AI Summit Report: How can artificial intelligence support early intervention?

On 14 and 15 October 2019, we held an Autistica summit at The Alan Turing Institute in London to develop research ideas about autism, early intervention and artificial intelligence. This is a report of the research ideas created at the workshop.

Autistic adults, parents of autistic children, researchers, health professionals and technology experts participated in the summit.

Why is this type of research needed?

Receiving interventions in the first years of life is often believed to be key to autistic people living a long, happy, healthy life.

Despite this view being widely held, there are a limited number of evidence-based early interventions for autistic children. As well as this, early interventions are very resource-intensive; for instance, they are provided by highly-skilled professionals whose time is scarce, who receive years of training and so are very expensive. As a result, evidence-based early interventions are not widely available, and these interventions are not equally accessible to everyone in society.

Rapid advancements in technology mean that it is becoming increasingly possible for professionals to help more people and make access to support more equitable with the use of algorithms and software that can speed up some of the most time-consuming elements of their work.

What are data science and artificial intelligence?

Billions of gigabytes of data are generated globally every day. Data science is the drive to turn this data into useful information and understand its powerful impact on science, society, the economy and our way of life. The study of data science brings together researchers in computer science, mathematics, statistics, machine learning, engineering, economics, philosophy, digital humanities, and other social sciences.

There is no single accepted definition of artificial intelligence or ‘AI’. Still, the term is often used to describe when a machine or system performs tasks that would ordinarily require brainpower to accomplish, such as making sense of spoken language, learning behaviours or solving problems. There is a wide range of such systems, but broadly speaking, they consist of computers running algorithms – sets of instructions in computer language – often drawing on data. Some branches of AI are described below:

Machine learning is a branch of artificial intelligence that allows computer systems to improve their performance by looking at examples, data and their own previous experience and interpolating or inferring information about the real world from examples it has been given.
Mathematical optimisation is when an algorithm can search for a huge range of options to calculate the best way to execute a plan. Optimisation is the process of selecting the best path through all the possibilities.
Robotics is an application of AI that combines machine learning to scan the room and understand the surroundings, inference to generate the next step in a sequence to progress towards a goal such as picking up an object, and a type of optimisation called pathfinding to quickly and efficiently coordinate its limbs.

Research ideas from the AI summit

Jump to: 1. Personalised profile building 2. Pain 3. Early detection of distress 4. Language and communication 5. Crowdsourcing intervention information 6. Improving diagnosis 7. Enabling environments 8. Sleep interventions

Why is this helpful?

Families often must wait a long time to receive the outcome of a diagnostic evaluation
Diagnosing developmental conditions like autism is complex and so diagnoses may be imprecise, incomplete or inaccurate
A tool that makes it easy to compile information from lots of sources may improve a clinician’s ability to make a diagnostic decision quickly, more accurately or to make more detailed provisions for the child in question

How would the research be designed?

Conduct a scoping review of the research literature and consult family and professionals to determine sources of information that have the potential to be clinically useful if they were input into a digital profile that diagnosing clinicians could access
Run a pilot study to determine if (a) any wearable sensors that automatically add data to a digital profile are acceptable to autistic children, (b) parents and professionals manually input information with the accuracy and frequency needed and (c) if either or both generate any clinically-useful information
Develop a minimally viable product that involves a dashboard where parents and professionals can upload and access the compiled data and view visualisations of their profile
Investigate if this product achieves any of the following desired outcomes: (a) reduction in wait time for diagnosis, (b) improvement in parent and children's experience of the process, (c) improvement in the professional experiences of the process or (d) increase in the accuracy or precision of diagnoses given

Who would be involved?

Autism researchers
Data scientists
Technology and health industry collaborators
Experts in wearable sensors
Children undergoing autism assessment or those who have been diagnosed and their parents
Representatives of each of the medical and educational professionals who may contribute data to a platform like this (e.g., GP, paediatrician, clinical psychologist, speech and language therapist, psychiatrist and educational psychologist)
Experts in data privacy who are experienced with health technology
NHS innovation and NHS digital collaboration

Things to consider

The need to create clear guidance on data ownership and data sharing permissions
Important to resource initiatives that build trust in the platform
Aggregated data from the platform would be a useful source of information for researchers, if individuals consent for their data to be used in that way
Designers should be engaged throughout to ensure the amount of effort/time needed to use the platform does not outweigh potential benefits that could come from using it

Why is this helpful?

It is often difficult to know when autistic people who speak few or no words are experiencing pain
Autistic people, including those who speak fluently, can sometimes find it difficult to identify when they themselves are experiencing pain
Improved recognition of pain may help autistic people and their caregivers to understand the source of an autistic person’s pain which allows them to make the necessary changes to their environment
Better identification of pain would help autistic people, caregivers and professionals to differentiate between behavioural responses to physical pain and distress
Improved pain recognition in autistic people would allow better referrals, diagnoses and interventions for co-occurring conditions that cause pain

How would the research be designed?

