Episode Transcript
[00:00:00] Speaker A: This podcast is brought to you by.
[00:00:02] Speaker B: The Gosh Learning Academy.
Welcome to this very special episode of Gosh Pods. Practicing Pediatrics. The way in which we practice medicine is changing every day. With the advent of new tools such as machine learning and AI, the future really is open and bright. Today we're going to look at just one aspect of this and talk to Dr. Micah Kosters about HPO human phenotype ontology.
So I am very excited today to be talking to Dr. Micah Gustas about human phenotypic ontology. Micah, do you want to start by introducing yourself?
[00:00:40] Speaker A: Yes, of course, Sarah. So, I'm Micah Koesters. I'm a consultant in Pediatric Immunology at Great Ormond street. And I'm also very excited to tell you a bit about human phenotype ontology. We recently organized the HBO Day, and I think it's really important that people actually know that HBO exists and what it actually can do for you if you're a clinician or a researcher, and how we can connect for the future HBO in our healthcare system. So, yeah, I would really like to tell people more about this.
[00:01:13] Speaker B: Yes, absolutely. And I think it's a really important conversation to have because, like you said, I don't think a lot of people know what HBO is. And so you've touched on this a little bit already. But what would you like people to get out of listening to our conversation?
[00:01:27] Speaker A: I would be happy if, after this podcast, at least know about hbo, that it exists, what it actually is, and how they can use it. I think if they sort of have that information after this podcast, I'm already super happy.
[00:01:43] Speaker B: Amazing. Well, hopefully we will do that. So I think we should start with the most obvious question.
What is hbo?
[00:01:52] Speaker A: Yeah, so HBO is, of course, one of the many abbreviations in medical world. So it stands for Human Phenotype Ontology. It's already quite out there for a very long time. It has been invented by a couple of German guys probably 20 years ago, but it's sort of taking off in the last five, 10 years, I would think. So it's a standardized vocabulary, and it's basically describes every single feature of human disease and genetic disorder. So any sign, any symptom that you can think of should actually be described and coded in this system. And if you do that in this way, in a very structured way where you, for instance, also have different trees with overarching diagnosis, for instance, you can have infection as a coat, but then you can also have an infection in your chest or an infection of your skin or it could be bacterial infection or it could be recurrent infection. So if you do this in a specific way, where we all choose the same system and the same language and how we describe every single feature and every single symptom of human disease and of the human body, then we can actually make one system that we can use to speak the same language. And not just for, like, research or for genetics, but basically any clinician in the world can then say, let's agree on this. So if a child comes in with a snotty nose, then this actually can be coded in human phenotype ontology, not like snotty nose, but like a medical official abbreviation. And in this way, you can collect data which can give you a very accurate diagnosis. You can also identify new genetic diseases. You can also compare and use more computer, you know, analysis, because you all have the same sort of definitions of how you describe something. And then you can also connect this to research and integrate it with other systems. So to combine and integrate HBO with other databases, that's like massive potential for bioinformatics and discovering new diseases, especially in my field, because I work in inborn areas of immunity. You have so many rare diseases that your human brain can't really, you know, think of. So you need something else there to help you that can provide machine learning and can provide you to direct you in the right diagnosis at the earliest phase possible, so that you can guide also the treatment that you have to give for the child.
[00:04:21] Speaker B: So would it be safe to say that it's a way of universally coding, and then people can then bring that information together and pool it and look for patterns within it?
[00:04:31] Speaker A: It's exactly that. It's like, actually quite crazy if you think about it, that we haven't agreed in the medical world how we call things, you know, so basically, the geneticists have been at the forefront of this. How to describe an ear that is a bit different in sheep, and how do you call this? Or a nose or an eye or an eyebrow? And so they have been already working on this system for many years. So they saw already the potential in how to connect it to the genetics. But the rest of the medical world, the other clinicians, they haven't seen this potential yet. So we sort of stumbled upon this from an immunology point of view. And then we noticed actually immunology is hardly in there. And we as immunologists should therefore define how we want to describe a T cell or how we want to describe specific conditions in the immune system. And therefore we have been working on this now in the last couple of years, because it starts with agreeing on this is the system, this is the language, this is how we call something. So yeah, that's exactly what you say. A universal system where we agree on a specific language to define human signs and symptoms.
[00:05:43] Speaker B: And I know that you have a metaphor that you use when you're trying to explain HBO to someone. Can you share that with us?
[00:05:51] Speaker A: So when we look at how we actually are collecting the signs and symptoms for patients, I would almost compare this as how do we make honey? How do we get to the right diagnosis? If you look at bees, bees are not just randomly flying around. If you have a beehive, then bees, they are only allowed to fly a certain sort of action radius around their beehive. And that's where they collect from the flowers the right ingredients, they bring them back to the beehive, and then in the beehive there's a very set up structure and the structure is decided by one bee only. And this is the queen bee, she decide what goes and what doesn't go. So they have an agreed language, they have an agreed system, they have an agreed culture. And if you don't comply to what the queen bee decides, you're basically out.
