Jo Carlowe Hello, welcome to the Papers Podcast series for the Association for Child and Adolescent Mental Health, or ACAMH for short. I’m Jo Carlowe, a Freelance Journalist with a specialism in psychology. In this series, we speak to authors of papers published in one of ACAMH’s three journals. These are the Journal of Child Psychology and Psychiatry, commonly known as JCPP, the Child and Adolescent Mental Health, known as CAMH, and JCPP Advances. Today, I’m interviewing Helen Minnis, Professor of Child and Adolescent Psychiatry at the School of Health and Wellbeing, University of Glasgow, and Alessandro Vinciarelli, Professor of Computational Social Intelligence at the School of Computing Science, University of Glasgow. Helen and Alessandro are authors of the paper, “The Use and Potential of Artificial Intelligence for Supporting Clinical Observation of Child Behaviour” recently published in CAMH. This paper will be the focus of today’s podcast. If you’re a fan of our Papers Podcast series, please subscribe on your preferred streaming platform, let us know how we did, with a rating or review, and do share with friends and colleagues. All listeners to this and indeed, any of the ACAMH podcasts, are eligible for a free CPD certificate. Do please visit acamhlearn.org for details of this, together with information on how you can access hundreds of hours of free talks, lectures, interviews, all of which you can also get free CPD certificates for. The URL is www.acamhlearn.org. Helen and Alessandro, welcome, thank you for joining me. Can you start with an introduction about who you are and what you do? Professor Helen Minnis Hi, I’m Helen Minnis. I’m Professor of Child and Adolescent Psychiatry at the University of Glasgow, and I’ve had an interest in attachment and attachment disorders for many years. And then, when I met my colleagues in computing science, I developed an interest, also, in artificial intelligence. Jo Carlowe Good. Alessandro? Professor Alessandro Vinciarelli I am Alessandro Vinciarelli. I am a Professor in Computational Social Intelligence, and this is a domain in AI that concerns the automatic analysis of behaviour for trying to understand automatically social and psychological phenomenon taking place in interaction. And that naturally led me to consider the detection of mental health issues as a potential area of interest. Jo Carlowe Thank you. So, today, we’ll be looking at your CAMH paper. Before we go into the detail, can you tell us what is currently known about the use of social artificial intelligence within child and adolescent mental health services? Professor Alessandro Vinciarelli So, there is a lot of research around the world in order to use artificial intelligence in general for the detection of mental health issues and in particular for our concerned children and adolescents, because there are a particularly vulnerable population, and at the same time, they are particularly used and familiar with digital technologies. So, from this point of view, they are a natural potential population that can benefit from these type of technologies. Now, there is a lot of work at the moment in research. There is no real application yet, because there is still a gap in terms of organising randomised clinical trials that really clinically validate. But there is a lot of research work and all indications show that artificial intelligence is capable, at least, to address the most evident cases, those that even, let’s say, the less experienced Doctors can deal with. While leaving, potentially, to the most experienced Doctor, to leave it to the professional skills of Doctors, the possibility to deal with the more grace – grey area type of problems, so where the problem is a bit more ambiguous and challenging. And very important, there is clear evidence that AI can address that part of the clinical work which is the most repetitive, the most tedious, the one that clearly doesn’t require particular professional skills. So, I don’t know if Helen wants to follow up on this? Professor Helen Minnis Yeah, I mean, I think that’s absolutely right, Alessandro. I think there’s huge potential and there are some really exciting research avenues, but as a Clinician, I don’t think these technologies have really reached us in the clinic yet. You know, the work that we’ve done on the School Attachment Monitor is, in that sense, kind of, state of the art, because these tools are not being used in the clinic. But one of the things I really learned from Alessandro is exactly that, that there’s a hu – there’s huge potential for AI taking away some of the repetitive tasks in child mental health assessment and then freeing us up to actually use our clinical brains. So, I really hope if you ask us that question in ten years’ time, we’ll be giving a very different answer and we’ll be using it effectively. Jo Carlowe We’ll focus a bit more, later in the podcast, on, sort of, AI use v. Clinician use. Let’s focus on the paper itself. So, this is “The Use and Potential of Artificial Intelligence for Supporting Clinical Observation of Child Behaviour,” recently published in CAMH. Can you set the scene for us? What did you