Transforming knowledge about human behaviour

The Human Behaviour Change Project team on a unique collaboration between behavioural and computer scientists.

Behaviour change is key to tackling the world’s health and environmental problems: obesity, cancer, sustainable living and countless others. Researchers, policy makers and practitioners are crying out for evidence: what works, when, where and how.

A mountain of research evidence about behaviour change is being produced, but more than humans can synthesise and access – and it is fragmented. The Human Behaviour Change Project (HBCP) aims to transform our capacity to make use of this evidence.

The HBCP’s vision is to build a largely automated system that will synthesise up-to-date research evidence in real time and use it to generate new insights about effective behaviour change interventions. The user will be able to elicit answers tailored to the behaviours, populations and settings specific to their interests. This interdisciplinary, collaborative project brings together cutting-edge developments in computer science, behavioural science and information science (including system architecture) and includes academic and industry researchers. The collaboration will enable us to interpret the literature more fully using machine learning and artificial intelligence.

We have an ambitious vision of success – when faced with a question about changing behaviour, instead of being overwhelmed by a large diverse literature, policy makers, practitioners and researchers will be able to get answers (based on real-time updated evidence) to questions such as:

  • Are methods that are effective in changing the dietary behaviours of young, educated populations living in Europe effective for populations that are older or less educated or living in Africa or Latin America?  If so, how effective?
  • Is the answer the same for different behaviours (e.g. dietary/smoking/hygiene/recycling/ water conservation/clinical)?
  • Does it matter how these interventions are delivered? Face-to face, via TV or smartphones,to community groups, etc?
  • What are the magnitudes of effects?
  • What about populations that are hard to reach and difficult to engage?

Why? And why now?
How did we arrive at this approach? Bitter experience! Our Principal Investigator, Susan Michie, consulted for the Department of Health in England, and was asked whether the strategies shown to be effective to help people manage long-term conditions would also be effective for the ‘general population’ to prevent ill health. The question was timely as there is lots of evidence –  searching Google Scholar for ‘behaviour change’ found over 50,000 results for the first three months of 2018. But, frustratingly, while there was lots of research evidence out there, the tools to assemble it efficiently and effectively just weren’t available.  

Later experience with the Public Health Interventions Advisory Committee of NICE (National Institute for Health and Care Excellence) – which made recommendations for the UK’s National Health Service and public health system – revealed that the time taken from a government minister posing a question to its being answered was often very lengthy. Up to two years might be spent specifying a researchable question, commissioning systematic reviews, developing guidance, field-testing and making recommendations. At the end of the process, the evidence we could draw on was often sparse or not relevant for the intended UK populations and settings and the minister had moved on anyway.

We need to be able to find and understand what the relevant evidence has to say more quickly, mining the vast amount of accumulating evidence from public health and the behavioural and social sciences. Recent developments in behavioural and computer science mean that this is now possible.

What we are doing is beyond the capacity of human effort; we need computer power to help disentangle the vast mass of research. So we will ‘share’ our understanding of behaviour change interventions with computers, which will entail translating the diverse syntax, definitions, terminologies and languages used by researchers working in the field, and developing shareable taxonomies that are machine-usable.

This involves a system for organising our knowledge – a behaviour-change intervention ontology. It’s a knowledge system that uses a controlled vocabulary to define the features of interventions – such as techniques, modes of delivery, target behaviour, context – and the relationships between those features. Using that vocabulary to annotate the entities in published intervention evaluations, we can train computers to do the job more efficiently and at scale.

Computers can learn to recognise and extract such information – we are already having some success in training computers to recognise behaviour change techniques such as ‘self-monitoring’. But even apparently simple questions like ‘How old were the participants?’ requires the computer to recognise words such as ‘age’ and ‘years’ and then associate the appropriate number with these words. When the computers can accumulate the basic information, they can then begin to predict results – effect sizes for particular combinations of behaviours, interventions, populations and settings. And excitingly, while humans organise this evidence in different ways in different disciplines, computers lack these biases and may detect relationships and associations that we have not noticed.

Challenges – working across disciplines
Developing the behaviour-change ontology raises technical, methodological and philosophical issues, and working with an interdisciplinary team to address and solve them has been central to our project.

What’s it like to work together? Our general advice on interdisciplinary working is:

  • Choose to work with people who are better at their discipline than you are at yours – they will push you to learn and to succeed.
  • Work with people you like or at least get along with.
  • Agree some practices (e.g. about frequency of meetings, and terminology) at the start.

But the first impression is how similar we are! It’s been interesting to see how much of the recent progress in artificial intelligence draws on the everyday language of psychology. Of course, this can lead to confusion if we are careless. For example, to a computer scientist a ‘learning model’ typically refers to how computers are trained to mimic aspects of human intelligence – a key step in artificial intelligence. In psychology, ‘learning models’ mean something completely different. So the ontology has proven to be an invaluable lingua franca for communication across the team. It is satisfying to see how the ontology facilitates active collaboration between all of the disciplines in the project. It allows us to share knowledge clearly and consistently, so that individual disciplines can then contribute useful results to the project.   

Our collaborations with major producers and users of evidence synthesis – in particular, NICE and The Cochrane Collaboration – are important. They have a keen interest in making the discovery and synthesis of evidence more efficient and are also working on classification schemas and investigating the potential of automation. We can learn from one another’s work and avoid reinventing wheels. Ensuring compatibility, in terms of data structures and domain models, will be vital to ensure that semantic data can be shared between organisations into the future.

