology is being
realised. This survey went live to
four main sections:
on the potential for artificial the nation at the start of 2018 and
intelligence in healthcare. In this we have captured throughout this What do we mean by AI in health
report we are not attempting to report the initial findings from the 1 and care? This section describes
recreate that content but rather 131 responses. how AI is broadly defined and shows
to address some of the concerns how, as AI evolves, it is becoming an
In order to present a rich picture
raised and outline some of the increasingly complex landscape.
of the nation’s ecosystem and
emerging policy in this arena within
bring to life the complex and
the UK.
multifaceted aspects of the
Results of the 2018 national survey
We developed a survey in industry, the report also highlights
collaboration with industry, a number of case studies that set
2 of AI technology in health and care,
and the defining characteristics of
academia and policy makers in an the scene for the work needed to
the first 131 solutions that were
attempt to capture the reality of scale up evidence-based solutions
submitted.
what technology is actually being that are safe, effective and offer
developed within the UK health value going forward.
and care sector, and to understand Real world analysis of feasibility
what complexity of artificial 3 and implementation based on
evidence from over 100 leaders
and pioneers working in the field.
This highlights the top barriers and
enablers for catalysing an ethical,
evidence-based market for AI-
enabled solutions in health and care,
and defines the issues that will set
the agenda for the sector over the
coming months and years.
A summary of proposed next steps.

4 This includes the key themes for
policy makers to develop a ‘Code
of Conduct’ for an AI-enabled
digital health and care market
going forward, and the regulatory
challenges that need to be
addressed.
EXECUTIVE
SUMMARY
Over the last few years numerous The survey The results shows that AI has • ground AI solutions in real Showcase Code of conduct
reports have been written about huge potential to transform ‘problems’ as expressed
The survey was developed with whole the health and care by the users of the health To show what can be achieved as To address these challenges, a
the opportunities and benefits
the input of the AHSN Network system. Unlocking value in data/ system; AI is embraced across health and number of workstreams have
artificial intelligence (AI) can
AI Initiative Core Advisory Group analytics was the top category care, this report showcases some already been initiated across
offer for healthcare. These have (individual members are listed • engage healthcare
(75%) addressed by solutions of the emerging examples of the UK health and care sector.
ranged from the Reform report1 in Acknowledgements) and sent professionals and create an
submitted for the survey, more complex AI methodologies These include the Topol review
illustrating the areas where AI out nationally via the AHSN ethical framework to enhance
followed by condition recognition currently being used and which on workforce; a set of principles
could help the NHS become Network and a number of AI and and preserve trust and
(60%) and organisational hold significant potential to and guidelines summarised in
more efficient to the report by innovation networks including transparency;
processes (50%). deliver impact at scale in NHS, a Code of Conduct for digital
Future Advocacy2 which reviews the AI community run by NHS • build capacity and capability; social care and, importantly, health innovations incorporating
the ethical, social and political Horizons. The survey results are Whilst impressive, the survey • ensure the regulatory in preventative health. As the intelligent algorithms; and
implications of AI in health and self-reported and were compiled shows that many solutions are framework is fit for purpose; complexity and capabilities of a number of initiatives to
and analysed by a team at Kent primarily in their infancy and have these projects increase, it is vital understand and unlock the value
medical research. • explore innovative new
Surrey Sussex Academic Health a long way to go before the true that the policy and organisational of data to provide maximum
funding and commercial
While acknowledging these Science Network, supported by potential of AI for health and care contexts, processes and benefit to citizens and UK plc.
models; and
reports exist, we felt there was Health Education England Kent can be realised. As one survey regulation evolve to keep pace.
a need to understand what is Surrey Sussex. This report is a respondent commented ‘AI is • focus on building a
collaboration between the AHSN still evolving... it won’t solve all sound data infrastructure
actually happening on the ground
Network, NHS England, NHS the problems healthcare faces at and high quality data
and what is being developed. We
Digital and the Department for the moment’ and we must avoid sets, underpinned by
also wanted to ask people within interoperability and sharing
Health and Social Care. the trap of ‘overhyping potential,
the health and care system who standards.
Survey respondents included unrealistic claims and poorly
use artificial intelligence (which
CEOs, senior managers and others thought out products.’
is summarily defined as a series Furthermore, momentum is
working across the AI ecosystem The survey revealed that realising
of advanced technologies that starting to build for unlocking
in England. They represented both the truly huge potential of AI open innovation through
‘Focus on building
enable machines to effectively large organisations with 250 staff
carry out complex tasks that to transform health and care establishing open data
or more (32%) as well as micro services will require overcoming
would require intelligence if ecosystems across health and
organisations with less than 10
completed by a human) what
stage of deployment their work
staff (28%) across private, public
and charitable sectors as well as
several key barriers, and
working together across the
care.
a sound data
AI ecosystem to:
has reached. academia.
infrastructure and
high quality data
sets, underpinned by
interoperability and
Harwich, S. and Laycock, K. (2018). Thinking on its own: AI in the NHS. Reform. Available at: http://www.reform.uk/wp-content/
sharing standards’
1
uploads/2018/01/AI-in-Healthcare-report_.pdf.
2
Fenech, Matthew, Strukelj, Nika and Olly Buston (2018). Future Advocacy and Wellcome Trust. ‘Ethical, social and political challenges of
artificial intelligence in health’. Available at: https://wellcome.ac.uk/sites/default/files/ai-in-health-ethical-social-political-challenges.pdf.
WHAT DO
There is no single, universally better understand clients’ Thanks to advances in AI and
agreed definition of AI, nor indeed current and potential future Big Data research, narrow AI
of ‘intelligence’. Broadly speaking, financial needs. technologies have the potential
intelligence can be defined for wide application in health and
WE MEAN BY
as ‘problem-solving’, and ‘an • Ambient (Intelligence) - social care, bringing benefits to
intelligent system’ as one which the application of several individuals, families, communities,
takes the best possible action in technologies (including and society as a whole. While
a given situation. Artificial or Augmented early examples from our survey
Intelligence, but also sensor
AI IN HEALTH
illustrate that much of this work
The ‘A’ of AI generally refers to networks, user interfaces, is at an early stage, current
one of the following: home automation systems, technologies support a more
• Artificial (Intelligence) – makes etc) to create proactive ‘smart’ general shift away from reactive
environments.
AND CARE?
it possible for ‘machines’ to care models to models that are
learn from new experiences, more personalised and proactive.
AI is generally classified into the
adjust outputs and perform
following types: But this is not without its
human-like tasks. It can be
• Narrow AI typically focuses on challenges in health and social
thought of as the simulation
a narrow task, or works within a care and more widely – ensuring
of human intelligence and
narrow set of parameters such these technologies are fit for
could include voice and visual
as reading radiology scans, or purpose, ensuring outputs are
recognition systems.
transparent and explainable, and
AI describes a set of advanced technologies that enable machines to • Augmented (Intelligence)
optimising hospital workflows;
ensuring people are trained in the
carry out highly complex tasks effectively – tasks that would require - outputs that complement • Strong or general AI is a use of these new technologies.
human intelligence, hypothetical concept which
intelligence if a person were to perform them. emphasising AI’s can refer to an AI that can learn
supplementary role. Examples to perform several different
include tools that support types of task, or to a sentient
radiologists in reviewing large machine with consciousness
numbers of scans, or that and mind.
support financial advisors to
Example of how simple components form modules

which then form a complex AI application
Complexity
There is a significant amount of effort being
devoted in the research space to map machine
learning and AI, but it has been challenging to
Clinical
scale in AI
categorise them according to their ‘intelligence’. decision
Thus far, attempts at categorisation have been support Complex
limited to looking at their generic ability to solve system application
new problems, and at the speed with which they
adapt to these problems.
A more straightforward way of understanding AI
is to classify AI systems by their complexity. A
‘Complexity Scale for AI’ can be seen in the boxed Inference User
engine interface Modules
section and compared with methods and case
studies revealed in our survey.
A glossary of AI-related terms used within the
Complexity Scale for AI can be found at Appendix 3
on page 50.
Components
Algorithm derived of modules
Link to EPR for Interface widgets Context-sensitive
by ML from
current case & resource files help
historical dataset
High complexity Middle complexity Low complexity

AI applications AI modules or AI reasoning methods
Devised by Jeremy Wyatt, Director and Professor of Digital Healthcare, Wessex Institute of Health Research,
components Clinical Advisor on New Technologies, Royal College of Physicians, Fellow of American College of Medical
Informatics & UK Faculty of Clinical Informatics)
• Autonomous vehicle • Natural language • Deep learning module
• Machine translation tool to SNOMED code • Ensemble methods (e.g. Random Forest Models)
processing module
• Care companion robot • Neural networks
• Image processing
• Chat bot • Object segmentation algorithm
module
• Surgical or pharmacy • Signal processing algorithm / filter The lowest level of the complexity The above provides an example state that they use ‘AI’. We would
• Text to speech module
robot • Generative adversarial networks scale comprises single specific of a complex AI application. also like to encourage those
• Knowledge based or
• Mammogram
expert system module • Time series analysis reasoning methods (e.g. neural investing in these technologies
interpretation system Algorithms in healthcare are not
• Signal processing & • Graphical models networks, pattern recognition to understand what type of AI is
• ECG interpreter a new phenomenon and have
classification module • Decision trees, rule induction e.g. CART algorithms). When these being developed, how complex it
• Diagnostic decision been deployed for decades.
• Recommender module • Clustering algorithm reasoning methods are combined is, and indeed question what the
support system What we have attempted to
• Classification algorithm with other functions (e.g. a ‘A’ in AI truly represents.
• Speech driven radiology show here is how technology
report tool with SNOMED • Regression – linear, multiple, logistic database or user interface), we
utilising intelligence within its
coded output • Inference engine for rules or frames get ‘modules’, which sit at the
algorithms can fall under many
• Argumentation, temporal or spatial reasoner e.g. QSIM next level of complexity and are
different subsections and with
• Text generator using DCGs the problem-solving components
varying degrees of complexity.
• Case-based reasoning algorithm of a system. At the top level of
We encourage developers and
complexity, we have applications
industry to be transparent as to
Devised by Jeremy Wyatt, Director and Professor of Digital Healthcare, Wessex Institute of Health Research, Clinical Advisor on New or packaged systems comprising
what complexity or methodology
Technologies, Royal College of Physicians, Fellow of American College of Medical Informatics & UK Faculty of Clinical Informatics) two or more of these modules
they are utilising when they
(e.g. an autonomous robot).
Complexity of
current projects
As part of the survey we asked complexity scale (see previous At the highest end of the
respondents to list some of the section), it can be seen that many complexity scale, 8% of solutions
AI methods employed in their of the current solutions are using employed machine translation
solutions. This enabled them ‘lowest complexity’ advanced methods. 20% of solutions
to be categorised in a way that statistical techniques rather than indicated they used ‘other’ AI
shows how solutions in the AI more complex AI applications. methods, including a range of
RESULTS OF
space vary greatly in terms of Classification and neural network chat bot solutions (considered
complexity. machine learning methods ‘highest complexity’ solutions).
were by far the most popular
By mapping some of the methods
techniques, used by 60% and 51%
THE NATIONAL
employed by survey respondents
of solutions, respectively.
against Professor Jeremy Wyatt’s
SURVEY ABOUT AI
The percentage of solutions
Case studies reporting using a method of AI
delivering
TECHNOLOGIES IN
Lowest complexity
value now Classification 60%
A range of case studies Neural networks 51%
HEALTH AND CARE