look at and why? Professor Helen Minnis Maybe starting with ‘why’? I mentioned at the beginning that I’ve always been interested in attachment and most – I find that most CAMHS Clinicians are interested in attachment. And we think about it a lot and we talk about it a lot, but we’ve never had the potential for actually measuring it empirically in the clinic. So, there’s been all of this research using, for example, the Strange Situation Procedure, the Manchester Child Attachment Story Task, other story stem measures, in research, but it never really translates into the clinic. Because these are tools that we can use them in the clin – they’re not that difficult to administer, but they’re very, very time consuming to read. And I’m also a Child Psychiatric Epidemiologist and another frustration is that none of the big epidemiological studies really have true measures of attachment, because in order to really measure attachment, you have to slightly stress the child. You know, to really measure attachment, you can’t just use questionnaires. So, we’ve never had easy measures of attachment in big epidemiological studies, so that’s another big gap that really frustrated me. And meeting Alessandro and hearing about social AI, I mean, do you want to say more about that side of things, Alessandro? So, I really saw the potential for that. Professor Alessandro Vinciarelli Yes, what was very good in collaborating with Helen is that somewhat, we started from an existing test, which is the Manchester Child Attachment Story Task, quite a common instrument for assessing attachment in children. And somewhat, we modified, slightly, the protocol to allow artificial intelligence to work. This is a very important point, I think, of this work, because in generally, the interaction with Clinicians, the tendency is to think, okay, we do a lot of thing the way we do it usually, and then you just add the AI at the end. While the best results can be obtain if you really modify, you create the conditions for artificial intelligence to work. So, there has to be a little bit of adaptation on both sides to really work well together. So, it’s a very disciplinary work from this point of view. And the key point is that at the end of the day, the way this task works is that you ask children to tell stories. The way they tell it and what they say is the target of observation of Psychiatrists. By looking at that, they go to a protocol, they observe. Essentially, by observing, they understand what is the attachment of children. And this is something that we have been doing a lot in social AI, in the sense that analysing automatically, with computational methods that are agnostic about the problem, i.e., observing data analysis, somewhat, you can learn, hey, through statistical methodology, to bridge the gap between the data itself, the words – the way the words are say, so the tone, the loudness, the type of intonation, etc., you can map that into the judgement of the experts. So, the machine learns from the Doctors and replicate the judgement of the Doctors, up to a certain extent, of course. Jo Carlowe Is there anything more you want to say about the methodology? Professor Helen Minnis The only other thing I would say about the adaptation of the MCAST itself was that I had trained in using the doll’s house MCAST. So, the way that the original Manchester Child Attachment Story Task works is that you sit with the child with a doll’s house and you have mummy doll and child doll, and as Alessandro says, you tell the child an attachment related story. So, for example, you might have child doll in the garden playing football and mummy doll in the kitchen and the child hurts their knee, and you’ve got to put a bit of emphasis on that. “The knee is bleeding,” so that there’s an attachment related stress. Securely attached child will bring mummy doll close immediately and smoothly and in insecure attachment, that proximity seeking will be less efficient. And I’d always been struck by, when we were rating these, that much of the rating algorithm rests on how that proximity seeking happens and what happens immediately afterwards. And so, what our colleague, Stephen Brewster, did was he put movement sensors into the dolls, so you could actually see physically and actually measure in terms of physical vectors the speed and smoothness of that proximity seeking. One of the things in the original paper, that was published in PLOS One, that algorithm met – leant a great deal on the child’s hand movements, because they were holding the tangible dolls. What we did was we went into primary schools and with about 130 children, who basically, took it in turns to do the original MCAST with the doll play or our computer game version. So, we developed a very simple computer game where the story was told by an actor on a laptop and there was literally a very basic fuzzy felt version of a, kind of, map version of a doll’s house in front, and it worked really beautifully. And so, children took turns to do each. So, for most of the children, about 100 of the children, we had both an original version and a computerised version and that’s how the algorithm was developed. So, I