View complete transcript
Using validated scales, such as the Rush Tic Rating Scale. In this episode, we explore a cutting edge approach. automated, video based tic detection using machine learning algorithms. I have with me Professor Alexander Munchau, from the Institute of Systems Motor Science, University of Lubeck, [00:01:00] Germany, and corresponding author of the paper, Automated Video Based Approach for the Diagnosis of Tourette's Syndrome, published in July issue of Movement Disorder Clinical Practice.
Welcome, Professor Munchau.
[00:01:17] Prof. Alexander Münchau: Hello, Hugo. Nice to speak to you and nice to see you. It's a pleasure to be here today.
[00:01:21] Dr. Hugo Morales-Briceño: Thank you. I wonder if we can start by you telling us about the idea of designing such a study and why was there a need to use a different approach for Tic detection?
[00:01:37] Prof. Alexander Münchau: First of all we were interested in the question whether it is possible to reliably detect movement, spontaneous movements in people generally. And this first step was done last year and was published also in the Movement Disorder Journal. And it was shown that using the algorithms that we have tested, It is possible to detect movements but then it wasn't clear, of course, whether these movements we detected [00:02:00] were Tics in patients or in healthy controls because we didn't compare that.
And this was now the second step to actually, to test different algorithms actually whether they are capable of distinguishing between spontaneous movements in healthy controls and tics, supposedly tics, in patients with Joule Stiller Tourette syndrome. There is still a question whether one can identify a single tic and say this is a tic as compared to spontaneous movement.
And so the question was, are there some features that can help us to distinguish the two groups and is this going to tell us something more about the nature of tics in addition to just distinguishing? Okay.
[00:02:37] Dr. Hugo Morales-Briceño: I understand that the assessment of the patients that were enrolled in your study were based on video analysis on facial movements identifying the characteristics of movements in these patients and that's something that would added to a machine learning algorithm to learn how to identify.
Tics. But [00:03:00] what were the main findings of this study? So how good was machine learning to say, Oh, this is Tic. This is a Tic that differentiate from healthy controls. And what sort of conclusions from these results can you draw from?
[00:03:18] Prof. Alexander Münchau: So I see this sort of, this study in a line of studies that need to be done to, to come to a point where we are actually able to distinguish a single movement hopefully. We cannot do this now. We can only say that the movements in group one are different from the movements in in another group, in group B or so.
And these groups are healthy controls versus Tourette patients. And so one of the questions, of course is are there certain features that allow us to distinguish the groups? And one feature is the excess of movement in Tourette patients. So are there simply more movements in these patients compared to healthy controls?
But of course this is not enough because then you could still say what you're measuring are just movements, but does it have anything to do with tics? We're not still we're still not there yet but there are some features the algorithm told us. in addition [00:04:00] to just the frequency of movements.
And this was the clustering, also the duration and the interval between movements. And we could now with the certainty around 90, 95 percent or so distinguish the two groups on the basis of these these clusters. So the occurrence of a series of movements in a given time, so the likelihood of such repetitive movements.
uh, Increased activity in the 5 year, 5 year period. To take the clustering as one important feature. There are also other features and now this, of course, needs to be taken further and then validated on a sort of Tic by Tic basis.
But as I just said, if clustering is an important point, then this already shows that a single Tic per se or a single movement cannot lead to a firm conclusion what it is, because you need. more than one movement to decide what the movement is. And this is one, one important message, I think, of the paper.
[00:04:53] Dr. Hugo Morales-Briceño: From what I read in the paper, though, there is also quite robust that these [00:05:00] differentiated features that you mentioned the number of Tic clusters and also the duration of them. So will be relevant to distinguish between patients and controls and seeing this clear use of these machine learning algorithms in identifying this patient and looking into future studies or actually implementing such technology in clinical practice.
What have you thought about this? What is the feasibility of implementing machine learning to help clinicians in the diagnostic process?
[00:05:38] Prof. Alexander Münchau: So first if we can confirm these data if if they are reproduced then this would be, I think, an important step to to further develop this device into a tool that allows us to distinguish not only tics versus spontaneous movements in healthy controls but also tic like movements in patients with functional movements others for example or the differential diagnosis [00:06:00] of myoclonal sometimes it might is it myoclonal tic usually this is quite straightforward on a clinical basis but sometimes it can be more tricky but the distinction between Tics, classical tics in Tourette syndrome and functional tic like movements in patients with functional disorders can be very tricky.
And of course, people can have both in a given person. And this is not only interesting from an academic point of view, but it's also treatment relevant because you would treat tics in a different way than functional tic like behavior. And therefore, this becomes more important. And of course, you need larger populations to do this.
To be certain that you can distinguish this and this needs to be a multi group effort. So this is step one. So differential diagnosis, then you could use these these algorithms for automatic analysis of video ratings. It can be a bit tedious, if you do this manually and it takes some time, then you could use variables and record Tics in more naturalistic settings.
So at home or in certain situations. This has also been done in other [00:07:00] disorders, Parkinsonism and tremor and so on. And because we know that the context, the social context of tics is very important. So then when one could develop this further into some variable solution that might be helpful.
[00:07:14] Dr. Hugo Morales-Briceño: Thank you, Professor Munchau. One of the other things that I found really interesting in the study is that actually this machine learning algorithm providing an output in estimating the probability of the Tic. And but despite being really good in doing that still the accuracy was not perfect and clinicians needed to sort of corroborate the presence of Tic in those cases, there was a low probability of Tic. And this is a way that clinicians need to supervise AI or artificial intelligence. Is this something that will become really important if [00:08:00] AI or machine learning algorithms come to the clinic?
[00:08:04] Prof. Alexander Münchau: for the time being, yes, I would have thought so. I think it is very important to still supervise it. For quite a long time, probably. And what we also suggested is some sort of hybrid solution in the sense that you use this as a tool. But then there are gray areas where you can't be, where the algorithm is not certain, particularly in those who have few Tics compared to healthy controls.
And then in these groups of patients then experts come into play and could then help and decide, to which group this particular person belongs. And also as regards clinical rating usually it is done, by two independent raters. And if they disagree, by more than 15 percent or so, as regards Tic frequency, then Then they have a consensus meeting and try to come to a point, and this could also be something where these algorithms help their tools additional tools in addition to clinical settings.
And of course, also for unexperienced colleagues who are not so well trained yet this might be very helpful also as [00:09:00] teaching tools. I think of it in the sense that the algorithm can only be as good as we are as clinicians, as experienced clinicians. And if we can feed the algorithms with the experience, of a given number of people who have done this for 20 or 30 years and all these years of experience can enter this algorithm.
And it doesn't make us superfluous in a sense, but it. It helps that all that we have collected across a lifetime is not lost, if I say bye bye to the setting and then all my experience is gone, this would be a shame and also this applies to other people, but if we can if we can collect it, save it then this might be helpful for others in the future.
[00:09:41] Dr. Hugo Morales-Briceño: You're, You're touching a very interesting point about how clinicians do phenomenological diagnosis and probably how AI or machine learning algorithms can point us to the very specific sort of details or nuance characteristics that [00:10:00] may help us to distinguish between chorea, myoclonus in Tics.
Or even determining the combination of these phenomenologies, which is complex in clinical practice. But hopefully this type of technology can help us to learn about these conditions better. Now, so this concludes today's episode, and I would like to thank again, Professor Munchau for his insights into the potential applications of AI based technology in the assessment of patients with Tourette syndrome.
[00:10:32] Prof. Alexander Münchau: Thanks a lot. Thank you very much.
[00:10:34] Dr. Hugo Morales-Briceño: Thank you, Professor Munchau. [00:11:00]