Conduct an online survey and interviews/focus groups to determine people’s perceptions, worries and questions about wearing physiological sensors in daily life in order to inform design of study and information materials
Run a pilot study to test which wearable sensors provide the best combination of acceptability, durability, reliability and signal detection
Conduct a larger study to identify the optimum combination of physiological, environmental and behaviour signal features that might be informative about pain
Scope out how signals can be used to give personalised early warnings for autistic people or to inform caregivers
Algorithms could be trained on those who can communicate verbally about their pain experiences and applied to those who can't

Who would be involved?

Pain specialists
Medical doctors
Autistic people
Education professionals
Data science and AI experts
Signal processing experts
Data interface designers
Autism researchers
An industry collaboration with a wearable sensor or data dashboard company
Physiotherapists
Occupational therapists

Things to consider

There are currently no validated objective physiological markers of pain that can be recommended for clinical use – any system would have to be predictive and may therefore be difficult to implement clinically
A composite measure of physiological, behavioural, and possibly neuroimaging measures could be developed
Extensive preliminary work to learn the acceptability of wearing a sensor among autistic people, detailed protocols about how to maximise the acceptability and reassuring people about sharing such detailed data would be needed before beginning work on this

Why is this helpful?

When an autistic child has a meltdown it can be difficult to determine what triggered it and so it is challenging to learn how to support autistic people to avoid or cope with the environments or situations that are distressing
Autistic people may be able to learn their triggers and how to self-regulate their emotions with the help of technology that can give them advance notice of a likely stress response
Quality of life may improve for autistic children and their families
Families could feel more empowered to access their community
Autistic behaviours could become more normalised and accepted in society if people feel more confident accessing their community

How would the research be designed?

Run a consultation between designers, engineers, autistic children and their caregivers to determine the type of information that would need to be collected and analysed with an emphasis on the feasibility of data collection methods
Determine the quantity and nature of data needed to test the feasibility of this idea
Conduct a feasibility study to collect data before and during acute distress and test if the data collected can be used to predict or demonstrate distress with reliability and validity in autistic children
Run a pilot trial and then a full trial to determine if the resultant system reduces the number of acute distress episodes experienced in a given time period and/or reduces overall levels of distress for autistic children and, if so, test which groups would benefit

Who would be involved?

Autistic children aged 3-8 years old and their caregivers
Autistic research consultants
Autism researchers
Education, health and social care professionals
Digital solution designer
Implementation specialist
Machine learning expert and/or data scientist
Film and video annotation expert
Statistician

Things to consider

There are no single objective measures that can be linked to meltdowns or distress and so multiple types of data would need to be combined to determine a probability of distress
Triggers and reactions are likely highly idiosyncratic for different individuals and so the system would need to be able to learn from an individual’s reactions and not be programmed based on normative data

Why is this helpful?

If successful, this could lead to better access to education via a web-based text simplification plug-in
This could improve educational attainment, employability and self-efficacy among autistic people who struggle to reach language proficiency or to identify non-literal language
The combination of natural language processing for text simplification and external linking for interpretation of non-literal language, deployed as a web plug-in would be novel

How would the research be designed?

Natural language processing could be employed to simplify and summarise text and identify and label idiomatic expressions
This would make it possible to add links to external sources that explain the history and meaning of the idiom, sarcasm, irony, metaphors used or of grammatical exceptions in text (e.g. Wiktionary)
Identify domains of application in collaboration with autistic young people
Find example datasets to use for training of simplification and summarisation models Identify sources of idiom definitions to link to
Develop prototype and evaluate it with end-users

Who would be involved?

Autism researchers
Linguistics or psychology of language researchers
Web developers
Autistic people, family members and caregivers to take part in the research
Autistic people and parents to consult on the research
Data scientists/AI experts (including those with natural language processing expertise)

Things to consider

This would need to be kept continually up to date as the use of metaphor and idiom are constantly evolving
Interoperability with other platforms, apps and devices would be key to its success

Why is this helpful?

There is a huge body of knowledge on how well early interventions in autism work sequestered away in the minds and stories of parents, schoolteachers and social care and healthcare professionals
This project could highlight that knowledge and fill gaps in ongoing research by understanding the combinations of interventions that work for the individual person
This platform could link up and support parents and caregivers by equitably sharing knowledge
A process could be developed to only share information that is supported by empirical evidence, or at a minimum, could label information to indicate the quality of evidence that supports it

How would the research be designed?

Design a transparent and open source platform to collect people’s accounts of their experiences of early interventions in autism
There will be a broad engagement with multiple key stakeholders to ensure the information is reliable and useful
The platform will use natural language processing to increase visibility of high-quality information and to link up people with similar experiences to share knowledge
Qualitative analysis frameworks will identify current gaps in the policy focus on autism

Who would be involved?