identified through the Decision trees 39%
survey at various stages
Clustering 36%
of maturity (from those at
Time series analysis 33%
research stage through to
examples with regulatory Ensemble methods 25%
approval and/or publicly Regression 25%

available) are listed in Graphical models 16%
Appendix 1. Generative adversarial networks 11%
This section presents key findings from 131 self-reported entries in These solutions are Knowledge based/expert systems 37%
delivering value to the Image processing 33%
response to our survey that began in Spring 2018. The information health and care sector in
has been used to create an online map that illustrates what sort of the following areas: Middle complexity
• Unlocking value in data/ Text to speech 9%
problems are being solved currently, who some of the key players analytics Natural language processing 38%
• Leveraging skills and
are, and how we can group or categorise current projects to help our capacity Highest complexity
• Organisational
understanding of the current reality of AI in health and care. Machine translation 8%
processes Other (please specify) 20%
• Condition recognition.
Unsure/Not applicable 9%
Diagnostics
Unlocking Condition With wide consensus that diagnostics presents
value in data/ recognition some of the strongest early AI use cases, we chose
to make it a special focus for this first AI map and
analytics 75% 60%
in focus
survey.
Across the range of diagnostics categories, AI is
already offering opportunities to free up workforce
capacity and to dramatically increase diagnostic
The percentage of accuracy. Taking advantage of the convergence
solutions aiming across diagnostics, personalised medicine and
data science, some organisations on the map are
to address each already seeking to mine big data sets to enable
specified category identification of individuals at the earliest stage of
disease, when interventions have a higher likelihood
of success.
Overall, 66% of the initial solutions featured on
our map indicated they contained one or more
categories of diagnostics. As can be seen below,
many early solutions are in diagnostic imaging/
radiology (25%), where digital imaging has been
Leveraging
in widespread use for a number of years. This
skills and compares to far fewer solutions listed in pathology
capacity (9%) and endoscopy (3%), where the digital and AI
43% solutions are only recently starting to emerge.
Organisational
processes Other
50% 24%
Key areas
We wanted greater insight into what types of
problems are being addressed across the range of
solutions. Survey respondents were able to select
44%
Not applicable
where AI 25% 21%

multiple entries from a list of four categories.
Results can be seen below.
The percentage
Unlocking value in data/analytics was the top 11% Other
of companies
can deliver
category (75%) addressed by solutions submitted Imaging/Radiology Genetics
for the survey, followed by condition recognition
(60%). Organisational processes were addressed
& Genomics by category of
by half of the solutions, reflecting the increasing 9% diagnostics
impact
3%
Pathology
20%
use of AI to automate routine clinical, managerial Endoscopy
and back office tasks (e.g. document management,
paperwork and scheduling).
Physiological
measurement
The percentage of solutions, which

indicated a point of care, delivering
in each point of care site
The percentage Yes 57%

of projects
linked to smart
Patient’s own connected
home 49%
devices
Other 13%
No/Unsure
Hospital 68% Residential care
43%
or nursing home,
assisted living
38%
Mobile, semi-mobile
units and other and Personal/
Point of
Another way that AI is enabling (8%), vendor/supplier managed
‘pop up’ style settings wearable
new models of care is by settings (19%) and mobile,
of no fixed abode 19% technology using remote diagnostic and semi-mobile units (19%), were
35%
care
monitoring capabilities to selected by the least number of
change where and how care is respondents.
delivered. We asked solutions to
Already, 57% of solutions within
indicate the points of care where
the survey say they able to link
they deliver services (multiple
Vendor/supplier to smart connected devices (e.g.
selections were possible).
Internet of Things). With super-
managed clinic Excluding those entries that did fast 5G broadband networks
facilities 19% not indicate a point of care, the being tested this year, it is likely
Medical majority of solutions reported that the number of IoT-enabled
transport delivering services in hospitals solutions offered in non-acute
vehicle 8% (68%), followed by a patient’s points of care will increase over
Community or own home. Care settings such the coming years as 5G networks
primary care as medical transport vehicles are rolled out more widely.
clinic 39%
Regulation Proprietary (closed source) 35%
Our survey aimed to capture a device. Devices meeting the

Unsure 19%
The percentage of
full range of AI activity across requirements can place a CE Prefer not to say 16%
solutions according
the health and care ecosystem, mark (or logo) on their product to Not applicable 14%
ranging from ongoing research show that the medical device has Other (Please specify) 8%
to type of licence
projects to fully scaled met the requirements as set out
GNU GLPv3 (open source) 4% the computational
commercial products and in the conformity assessment.
services. With new solutions The CE marking also means Apache Licence 2.0 (open source) 3% product has
coming to market regularly, it is that the product can be freely MIT Licence (open source) 2%
important for buyers (including marketed anywhere in the EU.
commissioners and consumers) In the United States, the Food
and users of AI to have and Drug Administration (FDA)
mechanisms for distinguishing provides medical device approval.
which solutions have the
When we asked our respondents
appropriate evidence base and
about regulatory status, only 18%
Licensing
are ready for ‘at scale’ adoption. Currently, 35% of solutions in Proponents of open standards,
of solutions indicated they had
In the UK, medical devices secured approval in the UK/EU or the survey have been developed such as the Apperta Foundation,
must demonstrate that they abroad. A further 23% indicated using proprietary (closed a not-for-profit community
meet the requirements set they were in the process of source) software, distributed interest company supported by
out in the Medical Devices securing approval. under licensing agreement to NHS England and NHS Digital,
Directive by carrying out a named users who are given maintain that liberating both data
conformity assessment. The authorisation to modify, copy and applications and making
assessment route depends and republish applications. The them portable and interoperable
on the classification of the source code for this software is eliminates lock-in, facilitates
not shared publicly for anyone innovation and competition,
to look at or modify. Proprietary and forces vendors to compete
software developers often on quality, value and service.
pride themselves on product A downside can include the
‘usability’ and providing a high significant capacity and capability
level of ongoing support for required to run open platform
maintenance, security, content ecosystems.
updates and training.
The percentage of projects with regulatory approval A further 19% are unsure what
In contrast, only 9% of licence their computational
respondents report using one of product uses altogether, and this
the following three open source needs to be explored further to
licences – GNU GLPv3, Apache understand the reasons for this.
Licence 2.0 and MIT Licence.
Open platforms are vendor
41% 23% 18% 18% and technology neutral and
are based on open standards,
meaning that any application
built on an open platform will
operate on an open platform.
Not applicable In process of No Yes

securing approval
Who responded?
Survey respondents included Respondents cite a broad range
CEOs (42%), senior managers of experience with AI, with many
(15%) and others working across indicating that they wear multiple
the AI ecosystem in England. hats when dealing with AI.
They represented both large
organisations with 250 staff or
more (32%), as well as micro
organisations with less than 10
staff (38%) across private, public
and charitable sectors, as well as
academia.
REAL WORLD
ANALYSIS ON Q: What describes your current
experience with AI? I evaluate
FEASIBILITY AND
AI 43%
I use AI
I procure AI
IMPLEMENTATION
53%
17%
In order to inform government policy and the AHSN Network AI

Initiative offer, we conducted a survey of 106 thought leaders and AI
pioneers during May and June 2018. In this section we outline survey
results, highlighting top barriers and enablers for catalysing an ethical, I regulate
AI 5%
evidence-based market for AI solutions in health and care.
I develop AI
Other 51%
25%
Game- The best AI-enabled solutions

always solve a valuable problem
or ‘use case’, as expressed by
Development in drug discovery
and medical research will also be
hugely aided by AI.
‘80% of all
changing
users - citizens, carers and
Respondent views ranged from
professionals. Working with
AI being ‘ubiquitous’, ‘pervasive’
users to understand their needs
and ‘high impact’ that will ‘replace dermatology
use cases
and then working with them to
front line tasks’ to rather less
prototype and test solutions
iteratively is key to refining the
optimistic predictions. Many see
AI as a tool to help doctors and all
diagnoses will be done
using AI within 3 years
product’s value proposition and
healthcare professionals become
ensuring successful uptake and
more efficient and deliver a
adoption at scale.
We asked respondents to identify
higher standard of care at less
cost to benefit patients. Most see - it will be better than
the areas where the strongest AI having a key role in helping to
early use cases are. make decisions across the board
and in better planning for scarce
dermatologists at
Overall, the views were clear that
the main game-changing use
resources. diagnosing’.
cases for AI will be in three key
areas in the immediate period: Respondent prediction
• Diagnostics
• Non-clinical (operational and

administrative efficiency)
• Health promotion and

preventative health.
The top three use cases are • ‘Translation into routine processes (e.g. document
explored in more depth below. practice, widespread use management, paperwork and
Treatments of clinical decision support scheduling). Machine learning will
Health and Diagnostics Non-clinical Keeping Diagnostics (accurate and
tools for complex diagnostics, increasingly be used to process
promotion and interventions (accurate (e.g. save up to date early detection) was cited
genomics and lifestyle advice’ images and texts.
preventative (including and early time with with medical overwhelmingly as a strong
health surgery) detection) administration) research early AI use case, with 94% of A reduction in administrative
• ‘80% of all dermatology
respondents citing it as either diagnoses will be done using staff overheads is expected, and
extremely important or very a positive view on how AI will
Extremely important 47% 40% 80% 66% 35% important. Some predictions from
AI within 3 years - it will be
better than dermatologists at impact clinicians also emerged.
survey respondents include: diagnosing’. • ‘AI and clinicians will work more
Very important 31% 28% 14% 23% 30% • ‘Huge impact in radiology closely as one team’
Use cases for non-clinical
for assisted reporting and
applications, for instance saving
Quite important 16% 26% 5% 9% 21% screening’
time with administration, were
• ‘Supervised machine learning -
Clinicians remain in control’.
• ‘Increased use in radiology and seen as extremely important
Somewhat important 5% 5% 1% 2% 10% other imaging applications, or very important by 89% of
particularly in prioritisation/ respondents. AI will increasingly
Not at all important 1% 2% 0% 0% 4% triage of scans to ensure these be used in the automation of
routine clinical and managerial
are brought to human attention
first’ tasks and for back office
Some predictions include:

• ‘Significant improvements in
workflow management and
Comments from respondents
include:
• ‘Move towards using AI as
• ‘AI will be instrumental in
detecting minuscule changes
in individual’s records (data),
Overall AI To gauge what factors might support the
development of AI in health and care, we asked
respondents to consider the extent to which the
enablers
making it possible to detect following actions or policies were important in
data analysis coupled with the a tool for early prevention
and catch problems even realising the potential of AI in health and care.
emergence of intelligent clinical and diagnosis using large
before they actually form. It will
decision support systems’ population level datasets i.e. The numbers below highlight actions or policies that
enable prevention in the most
identifying individual risk. Use respondents viewed as very or extremely important:
• ‘AI will become a standard literate sense of the word’.
of AI in demand management
part of devices and image and predictive modelling’ For an excellent overview of
management systems’.
AI use cases, refer to Future
• ‘Better allocation of resources
Health promotion and Advocacy’s Ethical, Social and
by earlier detection of patterns
preventative health was cited
as extremely important or very
and thus disease, with
Political Challenges of Artificial
Intelligence in Health and Care
92% 88% 87% 82% 81%
better targeted preventative
important by 78% of respondents. (April 2018)3, a report produced
strategies as a result’
Overall, respondents expect AI to with the Wellcome Trust.
be used in a more predictive way, • ‘AI will take a large amount 3
Ibid.
facilitating the shift from reactive of the early identification of
care to a more preventative disease, allowing clinicians Engagement Ethical frame- Capacity and Clarity around Education of
health model in which people are to focus on the complicated of healthcare work to build/ capability to ownership of healthcare
more empowered to take care of cases’ professionals preserve trust and deliver scope data professionals
their own health. transparency
AI enablers
We then looked at the relative importance of factors
impacting diagnostics, building on potential actions
suggested by our AHSN Network AI Initiative core
specific to
advisory group members.
‘Better allocation The percentages below show the key factors that
pioneers believe are very or extremely important
of resources by
diagnostics
to address in order to realise the potential of AI in
health and care:
earlier detection of
patterns and thus
disease, with better 93% 93% 85% 82% 78%
targeted preventative
strategies as a result’
Respondent feedback Data sharing for Support the Consistent Reviewing Work with
medical imaging spread of proven labelling governance models commissioners
for AI training innovations methods for (eg ST 11-7) in light to help them
imaging data of machine learning understand how
algorithms to buy AI-enabled
products and
services
Trust, privacy and ethics

Education of
healthcare Education
professionals of public
According to survey respondents,

the top two factors enabling the
• ‘The speed at which AI will
have an impact on healthcare
Educating healthcare
professionals and the public on
Extremely important 50% 37%
realisation of AI in health and care will depend very much on the the potential of AI in a balanced
are ‘engagement of healthcare public’s (and therefore the way was also raised as a key Very important 31% 28%
professionals’ and establishing government’s) trust in AI and issue by survey respondents.
an ‘ethical framework to build/ the company using patient This is central to achieving Quite important 17% 25%
preserve trust and transparency’. data to develop AI. This will not and maintaining trust in an
impact all AI products but a environment where there is much
Overall, respondents agreed with
significant proportion’ negative media coverage on
Somewhat important 2% 10%
the need for a clear governance
the risks of AI and its potential
structure to guide decisions
and build trust. This needs
• ‘There has to be first an
enabling framework within the
impact on workforce. A narrative Not at all important 0% 0%
around data sharing is needed.
to be underpinned by a clear NHS. This would include ethical There is also the need to engage
ethical framework to address considerations, the right for the public actively in order to help
such issues as transparency in human interpretation of the AI define the problems that need
algorithm development. algorithms’ Key comments include: Predictions for the future include:
solving and co-develop solutions
Comments from the respondents enabled by AI. • ‘We need a narrative around • ‘Greater public support for AI
• ‘We need transparency of
include: data sharing and trust …’ due to better understanding
algorithm development’
• ‘There is a need for widespread [of] how AI works’
• ‘[the potential of AI] is based • ‘We need public education
understanding of augmented and enabling regulatory • ‘There will be a new cohort
intelligence, predictive on the governance structure
developed and ability to forge frameworks’ of healthcare professionals
analytics, deep learning and that will be educated to think
machine learning’ trust’. • ‘It’s not enough to ‘educate’ how to empower their human
the public- we need active abilities with AI driven tools’.
participation of patients and
other interested parties at all
stages of the development
Engagement Ethical framework to process’
of healthcare build/preserve trust • ‘Much more needs to be
professionals and transparency done to educate healthcare
professionals, listen to/
Extremely important 58% 61% understand their concerns,
and get their buy in. At the
Very important 33% 26%

moment the conversation is too
polarised between naysayers
‘We need a
who say, “AI will never change
narrative around
Quite important 8% 9% healthcare significantly” and
techno-utopians who say, “AI
will replace all doctors and data sharing
Somewhat important 0% 3% nurses” - the reality is of course
much more nuanced than that.’ and trust …’
Not at all important 1% 0%
Respondent feedback
Workforce Evidence of effectiveness

knowledge of AI and regulation
Workforce opportunities will be in the new era of AI will also be • ‘Ability for enough people to Linking strongly with the theme 78% of respondents felt that • ‘Regulation needs to be
addressed in detail in the Topol essential, along with training in understand the back end of of trust, the requirement for regulation was extremely light touch, to allow patient
Review (being led by Dr Eric technical and legal aspects of AI. AI, and even perhaps learn the evidence of effectiveness of important or very important confidentiality but at the same
Topol and facilitated by Health coding within hospitals to help the digital health innovations in realising the potential of time allowing the industry to
Above all, securing clinical
Education England), but it is understanding, engagement internal management’ and intelligent clinical decision AI in health and care. Gaps in flourish so we can achieve
important to note that at a high and buy-in to the co-design support (algorithms) was a topic regulating AI-enabled products efficiencies in the fast time
level, education has come out as • ‘We need to understand and
and use of AI will be important that ran throughout respondents’ and services, and uncertainty possible’
a clear enabler. design the human computer
to leverage the potential of the comments. about the roles of the various
interaction and how algorithms • ‘Regulation is important, but it
87% of respondents indicated technology. This will not only regulators and when a product,
are used in practice’ A number of respondents called would be better to find a global
that building capacity and assuage clinicians’ fears and service or algorithm should
for the ability to explain the solution rather than country by
capability is extremely important concerns, but will ensure that the • ‘Understanding financial and be subject to regulation were
algorithm and providing enough country. Particularly concerned
or very important to achieving AI algorithms developed augment clinical pathways in more detail also strong themes. There is a
information to allow regulators to if the UK decides to go its own
AI’s potential. This includes basic (rather than replace) and increase will be important…’ clear need for a new regulatory
independently replicate results way post-Brexit, as the NHS
education on AI and its potential the accuracy of human clinical framework to keep up with
• ‘Helping to build interdisciplinary on a similar set of data, ensuring market isn’t large enough to
applications for senior managers decision making. Helping senior advances in AI.
teams so that clinicians with algorithms are safe and unbiased. be worth separate certification
and directors in clinical, decision-makers to understand
good ideas can have these Representative comments Comments included: beyond FDA [United States] &
management, commercial and and have realistic expectations
of what AI has to offer will also be realised by people with include: CE [EU]’.
procurement roles. Training in • ‘Legacy regulations will
important. computer programming skills’ • ‘Many AI-based tools will limit widespread adoption
areas such as user-driven design,
change management, ethics and Key points to consider from our • ‘Having highly skilled data struggle to get used through of diagnostics and health
having difficult conversations survey participants include: scientists involved is crucial’. lack of evidence and/or clinical prevention applications’
conservative behaviour’
• ‘[Government should] address
• ‘Clinical support should be barrier of regulation and
gained by discussion of the the ability to rapidly iterate,
Somewhat Extremely scientific case and justifying prototype, and validate
important important the technology. Just offering prospectively’
46% an algorithm lacks scientific
0%
credibility’.
Regulation of AI
Not at all
important Capacity and Extremely important 44%
1% capability Very important 34%
to deliver
Quite scope Quite important 17%
important Very
12% important Somewhat important 5%
41%
Not at all important 0%
Funding and
Given AI’s potential for system
wide impact, with funding
flows and incentives crossing
‘Understand
commercial models
organisational boundaries
and hierarchies, some
respondents also commented
on the opportunity to reimagine ‘value’ of public/
commercial models in the new
Despite the financial challenges
experienced in the NHS, ‘funding
models for AI development and
deployment’, which came in at
92% of respondents said
they believed ‘supporting the
era of AI: NHS data and how
and budget restraints’ featured
only tenth on the list of thirteen
12th and 13th (last) on the list
respectively.
spread of proven innovations’
is extremely important or very
• ‘Understand ‘value’ of public/
NHS data and how this can be this can be sold
factors affecting the potential of important to realising the sold to developers or used to
AI in health and care, with 68%
of respondents indicating that
This result could be reflective
of the early stage of the
potential of AI in diagnostics. generate additional income’ to developers or
funding was extremely important
or very important. Featuring even
development of the AI market
in health and care. In contrast,
• ‘Evaluate cross department
business models - who used to generate
in diagnostics, where the early
additional income.’
owns hospital-wide clinical
lower on the overall list of key
AI use cases are strongest and efficiency? For example,
factors were ‘NHS internal market
where we are already starting to will radiology purchase an
and procurement’ and ‘lack of
see products come to market,
clarity over appropriate business
AI product whose benefit is Respondent feedback
realised by reduced drug cost
in neurology? How do those
dots get joined up?’.
Lack of
clarity over
appropriate
business
models for AI NHS internal
Funding/budget development market and
Extremely
restraints and deployment procurement
important
Not at all
50%
important 1%
Extremely important 45% 34% 29%
Very important 23% 25% 34% Support
Quite important 27% 33% 21% the spread
Somewhat
important 2% of proven
Somewhat important 4% 8% 13% innovations in
Not at all important 1% 0% 3% Quite diagnostics Very
important important
5% 42%
Data quality, sharing Consistent labelling Data sharing for
and interoperability methods for imaging