don’t know if you wanted to say more about the development of the algorithm, Alessandro? Professor Alessandro Vinciarelli So, yes, the key point is that mental health leaves traces in the way we behave. There are traces that are particularly evident, like the one Helen mentions, so this proximity between dolls, which is very easy to capture, very easy to see. There are also traces that are less visible and artificial intelligence here acts a little bit like a microscope, because it has this ability to consider a lot of different variables and how they relate to one another, which is something that escapes our own cognitive abilities. And for that point of view, it can detect also traces that are less understandable. In general terms, they cannot even be described or expressed from that point of view, and still, they are sufficiently, consistently associated to the condition to allow automatically, to allow statistically, to make correct inferences. So, we analyse all types of behaviour about the dolls that represent, very semantically, what happens in the story. We analyse what they say. We analyse how they say it. We analyse facial expressions, and all these microtraces consider that we analyse, really, windows of 25/30 millisecond and then we do statistics out of the results we get from each one of them, all of them mathematically, statistically, they can map into a decision. Is the child secure? Is the child insecure? And we get accuracies between 75 and 80%, meaning that three times out of four we correctly made a decision, and of course, we talk about the cases that trouble you, with most evidence, there is a most clear trace of the condition in the behaviour of the child. Jo Carlowe Just staying with that. Clinicians, through their training and experience, they have implicit knowledge and an intuitive feel about what they are seeing. Your paper states, I know, that “social AI approaches can automatically detect a smile and estimate the probability that it is a manifestation of happiness.” Can social AI replicate this – the, kind of, implicit feel that Clinicians have? Professor Alessandro Vinciarelli As a matter of fact, this is what artificial intelligence does, right? Artificial intelligence, the name is a bit misleading, right? There is no intelligence in artificial intelligence, right? It is just mathematics. It is number crunching and what it does is that it learns to replicate human judgements about the data. So, human judgements can be very objective and simple. A lot of work, for example, about detecting cat in images, right? Whether there is a cat in an image, okay, that is very objective. So, you take thousands of images where there is a cat, thousands of images where there is not a cat and essentially, you learn the statistical properties of the two types of image and based on that, you learn to do it. And here happens exactly the same thing. So, we had experts analysing that 104 children that were involved in the experiments, expert that gave their judgement, right? “This child is secure.” “This child is insecure.” And what we do is that we teach the machines through a process that is called training, which has nothing to do with the training we have in the human sense. It is a mathematical process that simply takes mathematical factions and change the varia – their parameters to reproduce the best possible, the judgements of the experts. So, from this point of view, the machine does nothing else than learning, embodying, capturing this type of implicit knowledge that Doctors have about the behaviour of children. Professor Helen Minnis And I think – I mean, just to maybe add something to that. One of the things I’ve found really fascinating in working with Alessandro and his team is that I think AI and working with Computing Scientists can help us to understand what it is that we’re noticing when we make those intuitive, implicit judgements. So, I mean, I’ll give you an example from reactive attachment disorder. So, reactive attachment disorder is characterised by, really, a closing down of the attachment system. So, children are less likely to seek or accept comfort in an obvious way. So, they can become very withdrawn, and we sometimes talk about ‘frozen watchfulness’. So, there’s a, kind of, clinical implicit judgement that a Doctor or a Psychologist or a Speech and Language Therapist would make that a child is frozen, but what I’ve learned from Alessandro is that what – sometimes what the computer is noticing is, for example, the child’s leg movements. Now, as a Psychiatrist, and same with my other CAMHS colleagues, we’re often only really looking from the head upwards, or we think we are. We don’t realise that we’re actually taking in the whole Gestalt. So, for me, one of the really fascinating areas of potential next steps are to help us actually understand what it is we’re noticing when we make these human, kind of, Gestalt clinical judgements. Jo Carlowe Hmmm hmm, and what other key findings from the paper would you like to highlight? Professor Helen Minnis For me, the biggest and most exciting finding is the finding that particularly for insecurely attached children, a