Transcript ppt - UCL

The Human Behaviour-Change Project
Participating
organisations
A Collaborative
Award funded
by the
www.humanbehaviourchange.org
@HBCProject
This evening
Opening remarks from the chair
1. The need and vision
2. The preliminary ontology
3. The sciences working together
Question and answer session
4. Automated data extraction from papers
5. Machine learning
6. The querying system for users
7. The future
Question and answer session and discussion
Mary de Silva, The Wellcome Trust
Susan Michie, UCL
Robert West, UCL
Marie Johnston, Aberdeen University
Pól Mac Aonghusa, IBM Research
John Shawe-Taylor, UCL
James Thomas, UCL
Mike Kelly, Cambridge University
The Vision
Susan Michie
Professor of Health Psychology
Centre for Behaviour Change, UCL
Building the science of behaviour change
• The HBCP aims to revolutionise the ways in which we
• Build knowledge and understanding about behaviour change
• Use that knowledge to answer real-world questions
‘What works, compared with what,
how well, with what exposure, with
what behaviours, for whom, in
what settings and why?’
Building the science of behaviour change
• The HBCP aims to revolutionise the ways in which we
• Build knowledge and understanding about behaviour change
• Use that knowledge to answer real-world questions
• The method
• Harness cutting edge methodologies, to build
1. An Ontology of Behaviour Change Interventions
2. An Artificial Intelligence System, including Natural Language
Processing and Machine Learning
3. A User Interface including a query system
• Collaboration between behavioural, computer and information science
The problem
Volume of research
• >200 evaluations of behavioural interventions published each day
Reporting variability
• Studies are reported very variably so difficult to synthesise or to draw
theoretical conclusions about mediation and moderation of effects
Context sensitivity
• Much literature not directly relevant to specific contexts of knowledge users
Need for timeliness
• Typical time for study results to be included in systematic reviews 2.5-6.5 years
The need
Collaboration with
computer science
• … to efficiently advance
our understanding of
behaviour
• … to answer questions
from policy makers &
practitioners
The 4-year plan
1. A Behaviour Change Intervention Ontology for organising relevant
information from research reports
2. A Natural Language Processing system to find and extract that
information starting with the ‘use case’ of smoking cessation
3. Machine Learning and Reasoning Systems that integrate and
extrapolate from that information to generate new knowledge and
hypotheses about behaviour change
4. A User Interface that answers questions about behaviour change
interventions and explains its conclusions
The Preliminary Ontology of
Behaviour Change Interventions
Robert West
Professor of Health Psychology
Tobacco and Alcohol Research Group, UCL
The ‘big question’ in behaviour change
What works,
compared with what,
how well,
with what degree of exposure,
for whom,
in what settings
with what behaviours,
and why?
Unpacking the ‘big question’
Intervention and comparator
What is the behaviour change
intervention and how is it delivered?
Compared with what?
Exposure
What is the reach of, and
engagement with, the intervention?
Population
Whose behaviour does the
intervention aim to change?
Setting
What is the setting in which the
intervention is operating?
Mechanism
How does the intervention work?
Behaviour
What behaviour or behaviours is the
intervention targeting?
Unpacking the ‘big question’
Intervention and comparator
What is the behaviour change
intervention and how is it delivered?
Compared with what?
Exposure
What is the reach of, and
engagement with, the intervention?
Population
Whose behaviour does the
intervention aim to change?
Setting
What is the setting in which the
intervention is operating?
Mechanism
How does the intervention work?
Behaviour
What behaviour or behaviours is the
intervention targeting?
Unpacking the ‘big question’
Intervention and comparator
What is the behaviour change
intervention and how is it delivered?
Compared with what?
Exposure
What is the reach of, and
engagement with, the intervention?
Population
Whose behaviour does the
intervention aim to change?
Setting
What is the setting in which the
intervention is operating?
Mechanism
How does the intervention work?
Behaviour
What behaviour or behaviours is the
intervention targeting?
Unpacking the ‘big question’
Intervention and comparator
What is the behaviour change
intervention and how is it delivered?
Compared with what?
Exposure
What is the reach of, and
engagement with, the intervention?
Population
Whose behaviour does the
intervention aim to change?
Setting
What is the setting in which the
intervention is operating?
Mechanism
How does the intervention work?
Behaviour
What behaviour or behaviours is the
intervention targeting?
Unpacking the ‘big question’
Intervention and comparator
What is the behaviour change
intervention and how is it delivered?
Compared with what?
Exposure
What is the reach of, and
engagement with, the intervention?
Population
Whose behaviour does the
intervention aim to change?
Setting
What is the setting in which the
intervention is operating?
Mechanism
How does the intervention work?
Behaviour
What behaviour or behaviours is the
intervention targeting?
Unpacking the ‘big question’
Intervention and comparator
What is the behaviour change
intervention and how is it delivered?
Compared with what?
Exposure
What is the reach of, and
engagement with, the intervention?
Population
Whose behaviour does the
intervention aim to change?
Setting
What is the setting in which the
intervention is operating?
Mechanism
How does the intervention work?
Behaviour
What behaviour or behaviours is the
intervention targeting?
The BCI Ontology v1
West & Michie (2016) A Guide to the
Development and Evaluation of Digital
Behaviour Change Interventions in
Healthcare. London: Silverback
Cochrane’s PICO Ontology
We will map the BCI ontology where possible to Cochrane’s PICO
Ontology of clinical studies1
• Patient, Population or Problem
What are the characteristics of the population and the condition of interest?
• Intervention
What is the intervention under consideration for this patient or population?