data
medical imaging for
AI training
Extremely important 54% 69%

The importance of a sound data Respondents’ comments include:
infrastructure with high quality
• ‘The current datasets in Very important 31% 24%
data and the relevant standards
healthcare are patchy, dirty
on sharing and interoperability
came through as key factors in
and often incorrect! Garbage Quite important 11% 6%
in garbage out. Often data are
realising AI’s potential.
not digitised. The first action
in machine learning or AI is to
Somewhat important 3% 1%
Data quality clean up dirty data’
A key concern affecting the ability • ‘We need clean labelled un-
Not at all important 1% 0%
of AI to deliver on its potential gamed datasets’
is that of the quality of the data
• ‘We need data compatibility
itself, much of which is not
through labelling and
digitised or in machine-readable
standardisation’.
format. ‘Data readiness’ (getting
data ready for AI) was a key Data sharing and interoperability
theme.
Overall, the view emerged that Key points voiced by respondents • ‘We need distribution across
the underlying data infrastructure were: public and private sectors,
is not fit for purpose for AI with patient access to any
Open standards and requires standards to
• ‘The role of private companies
information generated and
Clarity around to promote data and developers including
facilitate data sharing and the ease of sharing this’
Data sharing ownership of sharing and ownership of and access to
development of appropriate
public and patient data and • ‘Clarification of concepts
framework data interoperability commercial models to leverage
how data sharing agreements around patients being curators
the value of public/NHS data.
are negotiated’ not owners of the (“their”) data;
This is an especially pressing
Extremely important 56% 54% 47% concern where public sector • ‘The underlying IT
compliance with GDPR but still
entities have entered into allowing retention of images/
infrastructure in the NHS is
Very important 24% 28% 31% agreements with companies to poor and not AI ready. We need
blood results/other data to
feed Big Data’.
process data. These datasets a large push to standardise IT
Quite important 16% 10% 19% often end up in proprietary
format or in difficult to access
formats and data sharing’
repositories. Intellectual property • ‘Robust development,

Somewhat important 4% 7% 2% of algorithms developed using testing and validation of AI
these proprietary data sets often is key. Without appropriate
governance it will be a liability’
Not at all important 1% 0% 1% rests with the companies (outside
the public sector/NHS)4.
4
Naylor, A. and Jones, E. (2017). Unleashing the potential of health and care data. Future Care Capital. Available at:
https://futurecarecapital.org.uk/policy/healthcare-data/.
Towards a
A number of respondents endorsed the NHS Digital and NHS England are also laying the
establishment of open data ecosystems in order groundwork for open innovation with a number of
to leverage insights from multiple datasets, initiatives including:
sustainable
enabling the real power of AI to come into its own. • Apperta Foundation, which recently published
For example, Transport for London provides a ‘Defining an Open Data Platform’
common API to access 80% of the UK’s transport
data. Thousands of developers (including the • Code4Health, which provides a home for the
ecosystem original Citymapper app) build on top of this open

API platform. Similarly, Open Banking, recently
introduced in the UK, will see the UK’s nine biggest
increasing number of open source projects
providing software suitable for use in health and
care
banks release data in a secure, standardised form,
so that it can be shared more easily between • International exemplars in the area of open
authorised organisations so they can then use it innovation platforms in the health and care space
to create more products and services to benefit include REshape Centre Radboud (Netherlands)
citizens. The intention is to put citizens in control and Boston Children’s Hospital (United States)
of their own banking data, providing an easier way • Closer to home, University Hospitals Plymouth
for them to move, manage and make more of their NHS Trust and Great Ormond Street Hospital are
money. exemplars in building open data ecosystems and
In order to unlock open innovation around data- fostering open innovation.
driven health and social care services, any open
data ecosystems must provide mechanisms for
data to flow safely and securely across disparate
Case Study
health and care organisations, whilst ensuring
informed consent and transparency. Enabling
citizens to ‘donate’ their consumer data (e.g. from
banking, retail, transport, telecommunications,
utilities, etc) and data from sensors and IoT-enabled Great Ormond Street Hospital DRiVE Unit (Digital
devices could also support citizens to stay healthy
and in their own homes for longer5.
Research, Informatics and Virtual Environments)
Momentum is growing to establish open data GOSH’s DRiVE unit provides a good example in a secure environment in the cloud that
ecosystems across health and care. This should of the type of open data ecosystem and is compliant with ICO and GDPR guidance
accelerate over time as forthcoming Industrial infrastructure required for exploitation of AI’s regarding the use of data for research.
Strategy Grand Challenge investments in initiatives potential within health and care. The DRiVE Clinicians, researchers and industry partners
such as Digital Innovation Hubs (connecting regional unit provides both a concept and a physical looking to address specific problems can come
health and care data with biomedical data in secure space dedicated to accelerating research together in secure virtual ‘workspaces’ to run
environments) and the Healthy Ageing Challenge and evaluation of new AI-enabled technology analyses and APIs. The data therefore does
start to bear fruit. and data analysis, with the aim of developing not leave GOSH’s control and governance,
scalable solutions for child health. Working providing full transparency. The GOSH team
with partners including University College and regulators have full data provenance,
London (UCL)/ Alan Turing, major industry including details of IP addresses accessing the
partners and NHS Digital, early areas of focus code and whether changes have been made.
will include machine learning, assisted decision During the development phase, the GOSH
making and the use of medical chatbots. team are also working to generate synthetic
datasets for innovators to test their early
GOSH’s open data ecosystem captures
algorithms on, prior to validation on real data.
and integrates data from multiple sources
5
Woods, T.M. and Kihlstrom, E. (2018). Data and the Future of Health and Social Care. Report, Proceedings and Key Recommendations.
Round Table, 17th November 2017. FutureHealth Collective. Available at: https://www.colliderhealth.com/future-health-collective.
Where are we now?

Whilst AI solutions are increasing is currently reviewing technical 1. Build fairness and transparency
in their complexity, most now and clinical safety requirement in digital health innovations,
delivering impact are on the low and design standards before algorithms and clinical decision
complexity end of the spectrum. publishing onto the library. support tools.
Understanding the vast potential Ultimately, the NHS must
of AI – as well as its limitations 2. Help identify the requirements
protect its reputation as an and standards that
- will be key moving forward. internationally trusted health and
As one survey respondent organisations and suppliers
care system, ensuring patient need to fulfil in order to show
commented, ‘AI is still evolving… safety and high quality care, and
it won’t solve all the problems that products are safe, secure
preserving the trust between and maintain public trust.
healthcare faces as the moment’ citizens, clinicians and the wider
and we must avoid the trap of health and care system. In order 3. Identify gaps within regulatory
SUMMARY
‘overhyping potential, unrealistic to do this we have collated a set and approval processes that
claims, and poorly thought out of principles outlined in a Code of need to be addressed to
products’. Conduct, which is in early stages accommodate developing
Top AI enablers include of development. technologies.
AND NEXT
engagement with health The purpose of the Code is By working collaboratively with
professionals, as well as to provide a source of clear academia, industry, innovators,
grounding the use of AI in principles and guidance for the commissioners and AHSNs to
real problems as expressed development of trusted digital iterate and continually update
STEPS
by citizens, carers and other health innovations and intelligent
these principles, we can go
health professionals. Providing algorithms within the UK NHS
some way to staying abreast of
mechanisms for improving data health and care sector. This
evolving technologies, helping to
quality and the underlying data code can be used by innovators,
catalyse the scale and adoption
infrastructure will also be key, industry, commissioners,
of intelligent technologies.
along with introducing a safe, academia and individuals,
Addressing these requirements
evidenced and transparent as a framework to support
with the right solutions will spur
approach to how algorithms and development and deployment of
collaboration across the NHS,
The analysis of survey responses, together with the constellation of innovations are developed. any DHI or intelligent algorithm
social care and other partners in
(IA). Whilst this code will ensure
Currently in the NHS, we have the ecosystem and build public
organisations in the AI map and illustrated by the case studies in this a number of programmes such
that the DHI/IA being developed
trust. Strong cross government
are in line with the principles and
report, reveals that AI in health and care is still at a relatively early as the Local Health and Care
values of the UK health and care
collaboration, including pooling
Record Exemplars, NHS Test resources and partnering on joint
system, it is still a requirement
stage. At the same time, there are many promising early use cases Beds and the forthcoming Digital
that the relevant regulatory
initiatives is also underway and
Innovation Hubs that give us the is the key objective of the AHSN
for AI in this space, especially in diagnostics. The health and care opportunity to test and refine
and/or approval processes are
Network AI initiative.
adhered to.
digital health innovations (DHIs)
AI ecosystem continues to grow at pace, with a range of promising and algorithms with our partners. This code, if followed, can ensure
that within the NHS and the wider
interventions in the pipeline, currently gathering evidence that they The NHS Apps Library and
UK health and care sector we
Digital Assessment Questions
are safe, effective and offer value prior to regulatory approval and are examples of how the NHS collectively:
widespread implementation.
This is where For individuals