machine learning algorithm is usually correct. And then, the second most interesting thing is that that still leaves huge scope for the Clinician. So, the way that I could imagine this being used in practice in CAMHS clinics would be a little bit like the way that Cancer Doctors, for example, use radiology, CAT scans for example. A Cancer Surgeon may look at a CAT scan and may think that there’s something on it, but they’re not going to just use the CAT scan. That’s going to be a tool that’ll help them, along with other clinical signs. So, for me, that’s the other really important finding, that the machine learning algorithms are really pretty good, but they’re not good enough to replace the Doctor or the Psychologist or the Speech and Language Therapist. That we’re always going to need the Clinicians to add the rest of the assessments to really decide on what’s going on with a child and make our formulations. Professor Alessandro Vinciarelli I think a very important point that is emerging from the work we have been doing together with Helen, and also, partly with work I do on other mental health related problem, like depression, etc., is that probably the next decade, this is general for artificial intelligence, will be the decade where we bridge the gap between AI and users. In this moment, all these technologies are somewhat developed a bit with a technology centric approach. What does it mean? It’s developed by Engineers for Engineers, for people that have an AI literacy and so, people that are fundamentally already adept at using these technologies. While the very important thing will be to develop this and partly with Helen, we did it already through this type of game we design in order to interface the test with the computers. It will be important to make it sure that these technologies communicate the results in a way that can be used to the Clinician, in a way that the Clinician can rely on it without over rely on it, so always while keeping the critical sense and the ownership of the diagnostic process. And this is going to be extremely interesting because it is about human AI teaming and collaboration, right, which is still a very widely open type of such question. So, I would say the most interesting point is that we have this gap to bridge. It is probably what is going to finally help us to have this type of work applied and used. Jo Carlowe Focusing on that, so this is the translation of AI into clinical practice, this is something you mentioned earlier, Helen, when you said that formal attachment assessments, such as the Strange Situation Procedure and the Manchester Child Attachment Story Task, they’ve been used for decades in clinical research. But because they’re costly and time consuming, they’re rarely used in clinical practice. So, do you envisage social AI technology bridging that gap and how so? How do you see that translation happening? Professor Helen Minnis To give you an example, we used to use the Strange Situation Procedure in our clinic, but we only had one person in our team who was trained in the rating of it, and she retired. It’s very expensive to train in the rating of the Strange Situation, and even when you have, it literally takes hours. We estimated that to rate, using human ratings, all of the videos that we gathered in the study, which were Manchester Child Attachment Story Task videos, it took about 500 hours. You know, we just don’t have that time and that’s why I think Clinicians don’t use these tools. So, if it were possible for tools to be developed in the collaborative way that Alessandro has discussed, where Clinicians could be given a, kind of, good enough AI rating of a Strange Situation, and we could get good at looking at these alongside the other information that we’ve got, then I think we could actually be in the era of using attachment to really inform our clinical assessments in a way that we just haven’t been able to in the past. I think it’s very exciting. We’ve been working with Attachment Researchers, so including Carlo Schuengel and Pasco Fearon, and there’s some really quite exciting work coming up using the Stange Situation, and also, our Computing Scientist colleague, Marwa Mahmoud. So, there’s some – you know, watch this space. I think there’s going to be some really interesting research coming out that hopefully, will have some real clinical relevance in this field over the next few years. Professor Alessandro Vinciarelli Yeah, one important thing to remember, because there has been a little bit in the last decade, this tendency to think that artificial intelligence is going to replace workers, which is absolutely not true. Now, whoever was serious about artificial intelligence knew this will not been – will not have been the case, right? The point of artificial intelligence is essentially, about making it faster, making it cheaper, making it more efficient. That is where artificial intelligence can intervene. It’s not about replacing Clinicians. It is about helping them to make their work not even better, but simply faster, with higher volume, etc., reducing the cost, which is extremely important in nowadays. So, this is where artificial