So the honey and the quality of the honey sits with the bees, how they fly. If they don't collect the right ingredients, then we're not getting to honey. But if, for instance, around the beehive there are not so many flowers because there are a lot of flowers destroyed, then the bees have to fly and fly even further out of the beehive in a way to collect enough ingredients to make the honey. If it's not available, you never get to the honey. So if you would translate this to getting to the right diagnosis, then basically what you need is for doctors to be really like willing to find all the right signs and symptoms, fly out, you know, far out to go, collect all the sort of evidence, all the signs and symptoms, bring it back to the beehive collected in a specific systematic way, and only then you can get to the right diagnosis. And so far in the sort of medical world, we have not decided who is the queen bee, what is the system that we are using and which is the language that we are speaking. So it's always a miracle that people get to the right diagnosis and even get to the honey. But the question is the quality of the honey good enough or can we actually even make it better if we would agree on one system?
[00:07:59] Speaker B: I love that. What a wonderful metaphor.
[00:08:02] Speaker A: Yeah.
[00:08:03] Speaker B: And so you, you've talked about this already, but let's dive into it, into a little bit more detail. So if this is used universally, how can HBO contribute to healthcare?
[00:08:15] Speaker A: Yeah, there's like huge amount of potential. It's not just for like diagnosis, but you can already imagine how things are done currently in healthcare. It's often, for instance, you get a letter, that letter is scanned in like a healthcare system. Sometimes it's even still on paper. You have paper files. So these are just scribbles of clinicians. If you would have a standardized system where these kind of, you know, flat string text information would be translated into a code, then you can actually diagnose all these diseases far more better than we are currently doing because then it would allow pattern recognition of the specific signs and symptoms. And then for an individual patient, you can also match them to other patients that are already known, for instance, in your hospital, or even wider on a national level, and can pinpoint children who have overlapping symptoms, who might already have a genetic diagnosis or a defined disease. And, and if your child is a full match based on this exact terminology, this might already direct you in the, you know, right diagnosis at a far earlier stage than otherwise would be possible. So that I think is a massive potential also for clinical decision. Because if, for instance, my system could already tell me, well, with these type of symptoms, then I would think this might be the diagnosis or this might be the differential diagnosis. Then I can say, oh, I'm only going to request these type of tests, for instance, blood test or imaging, or I'm going straight to genetics because I already feel that it should be going in this direction. So you could also say, I'm not going to do a lot of diagnostics. I'm going to target it specifically on the information I already have gathered from this patient. And then I'm only doing that what I think is the most likely diagnosis and the most likely direction this patient should be going in. If you think, for instance, more like research or medicine, like development of new medicine, if you would have a system where you could pull out all the patients who have the similar set of symptoms based on HBO coding, you can sort of say, ah, these are the children that I would like to have in a clinical trial, or these are the children that have no genetic diagnosis, but they're all the same in terms of sign and symptoms. So if we connect them together and then look at their genes, there might be overlapping mutations in their genes and that might lead to a new disease. So basically, basically discovery of new Rare diseases. And also what I think is really important, there is a huge amount of data present in healthcare, but how are we collaborating between the hospitals or national, international level with this data that we have? We're sitting on a huge amount of data, but we don't have a good way of exchanging this data or collaborating because of patients data safety, for instance, and how to exchange the data. If you only have coding, which is completely unidentifiable, then this is far more easy than to exchange this data and work together from this angle as well. I think also for patient education, you can sort of explain to them, look, you have this condition, but actually if you look at my other patients, they have overlapping symptoms and this child for instance, is five years older or 10 years older and has developed over time these type of new symptoms. And the likelihood of your child, child also going in this direction is 50% or 60%. But if I start this medicine now, today, because I have learned from the other patients, then actually we can make the direction of the clinical progression in your child different than the children because we already learned from those other children. So it's a huge amount of potential. And I think the next sort of thing would be machine learning and AI tools to actually integrate HBO with machine learning and AI and then integrated in the healthcare system as we currently have it.
[00:12:13] Speaker B: So is this something that's still theoretical or is it something that's actually being used at the moment on the shop floor?