• Comparison
What is the alternative to the intervention?
• Outcome
What are the relevant outcomes?
1http://linkeddata.cochrane.org/pico-ontology
The Sciences Working Together
Marie Johnston
Professor of Health Psychology
University of Aberdeen
Iterative interaction between the sciences
Ontology of
behaviour change
interventions
How can we
organise the
evidence?
Extracting and
interpreting the
evidence
What does the
evidence show?
computer science
information science
behavioural science
Making the evidence
accessible at scale in
real time
How can we make
the evidence usable?
Iterative interaction between the sciences
Ontology of
behaviour change
interventions
How can we
organise the
evidence?
Extracting and
interpreting the
evidence
What does the
evidence show?
computer science
information science
behavioural science
Making the evidence
accessible at scale in
real time
How can we make
the evidence usable?
Iterative interaction between the sciences
Ontology of
behaviour change
interventions
How can we
organise the
evidence?
Extracting and
interpreting the
evidence
What does the
evidence show?
computer science
information science
behavioural science
Making the evidence
accessible at scale in
real time
How can we make
the evidence usable?
Iterative interaction between the sciences
Ontology of
behaviour change
interventions
How can we
organise the
evidence?
Extracting and
interpreting the
evidence
What does the
evidence show?
computer science
information science
behavioural science
Making the evidence
accessible at scale in
real time
How can we make
the evidence usable?
The collaboration of 3 sciences
Iterative interaction between the sciences
computer science
information science
behavioural science
The Human Behaviour-Change Project
Questions and discussion
Knowledge Extraction from Reports
Pól Mac Aonghusa
Senior Manager of Social, Mobile and Decision Science
Research
IBM Research Dublin, Ireland
The task
• Can we teach a computer to be our Behaviour Change research
assistant?
• Interpret documents many times faster than a human - large, longterm memory
• Systematic identification of connections over everything it learns
• Unbiased assessment of evidence vs opinion – and consensus vs
disagreement
• Generate reusable knowledge that both humans and machine can
interpret
…. to answer real-world questions
Challenges
• Join knowledge from many sources – significant effort to identify
connections
• De-noise relevant knowledge – useful information represents
small proportion of total content
• Resolve content ambiguity – versus precise semantics of ontology
• Assign confidence to learned knowledge – assess evidence versus
opinion
• Connect rich semantic knowledge source to ML & AI – without
combinatorial meltdown
Starting points
• ‘Medical Recap’
• ‘Watson Oncology Advisor’
• Incorporating human faculties such as ‘Debating’
The Role of Machine Learning
John Shawe-Taylor
Professor of Computer Science
Centre for Computational Statistics and Machine Learning,
UCL
Artificial Intelligence (AI)
• Can computers be programmed to show human levels of intelligence?
• AI has been a dream of Computer Scientists since the birth of
automated computation, e.g., Ada Lovelace, Alan Turing
• First attempts at creating AI were focused on reproducing logical
reasoning in automatic programs
• Despite some successes (e.g., deep blue Chess playing system)
reducing intelligence to logic alone leads to a combinatorial explosion
of possibilities that defeats even the fastest machines
• General purpose AI seemed as remote as ever
Machine Learning (ML)
• ML develops algorithms to find patterns in data: based on
probabilistic analysis rather than logical inference
• Simplest ML tasks are supervised learning: data such as images
labelled with content (eg contains bicycle)
• Task is to feed this data to an algorithm that outputs a function to
classify new images (ie image contains/doesn’t contain a bicycle)
• There have been significant advances in solving these types of
problems: Support Vector Machines (SVMs), boosting and deep
learning are able to give accuracies similar to humans
ML 4 AI
• Turning an AI problem into a logical task can throw the baby out with
the bathwater:
• Richer representations have been shown to retain semantic information,
e.g., in natural language processing
• Furthermore the additional information contains patterns that machine
learning can for example use to learn to shortcut the combinatorial
explosion
• The approaches of supervised learning can be used directly or
combined to enable an agent to learn to act in a context
• Remarkable successes such as the IBM Watson (playing jeopardy),
DeepMind (playing Atari gams and Go), etc.
• HBCP will leverage these latest approaches to learn to populate
ontologies and suggest improvements to the structures
The user interface: aims and evaluation
James Thomas
Professor of Social Research & Policy
EPPI-Centre, UCL
The user interface
• Outputs of machine / human effort will be made available in an
online portal with two main aims:
• To enable widespread use of the knowledge generated
• To feed back into the AI system – and the ontology – a wide perspective of
views
• Two evaluations will be conducted:
• Which parts of a systematic review can be automated using the new system
(and how well)?
• Can the system transform the nature of evidence synthesis in terms of the
types of evidence utilised and the inferences developed?
The Future: Science and its Application
Mike Kelly
Senior Visiting Fellow
Institute of Public Health, University of Cambridge
Science
• Putting behaviour change on a properly evidence-based platform
• Policy problems and behaviour and behaviour change
• The current gap between knowledge and action
• The possibility of superseding common sense!
• Making scientific sense of mega information
• A new revolution in the evidence base
Applications
• Policy makers in public health and beyond – providing a platform for
evidence based policy interventions
• National public health bodies – Public Health England, Agency for
Health Care Research and Quality (AHRQ) USA
• The “What Works” Collaborations.
• National Institute of Health and Care Excellence (NICE)
• Academic colleagues
• Cochrane/Campbell Collaborations
• Information Scientists
The Human Behaviour-Change Project
Questions and discussion
www.humanbehaviourchange.org
@HBCProject