intelligence intervenes. And then, as I mention, in the best cases, it is also possible to get some extra information, because the type of analysis you can get out can give you the type of intuition that maybe will not be easily accessible at naked eye, right? So, once again, AI can act a little bit like a microscope in this case and help you to see things that you cannot see. Jo Carlowe Do you think policymakers understand that? Your paper emphasises the use of social AI to “improve clinical efficiency without replacing human clinical judgement.” But given the cost and time savings AI might present, might policymakers push for its overuse at the expense of Clinicians? Professor Helen Minnis I think that’s something that has to be thought about with any new tool, whether it includes AI or not. I mean, I think a good example would be that there are various profiling tools for neurodivergence, for example. People can go online and go to their GP, having done a profiling tool, and say, “This says that I’ve got ADHD,” and actually, that is not a clinical assessment. So, I think as far as policymakers are concerned, they have to be no more or less wary than they should be of any new tool. It’s got to be evidence-based. It’s got to have been developed with Researchers and Clinicians working in collaboration. We have to be really careful not to over rely on tools and also not to allow new tools to introduce bias. So, I think because AI technology has advanced so rapidly, there are perhaps greater fears about tools that involve AI, but we should have those fears about any new tool. Professor Alessandro Vinciarelli The regulation and the auditing, now, as it’s called, which is a process that for any technology, tries to identify and mitigate potential harms, in the case of artificial intelligence, is still a subject of research. So, for the moment, on the side of policymakers, there is an application of what they call a ‘caution principle’ right? Let’s go slow, let’s try to avoid, first and foremost, to avoid issues, to avoid problems. However, there is an enormous amount of work and in particular, we are going to work about exactly this technology to detect attachment in the framework of a project that is starting participatory approaches to the auditing of AI technologies. Which is in the framework of a big UK initiative, called Responsible AI, that really targets – is putting together several kinds of Researchers, where the point is, a bit in the spirit I mentioned earlier, to bridge the gap with the users. Let’s try to make sure that we avoid any possible harm resulting from these technologies and let’s do it by taking into account all people that are involved. So, in the case of mental health, will be the patients, the families of the patients, the Clinicians, the political authorities, the administrative authorities around healthcare, etc. So, from this point of view, it’s still an open issue, right? How are we going to assess and ensure? However, part of it now starts being established, one was mentioned by Helen, is avoiding biases. As I mentioned, AI learns from Clinicians, which means it learns the good things, so their expertise, etc., but also bad things. Biases, if there are any, is an obvious case. So, at least for our concern, learning to avoid the trap biases, and that the result is – well, doesn’t take into account any factor that is not related to the pathology. So, somewhat, at least some data, there has been a lot of advancement, so we can safely say that some of these approaches that we can measure it are unbiased, at least with respect to major characteristic, like age, genders, ethnicity and so on, right? But still, it is a very, how can I say, a very active research area. Still, we are interrogating ourselves, how can we ensure that these technologies are going to help with no harm, right? And it will be a constant interaction between policymakers, Clinicians, technology people like me that develop this, right? It’s a bit of a frontier that is moving. Jo Carlowe Thank you. How do you envisage social AI being further developed in the future to support child mental health? Professor Alessandro Vinciarelli So, somewhat, the progress goes, essentially, in two main directions that are parallel, but feed one another. The one is the consistent improvement of the methodologies. It is a technology, so there is a, kind of, technological progress that keeps going on. So, these analysis methodologies keeps becoming better and better. Working with less and less data, this is a very important thing. A lot of the successful methodologies today need an enormous amount of data, which makes their use impractical. So, we are working a lot to reduce that need for data. And this, let’s say, progress is irrespective of the particular area of publication. This is purely technological thing, right? In the future, the big issue, the big important problem, will be exactly how do we make it useable? How do we make it an asset for the users, right? Which introduces a number of problems that are independent even of the performance. It’s really about to make it, for example, explainable. To a certain extent, it will be