[00:12:20] Speaker A: Yeah, that's a good question. So if you look like sort of internationally, then there are some coding systems. So you have for instance Snomed or you have ICD 10 or you know, depending on which one you're using. So those are normally the type of coding that sits behind the signs and symptoms. The problem list as we have it currently in our electronic health record systems. And this is quite universally used across the world. But if you look at those type of systems, especially for the ultra rare diseases that I am dealing with in the inborn errors of immunity, but across Great Ormond street, you know, many of my other colleagues will have children with very rare diseases. These type of systems, they are not providing the, the deep sort of level of the granularity of the phenotype of the exact signs and symptoms as how I want to describe them. So we're missing the exact accurate data and how we gather then the information. It's not deep enough to sort of recognize the individual patients and match them to other patients. So we're missing a more granular system and human phenotype ontology can provide such a system. So what we have done now quite recently is we actually integrated in our electronic healthcare record system in our hospital, which is really unique. This is like hand built by our own team in Great Orman suite. So we actually have now integrated this option of human phenotype ontology for the first time in an electronic health record. Purely data driven. We have a far more granular way of collecting the data. So we are hopeful that over time, once people are starting to learn and start to use this sort of part in our electronic healthcare record system, that then this data will provide far more detailed information about our children in our hospital with quite rare diseases.
[00:14:17] Speaker B: Can you give us an example of how it's worked?
[00:14:20] Speaker A: Yes, I can give an example. For instance, we have a condition, well, a group of conditions, umbrella condition. This is called hyper ige syndrome. These are children, for instance, with allergies with eczema. So you would not per se immediately think, oh, they could have an underlying immunodeficiency condition. Because there could be lots of children with allergies and eczema. But these type of children, they come with specific infections, for instance, quite deep infections in the chest or boils on the skin. And there are different genetic causes why you could have an hyper ige syndrome. So how can you differentiate these sort of children and can you differentiate them already at the age of 1, at the age of 5, at the age of 10? Because what most doctors would see, they would recognize a patient once they hit the full textbook signs and symptoms. But that's not what we want. We want to target and diagnose a child at the earliest age possible, so that we might already be able to support the child to prevent infections, to prevent problems later in life. Like why would we wait for a child to have so many chest infections, infections that in these type of conditions you can actually develop a hole in your chest? You should not have a hole in your chest. So if we can already diagnose you at an early age, we might be able to prevent these infections and therefore damage to your chest. So what we did, we went through the whole HBO 3s and for this type of condition, hyper ige syndrome, and then the most common genetic diagnosis, which is doc 8 and stat 3. And then we made a list of the all the patients that we have, but also in the literature, like what kind of phenotypical features do they have and how frequent are they present and are they already present at an early stage? Some symptoms, they only appear at a later age. So you would not see them because a three year old should not have them. So then you're not scoring high enough maybe to hit a high hyper ige score, but you still have it and you should still be picked up at an early age. So we updated the HBO 3s for these two conditions and then re ran our own patients to say yes, present or no, absent all these signs and symptoms. And then we also went back to like, like what if I go back to this child when the child was 5 years of age instead of 10? The child is 10 now. But what if I go back to the clinical data, how this child presented at 5, would I still already be able to recognize this child with this condition? And maybe not for the human eye, but when you pull this data, so completely unidentifiable data, just the terminology through AI tools, then some of the AI tools will tell you, oh doctor, this is hyperig syndrome. And some of the AI tools were even saying this is dog 8 or this is stat 3 mutation causing hyper IgE syndrome. This child should go to a doctor. This child should probably have an immunologist and a geneticist involved in their care. So I think that's a very good example of how great the potential is of human phenotype ontology, but also AI tools and how we can use them in the clinical care.
[00:17:33] Speaker B: Yeah, can we talk a little bit more about AI, which is like the buzzword at the moment in medicine. So how does this system of HBO fit in to some of the upcoming AI tools that we have available?
[00:17:46] Speaker A: Yeah, I think at this moment of time still quite challenging because what we actually have in the clinical setting is a lot of random data that is not connected and not translated into a unifying language such as hbo. So the first step should actually be having integrated sort of behind the firewall in your healthcare system. Tax mining tools that are not open source but are like staying like in the safety, you know, the security of your own healthcare system, then you would be able to pull out only hbo. So only like the medical terminology translated into an agreed like one system. Then you will be able to pull that out of the setting of, you know, patient identifiable because it's not patient identifiable at that stage anymore. And then you could enable sort of different types of AI tools to see if they are capable of diagnosing a child with the correct diagnosis at this stage. We for instance tried different AI tools. We compared nine AI tools at the same point of time.