This report highlights where Artificial Intelligence is been clear which behaviour change Future developments in AI will
currently having an impact on the health and care processes were responsible for the also be game changing for
system. The survey results highlighted make it clear In the future AI will bring observed results7. epidemiological researchers
we are, but
that innovations falling under the umbrella term precision medicine into the who, since John Snow’s study
Thus, as a result of future AI
of AI, are being used to transform every aspect of realms of possibility for a far of the London cholera outbreak
developments, both quantitative
health and care delivery in the UK. It is incredibly greater range of conditions and in 1854, have been trying to
medical and qualitative socio-
exciting to witness this potential being unlocked, more nuanced interventions use mathematics to make the
where are
cultural health interventions will
but it is only the beginning. than it is now. As Neil Jacobstein, prediction of diseases less
become more accurate, targeted
The development of AI has a long history, with chair of the AI and Robotics chaotic and more deterministic
and effective.
the era of electronic computers beginning in the track at Silicon Valley think so that it can be more easily
we going? 1940s and the current enthusiasm for the use of tank Singularity University controlled. The ability to do this
AI dating from the publication of Norbert Wiener’s emphasises, truly personalised For clinicians, has always been hampered by
Cybernetics: or Control and Communication in the recommendations based on
consideration of an individual’s
care givers and the problems of modelling, data
and computation. However,
Animal and the Machine in the 1960s. It, therefore,
stands to reason that interest and investment in the entire genomic profile will change researchers as AI programs become
development of AI techniques will remain high for a the way we view medicine6. more computationally and
The delivery of health and
long time to come. technologically powerful and able
In addition, and perhaps more care is a fundamentally human
to make connections between
It is almost impossible to predict what health importantly, future AI tools will, endeavour. Even very far into
previously un-associated data
and care use cases these future developments enable a rapid shift towards the future it is impossible to
points Dr Gunjan Bhardwaj ,
will deliver because technology is cumulative. preventive medicine. As the imagine a world in which Artificial
chief executive of Innoplexus
Technology adoption follows the normal distribution McKinsey Global Institute report, Intelligence – be it robotic or
AG, attests that they will deliver
curve meaning that not everybody takes advantage Artificial Intelligence the next algorithmic – will replace the
the ability to generate and test
of the latest developments at the same time, digital frontier?, highlights, role of clinicians and carers
and the goal posts are always changing. As Garry novel hypotheses for almost all
medical professionals will completely. An artificial system is
Kasparov accurately points out in Deep Thinking it healthcare scenarios.
have enough information to highly unlikely to be able to, for
was once believed that if an AI chess game could provide genuinely personalised example, pick up on subtle shifts These new statistical modelling
play at grand master level, this would represent preventive recommendations to in things such as body language capabilities will grant us the
the unlocking of human level intelligence and yet individuals based on both medical which can give clinicians vital ability to forecast the spread
this was achieved in the early noughties. Now interventions and lifestyle and clues into how a patient is feeling. of disease, anticipate which
nobody thinks twice about playing chess on their environmental factors such as However, future developments patients will be most likely to
mobile phone because, as Nick Bostrom stresses nutrition and exercise. What’s will be able to give clinicians and succumb9 and deliver precision
in Superintelligence, once something works, it is more, constant monitoring and carers augmented intelligence, public health interventions with
no longer considered AI. This will be just as true associated feedback loops will enabling them to process a maximised effectiveness (Flahault
for health and care AI interventions as it is for the enable clinicians to get over the et al 2017).
greater amount of critically
AI developments of the past. Whilst it might seem hurdle articulated by Michie et relevant information in a shorter
cutting edge now to be able to use a symptom
al (2009) that whilst there have amount of time and giving
checker such as Ada, in the very near future this is
been many systematic reviews of them the time to focus on such
likely to just be routine.
behaviour change interventions higher-level cognitive tasks by
It is possible, however, to speculate about what that have demonstrated automating many mundane and
the results of these AI developments will be for all worthwhile benefits, it has not routine actions.
components of the health and care system, even if
it is not possible to say which exact developments
will deliver these results.
6
Ross, C. (2018). The Future of AI in Health Care. Psychology Today, 13 June 2018. Available at: https://www.psychologytoday.com/gb/
blog/the-future-brain/201806/the-future-ai-in-health-care
7
Michie, S., Fixsen, D., Grimshaw, J.M., Eccles, M.P., (2009). Specifying and reporting complex behavior change interventions: the need
for a scientific method, Implementation Science, 4:40. Available at: https://doi.org/10.1186/1748-5908-4-40
8
Forbes Technology Council (2018). How AI is transforming the future of healthcare. Available at: https://www.forbes.com/sites/
forbestechnologycouncil/2018/01/30/how-ai-is-transforming-the-future-of-healthcare/ #4faf820d3e60
9
McKinsey & Company (2017). Artificial Intelligence: the Next Digital Frontier? A discussion paper. Available at: http://bit.ly/2fRblkR.
Flahault, A., Geissbuhler, A., Guessous, I., Guérin, P. J., Bolon, I., Salathé, M., & Escher, G. (2017). Precision global health in the digital
10
age. Swiss Medical Weekly, 2017(1314), Article: w14423.

For the system

Finally, on the largest scale,
Making this vision
a reality The AHSN Network AI
Initiative: moving forward
future developments in AI will
To achieve this vision, it will be
deliver what Flahault et al 2017
paramount to ensure that only
describe as a ‘learning healthcare
the best and most evidence-
system’. A system where all
based AI innovations are allowed
component parts are equipped
to spread throughout the system, This first AHSN Network AI Through it all, AHSNs look By catalysing this AI ecosystem,
with the relevant data and
but that all innovations have an Initiative survey sets a baseline forward to continuing to play an the AHSN Network AI initiative
decision support devices that
equal chance of providing this against which progress can important role in shaping and seeks to accelerate the
analyse results in real-time using
evidence base. This is why the be monitored over time. The implementing the national digital development, validation and
machine learning techniques.
report not only focuses on the accompanying online interactive and AI agenda, working directly spread of AI in healthcare, which
These results are combined with
ways that AI is transforming map that visualises the AI with citizens and healthcare in turn has huge potential to
systems thinking methods to
health and care, but the steps hotspots will also help inform the professionals to define clearly improve care and cut costs within
deliver a healthcare system that
that are being taken to create development of an ecosystem the problems that could be the NHS.
constantly adjusts to ensure it is
an open, transparent, clearly to promote collaboration and addressed by AI in the future.
always following the principles By building a collaborative
regulated, and competitive knowledge exchange. The map
of the currently best known Using our regional connections, framework and ecosystem within
market for emerging health tech. will be regularly updated, and a
evidence11. convening power and expertise, which we encourage the adoption
Creation of this environment will number of activities rolled out to
AHSNs can also undertake of fair and transparent principles
As such, in the future, we will mean that AI, and other digital develop the ecosystem, sustain
commissioned work, for instance: we will help to build a trusted
be looked after by a health and health innovations, are given momentum around solutions, and
and safe market for intelligent
care system that is using the the opportunities necessary build awareness of the potential • expanding our innovation
technology in health and care.
most recent developments in to deliver the best possible of AI across health and care. exchange role to support
mathematics, statistics, electrical outcomes for all those receiving, uptake of a wider range of
Over time, the AHSN Network AI
engineering, neurophysiology, delivering and managing health evidence-based AI-enabled
Initiative will develop and evolve
information theory, physics and and care services. digital innovation working
as a mutually beneficial and self-
epidemiology12 to: (1) ensure with the NHS in our regions,
sustaining ecosystem of leaders
that primary, secondary and including with STPs and ACSs;
and innovators committed to
community health and care
using artificial intelligence and • supporting industry and
services are running as efficiently
other digital technologies to entrepreneurs and brokering
as possible, and; (2) that the
transform health and social care. links;
result of these hyper-efficient
and effective services is delivery • helping establish Digital
on the prime minister’s mission Innovation Hubs and other
to use data, artificial intelligence open innovation ecosystems;
and innovation to transform the
prevention, early diagnosis and • supporting implementation
treatment of chronic diseases to of the Topol Review and
improve the overall quality of life other government policy
of the nation’s citizens. recommendations; and
• helping partners including NHS

England, NHS Improvement,
the Office for Life Sciences and
NHS Digital accelerate uptake
and solve complex issues in the
digital and AI space.
Ibid.
11
12
Nag, N., Pandey, V., Hyungik, O., Jain, R.
(2017). Cybernetic Health. ArXiv. Available
at: https://arxiv.org/abs/1705.08514.
APPENDIX 1: Unlocking value

in data/analytics
CASE STUDIES myLevels
A recent start-up using AI
Brainwear – wearables in brain
tumour patients
Ada Health
Ada is an AI-powered
to provide personalised diet Developed by the Imperial personalised health platform.
recommendations based on College London Computational Ada helps the NHS in two ways
people’s unique sugar level and Academic Clinical Oncology 1) empower patients to self-care
responses. This stops sugar team, Brainwear provides for minor illnesses by providing
levels spiking and crashing – electronic activity monitors to access to high quality medical
helping people lose weight and adult patients with high-grade information within Ada’s app,
reducing their chance of getting gliomas undergoing chemo- thereby reducing patient demand
diabetes. radiotherapy, providing the for GP and A&E services and 2)
opportunity for continuous triage patients to ensure they get
Deontics
outpatient monitoring. The team the right care at the right time in
A clinical decision support and at Imperial is currently evaluating the right place, thereby releasing
patient-facing aid software results compared to conventional GP capacity and improving care
enabling quick collation of assessments of disease, and outcomes.
conventional knowledge to demonstrate value to help
including guideline, quality Brainwear become part of the
measures, clinical trial and standard repertoire of Neuro-
research papers into a standard Oncology trial outcomes.
executable format that can be
integrated with individual patient
data to provide evidence-based
treatment options.
Organisational processes
Space Finder Blackford Platform
Space Finder is an artificial Developed by Blackford Analysis
intelligence solution developed this platform is used to deploy
by Edge Health/ South Tees multiple vendor AI solutions in
Hospitals to help operating to medical imaging workflows. A
lists run more efficiently. Using single platform unlocks multiple
predictive analytics, Space Finder vendor solutions, with a single
creates optimal operating lists contract, support contact and
tailored to the surgeon, the deployment.
patient and the operation.
Leveraging skills Condition

