Last Wednesday (Jan. 26 2022), The Times newspaper published a supplement titled ‘Learning for the future: Interim Report’ by The Times Education Commission.
A case study titled ‘How artificial intelligence can tailor learning’, describes ‘a technological revolution’, focusing on the work of British educational technology company Century Tech whose founder Priya Lakhani, (November 2021), ‘the AI can predict with 96 per cent accuracy whether a child is autistic by analysing how they use the computer mouse or tap on the iPad screen’.
If true, this would make Century Tech one of the most important edtech companies to come out of the UK. I’ve spent over 20 years in the business of education and edtech as a founder, investor and advisor and I highlighted the article on Linkedin and wrote to Century Tech asking for information about the research. To date I’ve had no response but my Linkedin post has been viewed over 3000 times and generated more replies than anything I have ever posted, an indication that the wider edu community are as interested as I am.
I wanted to learn more about autism, or more correctly Autistic Spectrum Disorder (ASD), and any relevant research on how AI might be used to predict it.
There are many approaches to a diagnosis of ASD, from the NHS ; the National Autistic Society; the American Psychiatric Association Diagnostic and Statistical Manual of Mental Disorders (DSM-5) ; and another from the World Health Organisation’s International Statistical Classification of Diseases and Related Health Problems (ICD).
In broad strokes these all seem to indicate that:
- ASD is a spectrum disorder, i.e. there is a wide variation in the type and severity of symptoms people experience
- Diagnosis is complex and includes deficits in:
– Social-emotional reciprocity
– Non-verbal communicative behaviors
– In developing, maintaining, and understanding relationships
– Stereotyped or repetitive motor movements, use of objects, or speech
– Insistence on sameness, inflexible adherence to routines, or ritualized patterns or verbal nonverbal behavior
– Highly restricted, fixated interests that are abnormal in intensity or focus
– Hyper- or hypo-reactivity to sensory input or unusual interests in sensory aspects of the environment
- Symptoms may be present in young children, but not obvious (or diagnosable) until later in life
- ASD and intellectual disability frequently co-occur
So diagnosing ASD is medically complex and when you claim that a proprietary AI technology product can predict ASD with 96% accuracy, verifiable evidence is required. With as many as 1 child in every 44 having ASD, up to 200,000 of the 8.8m students from England may be on the ASD spectrum.
The UK National Autism Society confirms that some tech tools may help identify ASD, with published research in this area going back at least 20 years. Technologies and approaches include augmented reality tools and eye/gaze tracking apps, but these are simply an interim step to obtaining a detailed medical diagnosis.
The most relevant publicly available research around potential uses of AI to help support early predictors of ASD was funded by the US National Institute of Health and published in 2021 in JAMA Pediatrics. This was a study that used a gaze-tracking app to highlight bioindicators of ASD in babies up to 36 months old.
However, despite positive initial results, the study’s co-lead author noted, “We’re still at the beginning stages of this work”, with a larger 5-year follow-up study underway. This is a critical point, because none of the several previous ASD technology detection tools have had their initial positive results replicated. I have not found any peer-reviewed research papers from major institutions published in well-known journals like JAMA that have used how a student taps an iPad screen as their model to predict ASD.
Other recent AI ASD research looks at how AI techniques can be applied to the analysis of non-invasive radionomic techniques (medical imaging):
- ‘Can autism be diagnosed with artificial intelligence? A narrative review.’ Diagnostics, Richard, Chaddad, A., Li, J., Lu, Q., Li, Y., Okuwobi, I.P., Tanougast, C., Desrosiers, C. and Niazi, T.
- ‘Deep learning for neuroimaging-based diagnosis and rehabilitation of autism spectrum disorder: A review’. Computers in Biology and Medicine, Khodatars, M., Shoeibi, A., Sadeghi, D., Ghaasemi, N., Jafari, M., Moridian, P., Khadem, A., Alizadehsani, R., Zare, A., Kong, Y. and Khosravi, A.
There is also an FDA-approved app from Cognoa called Canvas Dx™ ( which uses AI techniques to assess video of a child (18 months to 5 years) along with a detailed questionnaire for parents and carers. The output is then reviewed by certified ASD specialists.
If Century Tech’s 96% predictive accuracy claim can be substantiated to help parents and schools to identify the need for a formal ASD diagnosis to the standards of the NHS, WHO, National Autistic Society, etc, they would be transformed from a high profile but unprofitable small edtech business to one of the most valuable UK edtech/medtech powerhouses, with a commensurate valuation.
The Times Education Commissioners include a self-declared ‘Who’s who’ of important voices in UK education. Under the headline, ‘What the evidence teaches us’, the report states that, ‘the evidence presented to The Times Education Commission suggests that the system as a whole is failing’. The organising idea here is that it requires evidence to make improvements to education.
The idea of evidence-led policy is hardly new, but it is when we look at edtech and the subset called AIED (AI Education)*. For many years the bulk of edtech research was organised and paid for by the companies making the products, effectively it was self-serving marketing spin rather than serious evidence to support the educational impact of the product let alone the econometric consideration of Opportunity Cost (the next best alternative). A recent shift in thinking about educational research has been spurred by the success of researchED and industry initiatives like Edtech Evidence Group, UCL Educate, the Edtech Demonstrator Programme and other initiatives locally and internationally.
The report’s strong evidence focus virtually mandates that to support their claim Century Tech should make their evidence available for scrutiny. Research needs to be replicated to be valid. This is not an easy decision – many companies view any potentially proprietary IP as their enterprise’s core commercial asset (the ‘gold dust’ in investor parlance). However, a greater argument can be made that unless a company can prove their edtech/AIED product or service, at a minimum, works as claimed, then schools should not spend taxpayers’ money on it.
The idea that most forms of IP protection are a key value driver is debatable given the high cost of obtaining patents and the cost of litigation against infringements. Instead, early-stage companies should invest in continuous innovation and iteration. One part of this complex puzzle would be for governments to mandate that all edtech/AIED used by government schools have an Open Data API that makes relevant, non-commercially sensitive, information available to scrutiny by independent third-party researchers. In addition, AI companies of all sorts, but particularly those in education, should open up their ‘black box’ products by committing to Explainable Machine Learning so buyers, educators and others can understand how complex machine learning algorithms make decisions.
With great power comes great responsibility as AI edtech pioneer Jose Ferreira, founder of Knewton, discovered after making claims like, ‘We know everything about what you know and how you learn best because we get so much data. And education is the highest-stakes media product in your life’. Knewton’s AI did not revolutionize education and it was sold for less than $17m to Wiley in 2019 (vs an estimated value of $300m+ in 2018).
I really hope Century Tech can substantiate their claim to have developed a highly accurate AI tool to help identify students potentially with ASD so they can then start the path to a possible formal medical diagnosis (or not).
Note: Minor updates made on 08/02/2022.
* the term first appeared in The International Journal of Artificial Intelligence in Education back in 1989. Its’ difference from edtech is well explained in Artificial Intelligence in Education, Wayne Holmes, Mya Bialik, Charles Fadel Center for Curriculum Redesign 2019.