They were quite scattered in the results that they were providing. So you still need to be actually an expert in your field, because some of the things that came out didn't make sense at all. For instance, if you have a condition suggested by AI tool which is X linked and therefore it should be a boy who is suffering from the condition because a girl has two X chromosomes, in principle you could have some inactivation, but in principle you should not have an X linked condition. Then if my AI tool is suggesting, doctor, this could be this condition, then I will would tell this AI tool often, I said, no, this is a girl. I told you, this is, this is not just a patient, she is a girl. And then of course the AI tool would apologize and say, okay, I'm very sorry, you're right, it's very unlikely. And then it would update the differential diagnosis. So you have to be still an expert in the field at this moment in time. Also, you don't know how the algorithm behind it actually works. If we as sort of our healthcare provider, as the nhs, if we don't have our own AI tools and our own data people making the algorithms and improving our own algorithms, if we are relying on what's out there at the moment in terms of available AI tools and we don't oversee and we don't have the knowledge set to understand what the coding behind it is in the algorithm, then we can't actually guarantee the outcome of the AI tool at that moment of time. So what we also did, for instance, we used the AI tools for this hyperig cohort, then a couple months later we used it again and actually some of the algorithms already improved in a time span of three months. So this is also how quick that field is moving. And I think we should move along as a healthcare system because if we don't, we might feel daunted by, you know, the availability of the AI tools and we feel they're not reliable enough or consistent enough. But even the difference of three months already for me, it was noticeable that some of the AI tools had become more accurate because if you enter the data and you enter good quality data, the algorithms are learning from what you are providing them and then sort of they recycle that and then if you ask them again a couple months down the line, they still have that information in sort of their AI brain and then they are using it for you in your benefit and then it actually improves. So you can also improve the AI tool by driving it, by feeding it your accurate data. So the accurate data is also really important.
[00:21:26] Speaker B: Yeah, it really does sound like it's the cutting edge of where we're going with medical technology and you've mentioned some of the challenges already, but are there any other challenges with doing something like this?
[00:21:40] Speaker A: I think the main challenge is that people don't like change. So we are very used to, this is how we do it. You know, we are a bit like sort of dinosaurs in healthcare where things are moving really fast outside the hospital or outside the healthcare system. And the moment we step into the hospital we have our mobiles on us. But are we using our mobiles, are we using the apps, are we using AI tools on our mobile? We're actually not. We're going back to sometimes paper files, we're going back to printers and fax machines and it's all going backwards almost. You step into the hospital and you feel they're going backwards like 20 years in time. So. And this is also of course a cultural things. We're quite weak, of course. We are humans, we are used to specific habits, we quite like that. We are as a doctor sort of making this decision, using our own human brain to sort of think about a patient. And if for instance, all of a sudden an AI tool is coming into my clinic and telling me, well, you should do this or it could be this, most people won't like that. And they would also fear it. They would worry that AI is going to take over patient care, which is actually not the case. If you look at the sort of pioneers in AI sort of in the world in the last 20 years already actually, then they would also tell you, yeah, there is a potential risk of patient safety and governance and all these things. So you have to have a very good security place in place. But equally it should allow for more empathy, it should allow for more patient interaction with the doctor because if you sort of can move away from your computer screen and you can use all the kinds of tools in the background while you're having a conversation with your patient or with your, the family, with the child, then you have more space to have a more human interaction, for more empathy. And this is something that is actually captured and sometimes AI tools can even tell you off if your patient consultation is less empathic than you could have been. So there's a lot of potential there as well to improve empathy and human conversation.
[00:23:53] Speaker B: Maika this has been just such a mind blowing conversation and I feel slightly overwhelmed but also kind of optimistic about how this can be used in healthcare and I really look forward to seeing it come to fruition in the next few years. Just to summarize everything, what would be your takeaway? Learning points from this.
[00:24:15] Speaker A: I think for me, the takeaway learning point is that at this moment of time we are not doing what we should be doing. We are not talking all in the same language, defining the signs and symptoms of every single patient in the world in the same way. And the human phenotype ontology, if we agree on this and if we work together on this on a national and international level, would finally allow us to have one system, one way to define every single sign and symptom of any patients anywhere in the world. And this would open a potential for every single patient to get the right diagnosis and to get the right treatment on time. I think if that's the only message I can bring across today, then I'm already happy.
[00:25:00] Speaker B: Amazing. Maika, thank you so very much for talking to us. It definitely sounds like a what's this space? Kind of topic and in the meantime I'll put resources down in the description below to the day that you ran recently a great Ormond street all about HBO and some of the videos and things that we have from that. I'll make sure that's linked in the description. But thank you so very much for talking with us today.
[00:25:21] Speaker A: Yeah, you're welcome.
[00:25:24] Speaker B: Thank you for listening to this episode of Practicing Pediatrics. We would love to get your feedback on the podcast and any ideas you may have for future episodes. You can find a link to the feedback page in the episode description or email us@digital learningosh.nhs.uk. you'll also find a list of resources and further reading in the description. If you want to find out more about the work of the Gosch Learning Academy, you can find us on social media on Twitter, Instagram and LinkedIn. You can also visit our website at www.gosh.nhs uk and search learning Academy. We have lots of exciting new podcasts coming soon, so make sure you're subscribed wherever you get your podcasts. We hope you enjoy this episode and we'll see you next time. Goodbye.