and capacity recognition
Virtual Lucy resonance imaging (MRI) scans Medopad Astrodem Detection, characterisation Kheiron Medical
captured for radiotherapy and risk stratification of breast
Virtual Lucy enables patients Suite of mobile applications, A predictive model developed Technology focussed on helping
planning. Radiotherapy is one lesions
to attend fracture clinics within remote monitoring devices and by Brighton and Sussex Medical radiologists detect breast
of the main ways head and neck
the comfort of their own home, wearable technology, including School which will help general Working with Cancer Research cancer earlier using deep
cancers are treated; radiation is
linking them with an end-to-end Apple watches, to connect practitioners (GPs) identify UK Imperial Centre this DeepMind learning, to address two key
used to kill cancerous cells and
cloud-based service providing patients with health care patients at high risk of dementia. project uses machine learning for problems: 1) ease the workload
prevent their recurrence. Through
support from referral to rehab. professionals. It uses machine 96,000 anonymised GP patient the detection, characterisation on overstretched screening units
analysis of the images used in
Lucy’s model, tested on 20,000 learning to analyse patient data records from the Clinical Practice and risk stratification of breast by serving as a second reader
planning radiotherapy, the aim
patients thus far, has been to improve diagnosis, enhance Research Datalink are being used lesions. The project aims to to support breast screening
is to design a computer-based
shown to reduce outpatient treatment, predict deterioration to train the model. demonstrate the feasibility of programmes and 2) increase
tool to automatically segment
appointments by 50%, whilst also and empower patients and applying machine learning for the accuracy of screening by
regions and improve efficiency in Automated analysis of retinal
ensuring that 100% of patients carers. potential future clinical benefit in reducing the number of false
radiotherapy units. imaging using machine learning
have a consultant making mammography screening. positives and false negatives, and
Docobo (Doc@home free text techniques for computer vision
decisions on diagnosis and Alder Hey Children’s Hospital, hence number of patients getting
symptom checker) SkinVision
rehabilitation. Ask Oli Chatbot Working with Moorfields Eye unnecessary biopsies.
The doc@HOME Cloud EPR Hospital NHS Foundation Trust, An awareness and tracking
Automated segmentation of This is a hospital artificial
platform provides a free text DeepMind have developed AI solution that supports individuals
head and neck tumour to aid intelligence chatbot to answer
symptom checker enabling both technology which could improve with the early recognition of
radiotherapy planning common patient questions.
clinical users and patients the the way eye diseases are melanoma, squamous cell
The aim of this proposition is
A partnership between ability to key in symptoms and diagnosed and treated and could carcinoma, basal cell carcinoma
to reduce the number of low
DeepMind and University be taken down an automated ultimately help the millions of and precancerous actinic
complexity questions asked
College London Hospitals triage to help determine the chief people across the globe affected keratosis. It is the first global
to staff and to allow questions
NHS Foundation Trust (UCLH), clinical complaint. by preventable or curable sight certified skin cancer application
to be answered in a timely
applying machine learning to loss. This includes exploring available for consumers.
and engaging manner. The doc@HOME is currently deployed
automated segmentation of retinal pathology using deep
platform has also been designed in the Innovate UK / NHS England
head and neck tumour volumes learning tools on 3D volumetric
for children. Analysis of the Internet of Things test bed:
and organs at risk on medical optical coherence tomography
questions allows for real time Technology Improved Health
images including computed and digital imaging of the fundus
quality improvement. Management research project
tomography (CT) and magnetic as well as the development of
(TIHM for Dementia).
models to segment, diagnose and
recommend treatment decisions
from macular optical coherence
tomography scans.
APPENDIX 2:
AI in the UK: ready, Code4Health Ethical, social, and
willing and able? Code4Health is an initiative political challenges of
House of Lords Select supported by NHS England artificial intelligence in
and NHS Digital to enable
Committee on Artificial health and care
FURTHER READING
Intelligence the best use of digital tools
and technology to deliver Future Advocacy and Wellcome
Report produced by the House safe, high quality, efficient This report looks at the current
of Lords Select Committee on and compassionate care. and potential use cases of AI
Artificial Intelligence to consider Code4Health aims to educate in healthcare and explores the
the economic, ethical and social and inform all participants in ethical, social, and political
implications of advances in AI. the health and care community: challenges that this would raise.
The report concludes that the citizens, patients, carers, health,
UK is in a strong position to care and digital professionals
be among the world leaders in about the possibilities digital Industrial Strategy:
the development of artificial technologies provide and equip building a Britain fit for
intelligence during the twenty
first century. The Committee
them with the tools, knowledge the future
and skills to collaborate to
state that Artificial intelligence, develop and implement high Department of Business, Energy
handled carefully, could be a quality digital solutions. and industrial Strategy
great opportunity for the British
This white paper sets out a
economy.
Defining an Open long-term plan to boost the
productivity and earning power
Artificial Intelligence: Platform of people throughout the UK.
Real Public Apperta Foundation The report sets out four areas
where Britain can lead the global
Engagement The Apperta Foundation believes
technological revolution: artificial
that innovative technologies are
Royal Society for the intelligence and big data; clean
the key to achieving positive
encouragement of Arts, growth; the future of mobility;
transformation in health care.
Manufactures and Commerce and meeting the needs of an
In its report the Apperta
(RSA) ageing society.
Foundation argues that open
This report is about getting digital platforms based on an
citizens involved in the discussion open platform can lower barriers
around artificial intelligence. to entry, stimulate innovation
The RSA makes the case for and enable successful start-
engaging citizens in the ethics ups to rapidly get to scale. The
of AI and shares a snapshot of Apperta Foundation report is an
public attitudes towards AI and attempt to propose a definition
automated decision-making. for an open platform, based on
They argue that the citizen voice standards that have been proven
must be embedded in ethical AI. to work worldwide including HL7
FHIR, SNOMED-CT, IHE_XDS and
openEHR.
Life sciences: Map of technology and Policy Paper: Thinking on its own:
industrial strategy – A data in health and care AI Sector Deal AI in the NHS
report to government The King’s Fund Department for Business, Reform
from the life sciences This map brings together case
Energy & Industrial Strategy
This report illustrates the areas
and Department for Digital,
sector studies from across England, where Artificial Intelligence (AI)
Culture, Media & Sport
highlighting some of the places could help the NHS become
Office for Life Sciences
that are experimenting with and The AI Sector Deal is the first more efficient and deliver better
This report, written by Life implementing new technologies commitment from government outcomes for patients. It also
Sciences’ Champion Professor to achieve better health and industry to realise the highlights the main barriers
Sir John Bell, provides outcomes or more efficient care. potential of Artificial Intelligence to the implementation of this
recommendations to government and Big Data. It outlines a technology and suggests some
on the long-term success of the package of up to £0.95 billion potential solutions.
life sciences sector. It was written
Medical devices: the
of support for the sector, which
in collaboration with industry, regulations and how includes government, industry The Topol Review:
academia, charity, and research we enforce them and academic contributions up
organisations. to £603 million in newly allocated Preparing the
Medicines & Healthcare
products Regulatory Agency
funding, and up to £342 million healthcare workforce
London: The AI Growth (MHRA)
from within existing budgets, to deliver the digital
alongside £250 million for
Capital of Europe MHRA is the designated Connected and Autonomous
future, Interim Report
Mayor of London competent authority that Vehicles. June 2018 – a call for
CognitionX was commissioned
administers and enforces the evidence
law on medical devices in the
by the Mayor of London to map Health Education England
UK. This online guide provides an
AI innovation across London
overview of MHRA’s investigatory A call for evidence from the
and identify the capital’s unique
and enforcement powers, which Topol Review team. The review is
strengths as a global hub of
ensure the safety and quality of looking at what training NHS staff
Artificial Intelligence. This report
medical devices. would need to undertake to use
outlines their findings and will
Artificial Intelligence technologies
inform the actions that the Mayor
and robotics.
will take to support the future
growth of AI across different
industries to drive innovation,
productivity and growth.
APPENDIX 3: This glossary of selected items have been ordered in relation to
GLOSSARY
the Complexity Scale for AI (described on page 10) developed by
Jeremy Wyatt.
This glossary has been developed using Wikipedia except where specifically indicated and referenced. For further detail, a
comprehensive glossary of AI terms can be found on Wikipedia at https://en.wikipedia.org/wiki/Glossary_of_artificial_intelligence.
High complexity – Care companion robot for keywords within the input,
then pull a reply with the most
AI applications Robot used to aid in caring for matching keywords, or the most
someone or to provide them similar wording pattern, from a
Autonomous vehicle with companionship. A robot database.
is a machine—especially one
An autonomous car (also known programmable by a computer—
as a driverless car and a self- capable of carrying out a complex Surgical or pharmacy
driving car) is a vehicle that series of actions automatically. robot
can sense its environment and Robots can be guided by an
navigate without human input. Robot used to support surgical or
external control device or the
Autonomous cars combine a pharmacy tasks such as making
control may be embedded within.
variety of techniques to perceive the right shaped cavity in the
their surroundings, including femur for a hip replacement
radar, laser light, GPS, odometry,
Chatbot or selecting the right type and
and computer vision. Advanced number of pills to fil a medicine
A chatbot (also known as a
control systems interpret bottle in a pharmacy. A robot
talkbots, chatterbot, Bot,
sensory information to identify is a machine—especially one
IM bot, interactive agent, or
appropriate navigation paths, as programmable by a computer—
Artificial Conversational Entity)
well as obstacles and relevant capable of carrying out a complex
is a computer program which
signage. series of actions automatically.
conducts a conversation via
Robots can be guided by an
auditory or textual methods. Such
external control device or the
Machine translation programs are often designed
control may be embedded within.
to convincingly simulate how
tool a human would behave as a
Machine translation (MT) is conversational partner, thus Mammogram
a sub-field of computational passing the ‘Turing Test’ (a test interpretation system
linguistics that investigates the proposed by Alan Turing in 1950,
use of software to translate text of a machine’s ability to exhibit A system using AI that assists
or speech from one language to intelligent behaviour equivalent doctors in the interpretation
another. A deep learning-based to, or indistinguishable from, that of medical images, specifically
approach to MT, neural machine of a human). Some chatbots use breast screening mammograms.
translation has made rapid sophisticated natural language Most systems highlight
progress in recent years. processing systems to parse suspicious areas for human
the meaning of the user’s input, attention, though completely
but many simpler systems scan autonomous systems are being
developed.
ECG interpreter Middle complexity Text to speech module Low complexity Neural networks of a signal processing algorithm
/ filter is analysing physiological
Systems using AI that assist – AI modules or A text-to-speech (TTS) system – AI reasoning Artificial neural networks (ANNs) signals to detect a QRS complex
doctors by generating a prose converts a digitised text string or connectionist systems are
report describing the key findings components into an audio signal that sounds methods computing systems inspired by
in an ECG signal.
in an electrocardiogram. like spoken speech. A common the biological neural networks