important that it says we think this child is insecure for this, this and that reason. This is something that current methodologies are not capable to do. You get some indications, etc., but it’s a bit difficult to interpret them and to make sense of them. So, this is going to be the biggest problem, the most important problem, in bridging this gap, how to translate a decision into an explanation, right? I get to that decision because I observed this, this and that, right? And unfortunate, it is – how can I say? If you satisfy that need, you lose performance. So, there is this interesting trade-off to be found between working well and working in an explainable way for the moment, right? So, this will be, I think, the biggest problem, the most important problem to be solved in the next few years. Jo Carlowe Hmmm hmm. Professor Helen Minnis Where I think AI could be a real growth point in child and adolescent mental health is for those, kind of, quiet, invisible problems that we’re not very good at noticing with the naked eye. And so, I’m particularly interested, I think I mentioned at the beginning, Jo, that I’m particularly interested in attachment disorders. These are disorders that are only ever diagnosed in children who’ve been abused and neglected. And for some reason, Clinicians, as Clinicians, we’re not very good at noticing those. So, on the one hand, there’s reactive attachment disorder, which is characterised by closing down and children being very withdrawn. And we’re – somehow, our human brains are not very good at spotting the quiet children who sit at the back of the class and don’t ask for help. And then, we’re also not very good at noticing those indiscriminately friendly children who will wander off and chat to people that really, they shouldn’t be trusting, because that, on the surface of it, just looks like friendliness. So, I think there are areas in child mental health where these technologies could particularly help us, and it’s those quiet areas. I guess also children who are non-verbal, young infants, children with intellectual disabilities, who are unable to tell us how they are feeling. Clinicians who work with those groups of children, we rely very much on observation, and I think AI technologies applied to these observational techniques could be real growth points in the future. Jo Carlowe Thank you. Are you planning any follow-up research or is there anything else in the pipeline that you would like to share with us? Professor Helen Minnis Well, we’re really keen to do some more work on the School Attachment Monitor, but there is also some very exciting work about to start on attachment in general and AI. And I mentioned that Carlo Schuengel, Pasco Fearon, Marwa Mahmoud and ourselves, we’re all involved in that, so really, watch this space. And I think – I mean, I think there is work in other branches of mental health, using AI. I mean, I don’t know if you want to say anything about that, Alessandro? Professor Alessandro Vinciarelli Yes, somewhat. So, this is a world that can continue for decades and it has, at the same time, a methodological aspect. So, this data is extremely interesting to push these methodologies to their limit and to develop new methodologies to take into account. And another line of research, and it is something that we are going to publish, to present, is things that are coming. So, somewhat, for example, how aspects that are not attachment can help us to understand attachment. So, do secure and insecure children manifest different types of emotions, for example, and can this help us to make the decision? Is it possible to be multimodal, taking into account jointly, all ways of manifested attachment? So, what they say, how they say it, the gesture, facial expression, etc., and how to combine all of this. Something that, as humans, we do very easily, we don’t even realise we do it, but for a machine it’s a bit more challenging. Jo Carlowe Hmmm, brilliant. Finally, what are your take home messages for our listeners? Professor Helen Minnis I think for me, I would really encourage my CAMHS colleagues to be res – as respectful of artificial intelligence as you would be of any new technology, but let’s embrace it. Let’s become AI literate. Professor Alessandro Vinciarelli I think very important, don’t be afraid. It is not true that artificial intelligence is going to replace people. It’s going to help us to be more effective, more efficient. It’s going to eliminate the repetitive and tedious parts of our jobs, but it’s not going to replace us. So, embrace it, get ready to welcome it and let’s try to do the best altogether. Very much important for us technology-oriented Researchers, it’s very much important to collaborate in order to find the best way to make it work. Jo Carlowe Wonderful. Thank you, Helen and Alessandro. For more details on Professor Helen Minnis and Professor Alessandro Vinciarelli, please visit the ACAMH website, www.acamh.org, and Twitter @ACAMH. ACAMH is spelt A-C-A-M-H, and don’t forget to follow us on your preferred streaming platform, let us know if you enjoy the podcast, with a rating or review, and do share with friends and colleagues.