Shared tools and terminologies
Our aim is to make knowledge easy to share. This is trickier than it sounds because in the behavioural and social sciences words for the same things are used differently in different studies. For example, if we attempt to integrate evidence from studies of ‘behavioural counselling’, we might be combining evidence from studies that involve ‘educating patients’ or, by contrast, ‘feedback, self-monitoring and reinforcement’ while a systematic review of ‘self-monitoring’ might miss studies of ‘daily diaries’ (Michie et al., 2008). Precise labelling and definitions make it easier to combine like with like.

A key objective of the project is to enable replication through shared methods. The physical and biomedical sciences have made outstanding progress when they have adopted shared tools for structuring evidence, as in classification systems like the periodic table, or for collecting data, for example the use of shared neuro-imaging methods. There is some evidence that individuals trained in shared behaviour change techniques are more successful in recognising the content of such interventions, and might therefore be more able to replicate the content (Johnston et al., 2018).

Indeed, shared tools and terminology are essential even to know whether an idea or method is new rather than having been simply relabelled: ‘…we cannot have a paradigm shift if we do not have a paradigm. A paradigm requires some shared methods or theories’ (Johnston, 2016). Moreover, shared explicit systems can advance thinking by pointing to gaps and indicating what is needed to fill them. As we expected, the behaviour change technique taxonomy has pointed to the lack of strong evidence for a large number of behaviour change techniques in current use.

Next steps
This project will advance the field in three ways. First, it will make it possible for researchers and decision makers to identify relevant evidence quickly, without sacrificing the completeness of the research identified. This will make it possible to conduct systematic reviews of the domains covered by the HBCP with an efficiency that has hitherto been impossible. Second, it will provide a new model for evidence discovery and synthesis that can be a template that other areas can follow. The technical and software architecture of the platform we are building will be open source and free for others to use and build upon. Third, it will advance the field of synthesis methodology itself, by providing new approaches for synthesising research that are based on machine learning rather than the more traditional meta-analysis that the field is used to. These methods will blend the precision and strength for causal claims that synthesising findings from RCTs can bring, with newer analytic methods used for ‘big data’.

This is just the beginning of what will be a long journey. Once we have established that this kind of work can be done and adds value, it can be extended to cover a wider range of research designs and domains. We are setting up an extensible system that others will be able to contribute to and take forward; this is essential, as the scale of problem is such that no single research team will be able to solve it on its own.

We are starting with smoking cessation to help us prove the technologies and evaluate the overall impact of the approach. We plan to move on to physical activity, alcohol consumption and dietary consumption, incorporating literature from a wider range of research designs and analysing data at the individual rather than study level. This could have significant potential impact for practitioners seeking personalised next-best action recommendations for their patients.

Extending the artificial intelligence capabilities of the system is also on our list of interesting directions. What if we could simulate possible outcomes of a behaviour change experiment using existing data? Artificial intelligence has transformed activities such as drug discovery, where computers use research data about molecules to streamline new-drug research by proposing candidate compounds with desired clinical effects. We are all excited by the prospect that a future version of our system to simulate the effects of policy-driven behavioural interventions might lead to better value for money, better ways of enabling people to change their behaviour and better health outcomes.


Organising evidence about interventions effectiveness: Cochrane’s PICO project

Developing a typology of features of the proximal environment that influence behaviour: TIPPME project

Evidence of the targeted behaviours and how they change; for example, how smoking changes over time and conditions:

Advancing interventions using theoretically based lab and real-life tests of hypotheses and evidence from the collaboration of behavioural and cognitive science, to gain better understanding of the processes of behaviour change, especially non-conscious processes: ‘Behaviour change by design’ programme

Examining how behaviour change intervention content may be personalised in line with the more general aim of ‘precision medicine’: The Science of Behavior Change (SOBC) research network in the USA

As for the evaluation of behaviour change interventions, HBCP focuses on RCTs and is likely to gain from work by:

Photo – The Human Behaviour Change Project team, including: 

Principal Investigator: Professor Susan Michie, Director of the Centre for Behaviour Change and of the Health Psychology Research Group at UCL
James Thomas: Professor of Social Research and Policy at the EPPI-Centre, UCL
John Shawe-Taylor: Professor of Computer Science at UCL
Pol Mac Aonghusa: Senior Research Manager at the IBM Research Lab in Dublin
Marie Johnston: Emeritus Professor of Health Psychology at the University of Aberdeen and a registered Health and Clinical Psychologist.
Mike Kelly: Professor and Senior Visiting Fellow in the Department of Public Health and Primary Care at the Institute of Public Health and a member of St John’s College at the University of Cambridge
Robert West: Professor of Health Psychology and Director of Tobacco Studies at the Cancer Research UK Health Behaviour Research Centre, UCL

Key sources
Johnston, M. (2016). A science for all reasons: A comment on Ogden (2016). Health Psychology Review, 10(3), 256–259.
Johnston, M., Johnston, D., Wood, C.E. et al. (2018). Communication of behaviour change interventions: Can they be recognised from written descriptions? Psychology & Health, 33(6), 713–723.
Michie, S., Johnston, M., Francis, J. et al. (2008). From theory to intervention: Mapping theoretically derived behavioural determinants to behaviour change techniques. Applied Psychology: An International Review, 57(4), 660–680.
Michie, S., Richardson, M., Johnston, M. et al. (2013). The behavior change technique taxonomy (v1) of 93 hierarchically clustered techniques: Building an international consensus for the reporting of behavior change interventions. Annals of Behavioral Medicine, 46(1), 81–95.
Michie, S., Thomas, J., Johnston, M., et al. (2017). The Human Behaviour-Change Project: Harnessing the power of artificial intelligence and machine learning for evidence synthesis and interpretation. Implementation Science, 12(1), 121.

BPS Members can discuss this article

Already a member? Or Create an account

Not a member? Find out about becoming a member or subscriber