Natural language example is the speech output Deep learning module that constitute animal brains. Generative adversarial
Diagnostic decision to SNOMED code from a satnav. Deep learning (also known as Such systems ‘learn’ to perform networks
tasks by considering examples,
support system processing module deep structured learning or Generative adversarial networks
Knowledge based or hierarchical learning) is part of a generally without being
(GANs) are a class of artificial
The use of health information Natural Language Processing programmed with any task-
broad family of machine learning
technology including algorithms (NLP) is a linguistic technique expert system module specific rules. For example,
intelligence algorithms used in
methods based on learning data unsupervised machine learning,
and software to analyse complex that enables a computer A knowledge-based system in image recognition, they
representations, as opposed implemented by a system of two
patient data to assist healthcare program to analyse and extract (KBS) is a computer program might learn to identify images
to task-specific algorithms. neural networks contesting with
professionals by suggesting a meaning from human language. that reasons and uses a that contain cats by analysing
Learning can be supervised, each other in a zero-sum game
differential diagnosis. Clinical NLP, using SNOMED knowledge base to solve complex example images that have been
semi-supervised or unsupervised. framework.
CT’s concepts, descriptions and problems. The key feature of a manually labelled as ‘cat’ or ‘no
Deep learning architectures
relationships, may be applied to knowledge-based system is that cat’ and using the results to
Speech driven repositories of clinical information
such as deep neural networks,
identify cats in other images. Time series analysis
the knowledge is represented deep belief networks and
radiology report tool to search, index, selectively explicitly using a knowledge recurrent neural networks have They do this without any prior A time series is a series of data
with SNOMED coded retrieve and analyse free text. model based on the domain been applied to fields including knowledge about cats (e.g., that points indexed (or listed or
These techniques can be used they have fur, tails, whiskers and
output to extract SNOMED CT encoded
ontology. This allows them to computer vision, speech graphed) in time order. Most
generate plausible explanations recognition, natural language cat-like faces), such as would commonly, a time series is a
Use of speech recognition data from free-text patient of their advice. Other algorithmic be necessary in a knowledge-
processing, audio recognition, sequence taken at successive
technology to create records and support the retrieval or statistical advice generators based system. Instead, they
social network filtering, machine equally spaced points in time.
radiology report with SNOMED of clinical knowledge documents. lack this feature. automatically generate
translation, bioinformatics and Time series analysis comprises
(Systematized Nomenclature identifying characteristics
drug design, where they have methods for analysing
Of Medicine) coded output
Image processing Signal processing & produced results comparable from the learning material that time series data to extract
for diagnoses etc. Speech they process using supervised
recognition is the inter- module classification module (and in some cases superior) to
learning.
meaningful statistics and other
human experts. characteristics of the data, often
disciplinary sub-field of In computer science, digital Signal processing concerns to predict future results.
computational linguistics that image processing is the use of
develops methodologies and
the analysis, synthesis, and
Ensemble methods Object segmentation
computer algorithms to perform modification of signals, which
technologies that enables the image processing on digital (e.g. Random Forest algorithm Graphical models
are broadly defined as functions
recognition and translation of images. As a subcategory or conveying ‘information about Models) This is an algorithm used for A graphical model or probabilistic
spoken language into text by field of digital signal processing, delineating objects in digital graphical model (PGM) or
the behaviour or attributes of
computers. It is also known as digital image processing has Random forest models are an
some phenomenon’ such as images. structured probabilistic model is
automatic speech recognition many advantages over analogue ensemble learning method
sound, images, and biological a probabilistic model for which a
(ASR), computer speech image processing. It allows a for classification, regression
graph expresses the conditional
recognition or speech to text much wider range of algorithms
measurements.
and other tasks, that operate Signal processing
dependence structure between
(STT). It incorporates knowledge to be applied to the input data by constructing a multitude algorithm / filter random variables. They are
and research in the linguistics, and can avoid problems such as Recommender module of decision trees at training
commonly used in probability
Signal processing concerns
computer science, and electrical the build-up of noise and signal A tool to generate time and outputting the class
the analysis, synthesis, and theory, statistics—particularly
engineering fields. distortion during processing. recommendations driven that is the mode of the classes
modification of signals, which Bayesian statistics—and machine
Since images are defined over by algorithms applied to (classification) or mean prediction
are broadly defined as functions learning.
two dimensions (perhaps more) input data. Examples are the (regression) of the individual
conveying “information about
digital image processing may recommendations about films trees. Random decision forests
the behaviour or attributes of
be modelled in the form of you might watch from Netflix help to correct for decision
some phenomenon” such as
multidimensional systems. or books you might like from trees’ habit of overfitting to their
sound, images, and biological
Amazon. training set.
measurements. A typical example
Decision trees, rule graphics. Cluster analysis itself Regression – linear, Inference engine for Text generator using from a ‘case base’ that somehow
is not one specific algorithm, but resemble the new case whose
induction the general task to be solved.
multiple, logistic rules or frames Directive Clause data is provided. This could allow,
Decision tree learning uses a An example of a clustering In statistical modelling, An inference engine is a Grammars (DCGs) for example, an oncologist to
decision tree (as a predictive algorithm is a Kohonen feature regression analysis is a set component of the system that Natural language generation locate several historical patients
model) to go from observations net which is a self-organising of statistical processes for applies logical rules to the (NLG) is the process of generating with a similar disease and stage
about an item (represented in map that is trained using estimating the relationships knowledge base to deduce new reasonably natural sounding text as their new patient, and then
the branches) to conclusions unsupervised learning to produce among variables. It includes many information. The first inference from data, such as a patient data use the response to different
about the item’s target value a low-dimensional (typically techniques for modelling and engines were components of base or data entry form. A simple treatment options of these
(represented in the leaves). It is two-dimensional), discretised analysing several variables, when expert systems. The typical example is systems that generate historic patients to help choose
one of the predictive modelling representation of the input space the focus is on the relationship expert system consisted of a form letters such as contracts in the therapy for their new patient.
approaches used in statistics, of the training samples. between a dependent variable knowledge base and an inference a lawyer’s office. These usually
data mining and machine and one or more independent engine. The knowledge base do not typically involve grammar
learning. Tree models where variables (or ‘predictors’). More stored facts about the domain.
Classification rules, but more complex NLG
the target variable can take a specifically, regression analysis The inference engine applies systems dynamically create
discrete set of values are called algorithm helps one understand how the logical rules to the knowledge text. As in other areas of natural
classification trees; in these tree In machine learning and typical value of the dependent base using data eg. about a language processing, this can be
structures, leaves represent class statistics, classification is the variable (or ‘criterion variable’) patient and deduced conclusions done using either explicit models
labels and branches represent problem of identifying to which changes when any one of or advice, eg the diagnosis for of language (e.g., grammars) and
conjunctions of features that lead of a set of categories (sub- the independent variables the patient. the domain, or using statistical
to those class labels. Decision populations) a new observation is varied, while the other models derived by analysing
trees where the target variable belongs, based on a training set independent variables are Argumentation, human-written texts.
can take continuous values of data containing observations held fixed. Regression analysis
(typically real numbers) are called (or instances) whose category is widely used for prediction temporal or spatial
regression trees. CART refers to membership is known. and forecasting, where it has reasoner Case-based reasoning
Classification And Regression Examples are the diagnosis substantial overlap with the
Spatial–temporal reasoning is an
algorithm
Trees. for a given patient, based on field of machine learning.
area of artificial intelligence which Case-based reasoning (CBR),
observed characteristics of Regression analysis is also used
draws from the fields of computer broadly construed, is the process
Clustering algorithm the patient (gender, blood to understand which among the
science, cognitive science, of solving new problems based
pressure, presence or absence independent variables are related
and cognitive psychology. The on the solutions of similar past
Cluster analysis or clustering of certain symptoms, etc.). to the dependent variable, and
theoretic goal—on the cognitive problems. It has been argued
is the task of grouping a set of Classification is an example to explore the forms of these
side—involves representing that case-based reasoning is
objects in such a way that objects of pattern recognition. In the relationships.
and reasoning spatial-temporal not only a powerful method for
in the same group (called a terminology of machine learning, knowledge in mind. The applied computer reasoning, but also a
cluster) are more similar (in some classification is considered an goal—on the computing side— pervasive behaviour in everyday
sense) to each other than to instance of supervised learning, involves developing high-level human problem solving; or, more
those in other groups (clusters). It i.e. learning where a training control systems of robots for radically, that all reasoning is
is a main task of exploratory data set of correctly identified navigating and understanding based on past cases personally
mining, and a common technique observations is available. The time and space. An example experienced. Case based
for statistical data analysis, used corresponding unsupervised is a Qualitative Simulation and decision support systems use a
in many fields, including machine procedure is known as clustering Modelling (QSIM). computerised similarity matching
learning, pattern recognition, and involves grouping data into algorithm to offer the user cases
image analysis, information new categories based on some
retrieval, bioinformatics, data measure of inherent similarity or
compression, and computer distance.
Acknowledgements
Authors: • David Thompson, Programme • Dan Gosling, Policy Adviser,
Manager, South West AHSN Arden and Greater East
• Melissa Ream, Adviser, Kent Midlands Commissioning
Surrey Sussex AHSN and AHSN • Emma Smyth, Director of Support Unit
Network AI Initiative Communications, Health
• David Cox, National Medical
Innovation Network
• Tina Woods, Founder & CEO, Director’s Clinical Fellow (AI
Collider Health; Adviser to • Guy Boersma, Managing and Robotics workstream
AHSN Network AI Initiative Director, Kent Surrey Sussex of the Topol Review), Health
AHSN, and AHSN AI Initiative Education England (HEE)
• Indra Joshi, Digital Health and Lead within the Digital and AI • Des Holden, Medical Director,
AI Clinical Lead, NHS England INN Kent Surrey Sussex Academic
• Laura Day, Analyst, KSS AHSN Health Science Network
• Mark Avery, Director of Health
Informatics, Eastern AHSN • Dominic Cushnan, AI
Lead (Social & Digital
With special thanks to: • Melissa Ream, Adviser, Kent Transformation), NHS Horizons
Amy Baldwin, Ben Middleton, Surrey Sussex AHSN and AHSN • Dominic King, Medical
Des Holden, Dominic Cushnan, Network AI Initiative Director and Head of Product,
Eleanor Whittle, Genni Ansell- DeepMind
• Mike Burrows, Director
Apicella, Guy Boersma, Harpreet • Eleanora Harwich, Director of
Coordinator, The AHSN
Sood, Jan Steel, Jeremy Wyatt, Research and Head of Digital &
Network
Jess Morely, Lee Roberts, Manny Tech Innovation, Reform
Penuel, Mary Buckingham, Matt • Tara Donnelly, Chief Executive, • Finn Catling, Machine
Fenech, Maxine Mackintosh, Health Innovation Network, Learning Researcher & Clinical
Neil Sebire, Oli Blackaby, Patrick and Digital and AI Lead for The Entrepreneur, NHS England
Mitchell, Rick Fell, Sally Chapman, AHSN Network • Graham Sumner, Client
Sam Jackson, Shankar Sridharan,
• Tony Davis, Commercial Executive, IBM (UK)
Simon Eccles, Sophie Taysom,
Tony Newman-Sanders and Director, West Midlands AHSN • Guy Boersma, Managing
Tracey Watson. Director, Kent Surrey Sussex
AHSN AI Initiative Core Academic Health Science
Advisory Group Network and AHSN Network AI
With additional Initiative Lead
• Adrian Price, Policy Lead
support from: (Innovation and Horizon • Hugh Harvey, Clinical Director,
AHSN Network Digital Scanning), Department of Kheiron Medical
and AI INN Health and Social Care (DHSC) • Iain O’Neil, Deputy Director
• Bruce Potter, Chairman, Blake (Digital Strategy), Department
• Alan Davies, Director of Digital,
Morgan LLP of Health and Social Care
Innovation Agency
(DHSC)
• Chris Robson, Director,
• Axel Heitmueller, Managing Akeso & Co; NHS Supply • Indra Joshi, Digital Health and
Director, Imperial College Chain: Diagnostic, Pathology AI Clinical Lead, NHS England
Health Partners & Therapy Technologies • Jay Burrell Chief Technology
and Services, NHS Business Officer, GE Healthcare
• Ben Bridgewater, Chief
Services Authority • Jeremy Wyatt, Director and
Executive, Health Innovation
Manchester • Claire Novorol, Chief Medical Professor of Digital Healthcare,
Officer, Ada Health Wessex Institute of Health
Research (WIHR), University of • Mihaela van der Schaar, Man • Siddhartha Chaturvedi, Senior
Southampton Professor of Quantitative Partner Development Manager
• John Ainsworth, Professor Finance, Oxford Man Institute, (Healthcare & Life Sciences),
(Health Informatics), University Engineering Sciences, Microsoft (UK)
of Manchester University of Oxford • Tahreema Martin, Clinical
• John Kellas, Co-Director, This • Mike Tremblay, Director, Cassis Lead - Digital Medicine, Health
Equals; Consultant (Innovation Limited; Research Fellow, Education England (HEE)
and community engagement), University of Kent • Tara Donnelly, Chief Executive,
Bristol Health Partners • Nasir Rajpoot, Professor in Health Innovation Network
• Kenji Takeda, Director (Azure Computational Pathology, • Tina Woods, Founder & CEO,
for Research), Microsoft (UK) Head of Laboratory (Tissue Collider Health; Adviser to
• Laura Douglas, Co-Founder & Image Analytics), University of AHSN Network AI Initiative
CEO, myLevels Warwick
• Tony Newman-Sanders,
• Maja Pantic, Professor of • Natalie Banner, Lead, National Clinical Director for
Affective & Behavioural Understanding Patient Data, Diagnostics, NHS England
Computing, Imperial College Wellcome Trust
• Zoe Greenwell, Category
London, Research Director, • Nicola Strickland, Consultant Tower Manager – NHS Supply
Samsung AI Research Centre Radiologist & President (Royal Chain: Diagnostic, Pathology
• Manoj Ramachandran, College of Radiologists), & Therapy Technologies
Consultant Orthopaedic Imperial College Healthcare and Services, NHS Business
Surgeon and Honorary Reader NHS Trust Services Authority
& Royal London Hospital, Barts • Nuno Godhino, Chief
Health NHS Trust, Co-founder, Technology Officer and Partner,
Viz.ai Empericus
• Mark Hinton, Chief Technology • Richard Barker, Chairman,
Officer, Image Analysis Group Health Innovation Network
• Maxine Mackintosh, PhD • Richard Vdojak, Principal
Student in data science and Scientist & Innovation Lead
dementia, UCL (Data Science/AI), Philips
• Melissa Ream, Adviser, Kent • Shah Islam, Clinical Research
Surrey Sussex AHSN and AHSN Fellow, Division of Brain
Network AI Initiative Sciences, Imperial College
London
About The AHSN Network

