Labelling learning difficulties
Interview with
Now the school term’s begun here in the UK, and all over the country kids are going back or starting school. Some children have the extra challenges of learning difficulties to contend with, and might have been diagnosed with conditions including attention deficit hyperactivity disorder or ADHD, dyslexia, or an autism spectrum disorder. But how helpful are these categories for kids who are struggling at school? Using machine learning, Cambridge scientists have studied hundreds of children with these diagnoses, and identified clusters of learning difficulties which did not map to the diagnoses they had received. Chris Smith spoke to cognitive neuroscientist Duncan Astle, who led the study. Firstly Chris asked Duncan why the team did this study...
Duncan - Well, many children struggle with learning. There’s disorders like ADHD which was mentioned, so prevalence rates for those kinds of those things vary from 3 to 8 percent. But, actually, the proportion of children who struggle to learn is much higher than that. Governments statistics are around 30 percent of kids don’t meet their age expected levels by the end of primary school.
We’re really interested in what are the underlying causes or the underlying roots to being a struggling learner. And the way that you would normally study that would be to choose a group of kids that you’re interested in, say kids with ADHD and you would compare them to all the kids who don’t. But we began to think that there was some real problems with relying solely on that approach. The first one is that it assumes who have the same diagnostic label are all the same as each other.
Chris - You're genuinely comparing apples with apples but they actually might not be?
Duncan - That’s true.
Chris - Yeah. So who did you recruit then?
Duncan - Well, we set up a centre which we called the Centre for Attention Learning and Memory and to that centre professionals in education, in clinical services could refer children to us that they thought were struggling. They didn’t have to have a diagnosis but they could, or they could have multiple diagnoses. They just had to be struggling in the areas of attention, learning, or memory, so intentionally generic.
Chris - And how did you study them?
Duncan - Each family would visit us and the team of research assistants and PhD students would spend several hours doing different kinds of cognitive assessment with the children. And we would have behaviour ratings for the parents to fill out and most of the children would go through the MRI scanner.
Chris - So you’ve got what their performance is like, what their performance is like as in their track record, what their performance is like in your tests and a brain scan to go with. And what, you then marry all that information, or at least you’re asking a computer to marry all this information together and look for common ground or differences between them?
Duncan - Essentially, yes. We took many of the cognitive assessments and we fed them into a machine learning algorithm. Machine learning sounds kind of fancy, but actually we use it all the time in our everyday lives. Every time you type something into a search engine, behind the scenes there’s an algorithm which is learning about that information you’re feeding it, and you might notice that appear in your adverts that correspond to what you’ve searched for. Our machine learning algorithm was learning about the cognitive data that we fed in from these kids and what it learnt was that there were different profiles for children. Children with different profiles of cognitive difficulty.
Chris - What does that mean in practical terms? When you say different profiles are you saying say I had a diagnosis of ADHD, I’ve got that label but actually my ADHD may be quite different from the label of ADHD you might receive, for example?
Duncan - Yeah, exactly. Because one thing we could do is then check about what the machine learning algorithm has learnt and see whether it’s really learnt the categories, the diagnosis the kids came with. And the data showed very clearly that’s not what the machine learning was learning. Children who had a diagnosis of ADHD could have very different profiles from each other. They could have very different cognitive strengths and difficulties and that’s a real challenge in trying to think about how we support those kids.
Chris - Does this mean then our categorisation is just wrong? We’re putting people into narrow bins of problems and, in fact, it’s much more subtle than that and we need much sort of narrower but wider categorisation?
Duncan - I think it means the diagnoses, we’re not thinking about them in the right way. So they’re not like kind of medical diagnosis, they’re much less discrete and clean than that, and we don’t really understand what the underlying causes are. We still think that a diagnosis is a real landmark moment for children of families when they get some professional recognition for the challenges they’ve been facing. The question is how do practitioners then best support those kids, and how do we equip those practitioners to do that? And the answer is simply knowing the diagnostic label itself isn’t enough information to go on.
Chris - Indeed. At the end of the day we’re dealing with an individual here that’s got a problem that they want help solving isn’t it? So does your tool give us a better insight into okay, we can identify where this person’s weaknesses are so we can then go to the classroom and say if you augment the training in this direction, or give this person extra aids, perhaps more stimulation, more practice in this area this will help to develop this area that they are clearly deficient in?
Duncan - Well, we believe so. So, for example, in the data a large proportion of the children have problems in short term or working memory. Those children could have a diagnosis of ADHD, they could have a diagnosis of ASD, or they could have no diagnosis at all.
Chris - Autism spectrum disorder?
Duncan - Autism spectrum disorder, or they could have not diagnosis at all. But we know that if you try and reduce working memory demands in the classroom, then kids with poor working memory will do better. We already know that there are some interventions out there that are effective with these kinds of cognitive difficulties, it’s just that they seem to cut across the traditional diagnostic boundaries that we’ve kind of hitherto believed in.
- Previous Scooping up space junk
- Next Editing the mitochondrial genome
Comments
Add a comment