Translators to collect information in different languages
An advisory board with representation of policy makers, clinical professionals, education professionals, people lived experiences (autistic teenagers, parents)
Web developer
Researchers with expertise in qualitative and quantitative methods, autism, AI and citizen science

Things to consider

A critical challenge will be to develop a process whereby unevidenced and/or dangerous information is not shared via the platform or is appropriately flagged
Consideration needs to be given to ensure a platform like this serves the needs of different communities such as those from different cultures, regions or clinical profile

Why is this helpful?

Autism is diagnosed on the basis of observations of behaviour and accounts of developmental history - this requires a lot of time from highly trained professionals. As a result, there are long waiting times for publicly-funded diagnosis services and private services are expensive
Better prioritisation and triage of children referred for diagnosis could improve diagnosis waiting times by determining which children require specialist teams to make a diagnostic decision and where a decision might be more straightforward
If clinicians have access to a wider variety of and richer information about a child that may give a more accurate and precise diagnosis and prognosis to families
Could improve parents’ perceptions of the experience of undergoing diagnostic assessment

How would the research be designed?

Scope what similar work has been done already
Run a participatory workshop and wider consultation about ethics and data management
Agree ‘in principle’ access to data sources, including NHS, education and social care
Create platform for collection and analysis of different data sources (e.g., questionnaires, videos, health, education and social care data)
Use existing databases to pilot a model list and test if a likelihood of diagnosis score can be calculated from available data sources
Conduct simulation work to examine sampling biases
Compare the likelihood score to expert judgement based on same data or full assessment

Who would be involved?

Children up to 7 years referred for autism assessment
Health economist
AI expert
Developer for the app and/or platform
Parent of an autistic child
NHS managers
Professionals
Designers
Commissioners
Specialist autism diagnosticians (paediatrician, clinical/education psychologists, psychiatrists and speech therapists)

Things to consider

It is important that to justify the implementation of a system like that it must not increase a burden on professional and family member’s time and should produce outcomes that are better than currently achieved
Safeguards would need to be put in place to stop under-resourced services using the likelihood score instead of a full clinical assessment in diagnostic decision making
This could make use of under-used previously collected data

Why is this useful?

Autistic people experience sensory sensitivities which mean aspects of their sensory environment can cause distress. Control over the sensory environment can be enabling for autistic people
The link between physical space, AI and autistic children has been relatively unexplored
A system in which aspects of the sensory environment are updated in response to people's reactions could increase access to physical environments
Understanding how to design enabling environments for autistic people has applications in home environments, educational, health and commercial space
Open learner models have not been applied in this way before

How would the research be designed?

Combining open user models with architectural space
Identification of sufficient and necessary data for allowing their interpretation of a spectrum of experiences in situated environments (positive and negative)
Co-designing of model and sharing with children
Visualisation of information related to the agents interacting within the environment
Outcome 1: prototype of an enabling environment
Outcome 2: multimodal environment able to accommodate all types of communication profiles
Outcome 3: multimodal data mapped to children’s forms of expression

Who would be involved?

AI expert
Architect
Special education experts
Adviser and community liaison
A clinical psychologist
Data visualisation specialist
Data scientist, machine learning or analytics specialist
NHS network professional

Things to consider

Privacy considerations
Iterative development is needed – so this would require enough time built into planning
Rather creating a static profile of sensory needs this would need to focus on evolving and contextual sensory needs
It would be worth exploring if changes made to design principles based on autistic insights also result in benefits for other groups

Why is this useful?

Approximately 60% of autistic people have chronic insomnia
There are life-long adverse outcomes associated with poor sleep
No objective measures tailored for autistic people to support naturalistic or trial studies of sleep interventions
Objective monitoring and ecological measures allow personalised tailoring to enhance sleep
An automated measure would provide rich data with relatively low researcher time needed

How would the research be designed?

A series of co-production workshops with multiple stakeholders on remote monitoring technology for home settings
Pilot testing of co-produced recommendations in children with identified difficulties with sleep, recruited via NHS neuro-developmental services

Who would be involved?

Autism researchers
Industrial designers
Systems engineers
Industry partners
Participatory autism research collective (PARC)
AI and informatics partners
Human-computer interaction specialists
Process evaluation expert
Autistic and family member research consultants

Things to consider

It is understandable that people would have privacy concerns about collecting such detailed data about sleep but it will be important to consider the trade-off between privacy concerns and the negative consequences of chronic insomnia

What happens next?

This workshop is only the beginning of the process. The summit was successful at identifying research priorities that are shared between the autism community, the data science and the autism research communities but the next step is to evolve these questions into fundable research proposals that will lead to longer, happier and healthier lives for autistic people.

How can you get involved?

Follow us on social media for the latest research news and developments on this project or Join our Autistica Network for updates.

If you think you can help in making this research happen, email Lorcan Kenny at lorcan.kenny@autistica.org.uk.

This summit was generously supported by The Paul Foundation.