Understanding evidence: machine learning and artificial intelligence in the Human Behaviour-Change Project

Presentation video:




Oral session: Understanding and using evidence (3)


Wednesday 23 October 2019 - 14:00 to 15:30


All authors in correct order:

Michie S1, Johnston M2, Thomas J1, Mac Aonghusa P3, West R1, Kelly M4, Shawe-Taylor J1
1 University College London, UK
2 Aberdeen University, UK
3 IBM Research Ireland, Republic of Ireland
4 University of Cambridge, UK
Presenting author and contact person

Presenting author:

Marie Johnston

Contact person:

Abstract text
Background: the questions that most Cochrane Reviews address investigate the evidence of interventions that either directly target behaviour change (e.g. smoking cessation, increasing physical activity) or indirectly (e.g. reducing blood pressure, improving quality of healthcare). In consulting evidence, policymakers and others require rapid and robust answers to variants of ‘What works, compared to what, how well, for whom, in what settings, for how long, for what behaviours and why?’ Relevant evidence is produced at an accelerating rate but is fragmented and growing more rapidly than humans can synthesise and access.

1) to create an automated system that makes better use of the vast amount of accumulating evidence from behaviour change intervention evaluations and promotes the uptake of that evidence into a wide range of contexts;
2) to harness and integrate the power of machine learning and artificial intelligence (AI) with the contribution of system architecture and the expertise of behavioural science.

Methods: the Human Behaviour-Change Project (HBCP; www.humanbehaviourchange.org), a collaboration of behavioural scientists, computer scientists and system architects, is building an AI system to scan the world literature on behaviour change, identify key information and convert this into knowledge that can answer complex questions. To train the AI system, behavioural scientists are building an organising structure (the Behaviour Change Intervention Ontology; BCIO). Using both machine learning and rule-based algorithms, the AI system will extract, synthesize and interpret relevant information.

The HBCP is:
1) developing the BCIO, working with colleagues in Cochrane and building on Cochrane’s PICO ontology;
2) developing and training algorithms to ‘read’ the annotated intervention evaluation reports; and will
3) build a system for automatic evidence identification and the tools for information classification and extraction;
4) create user and machine interfaces for interrogating and updating the knowledge base; and
5) involve patients and members of the public in evaluating the usability of the knowledge system.

Results: as part of the over-arching BCIO, we have developed four of the nine component ontologies: the Behaviour Change Technique Taxonomy and ontologies of Mode of Delivery, Intervention Setting and Intervention Population.

Conclusions: the project is on course for developing an automated system to 1) answer questions with up-to-date evidence tailored to user need and context and 2) generate new hypotheses and advance our understanding of human behaviour. The first behavioural domain will be smoking, followed by physical activity.

Patient and public involvement: patients and members of the public will be involved in evaluating the usability of the knowledge system when further development work has been completed.