AI Initiative
Working to create Innovation through Connecting
better, safer and collaboration people through
more sustainable We support the market technology
healthcare development and spread of
The healthcare system must
Internet of Things, Machine
place less reliance on face-to-
The Academic Health Science Learning and AI technology by
face interaction to be able to
Network (AHSN) AI (Artificial strengthening collaboration
continue delivering the best care
Intelligence) Initiative, is working between citizens, industry,
for patients. The programme
to improve the quality of people’s academia and the health and
aims to develop and drive
lives and address urgent care sector.
innovation and adoption of AI
challenges faced by the health
The programme looks to work technology throughout the
and care system. We believe in
with everyone who wants to bring NHS and social care. It aims to
creating better, safer and more
the benefits of AI into health improve care at a lower cost to
sustainable healthcare through
and social care and incorporates healthcare commissioners and
innovation in technology.
user-driven design into solution providers and promote healthier
Our diverse perspectives development and deployment. lifestyles
and expertise span clinical, Over time, the programme will
We believe technology can be
technological, legal and ethical mature and evolve as a mutually
used to improve the quality
areas as well as practical on- beneficial and self-sustaining
of people’s lives and address
the-ground experience from ecosystem of leaders and
urgent challenges faced by the
innovators in the NHS. innovators committed to using
health and care system. This is
Kent Surrey Sussex AHSN, led by artificial intelligence and other
why we are working to create
Managing Director Guy Boersma, digital technologies to transform
an ecosystem which will enable
hosts the AI Initiative on behalf of health and social care.
citizens, innovators and health
all 15 AHSNs across England. We aim to help the UK become and social care staff to connect
the leader in what the Life and create sustainable health and
Sciences: industrial strategy social care solutions.
refers to as ‘entire new markets’,
Find out more at
drawing on the assets of our
www.ahsnnetwork.com
world class universities, our
market leading AI SMEs (c.50% of
global AI SMEs are UK based) and
our comprehensive NHS health
dataset.
@AHSNNetwork
The AHSN Network
www.ahsnnetwork.com
info@ahsnnetwork.com

AI Strategy

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

AI Strategy

Enviado por

Direitos autorais:

Formatos disponíveis

Accelerating

10: What do we mean by AI in health and care?

14: Results of the national survey about AI

22: Real world analysis: feasibility and

38: Summary and next steps

44: Appendix 1: Case studies

48: Appendix 2: Further reading

52: Appendix 3: Glossary

62: About The AHSN Network AI Initiative

Matt Hancock Lord O’Shaughnessy

Photo attribution: Chris McAndrew [CC BY 3.0 (https://

A summary of proposed next steps.

Example of how simple components form modules

High complexity Middle complexity Low complexity

HEALTH AND CARE

approval and/or publicly Regression 25%

where AI 25% 21%

The percentage of solutions, which

The percentage Yes 57%

Regulation Proprietary (closed source) 35%

Our survey aimed to capture a device. Devices meeting the

Not applicable In process of No Yes

In order to inform government policy and the AHSN Network AI

Game- The best AI-enabled solutions

• Non-clinical (operational and

• Health promotion and

Some predictions include:

Trust, privacy and ethics

According to survey respondents,

Very important 33% 26%

Workforce Evidence of effectiveness

Data quality, sharing Consistent labelling Data sharing for

and interoperability methods for imaging

Extremely important 54% 69%

repositories. Intellectual property • ‘Robust development,

ecosystem original Citymapper app) build on top of this open

Where are we now?

This is where For individuals

age. Swiss Medical Weekly, 2017(1314), Article: w14423.

For the system

• helping partners including NHS

APPENDIX 1: Unlocking value

Leveraging skills Condition

APPENDIX 3: This glossary of selected items have been ordered in relation to

in an electrocardiogram. like spoken speech. A common the biological neural networks

About The AHSN Network

Você também pode gostar