00:00:04: The Biorevolution podcast.

00:00:06: Your hosts,

00:00:07: Luise von Stecho

00:00:08: and Andreas Horchler.

00:00:13: Welcome to the Biorevolution podcast.

00:00:15: Today, the rare lens, AI-based image recognition for rare disease diagnosis.

00:00:22: Rare on a global scale appears misleading at first glance because more than three hundred million people are affected worldwide and suffer from rare diseases.

00:00:32: The problem, there are many, and diagnosis is not that easy.

00:00:37: Topic today.

00:00:38: Hello, Izzy, how are you?

00:00:39: Hi,

00:00:39: I'm excellent.

00:00:40: Thanks for the nice introduction.

00:00:50: We brought those who follow us on the video, might have seen already.

00:00:56: We have guests today.

00:00:57: Yeah, we couldn't hide them.

00:00:59: No?

00:01:00: You could have actually come out from under the table.

00:01:02: That would have been a nice... We come in again.

00:01:06: We have Benham and Adela today from Unibon and they will say a few words about what they can actually do to improve rare disease diagnosis.

00:01:15: But let's start with a quote as we usually do.

00:01:17: Absolutely.

00:01:18: Yes.

00:01:19: So today I brought a quote, which is actually not from a person, but from a position paper from Orphanet.

00:01:24: Orphanet is a very important database for orphan or rare diseases where you have disease information and a lot of other resources.

00:01:33: And what they say in this paper from two thousand twenty two is for patients affected by rare disease, obtaining a timely and accurate diagnosis is key to accessing the appropriate medical expertise, which I think it's very important and something that we'll talk about in a second.

00:01:48: But maybe we can introduce what image recognition and next generation phenotyping actually has to do with rare disease diagnosis.

00:01:56: And therefore maybe I can hand over to our guests to introduce themselves a little bit and what they bring to the table when it comes to rare disease diagnosis.

00:02:04: Maybe we start with Adele.

00:02:05: Yeah,

00:02:05: sure.

00:02:06: So my name is Adela.

00:02:08: I have a PhD in immunology, and my research was focused on the role of macrophages and monocytes, so immune cells in atherosclerosis, so all about cardiovascular disease.

00:02:20: But now at the University of Bonn, I have the official job title, Medical Science Liaison, and I work mostly on the Gestaltmetscher Project, which we will also talk about more here.

00:02:33: And my role is to help facilitate the transfer of our Gestaltmetscher tool from academia to the real world, to the patients and to the clinicians.

00:02:45: Excellent.

00:02:45: Good morning.

00:02:46: My name is Ben.

00:02:47: I have a PhD in astrophysics up in the sky.

00:02:50: I'm

00:02:52: expected.

00:02:53: Sorry.

00:02:53: No worries.

00:02:55: We don't fly that high usually.

00:02:59: I prefer to come back down to earth.

00:03:01: Okay.

00:03:01: I have years of experience working with big data, machine learning, in particular, a lot of image analysis.

00:03:07: And I'm happy that I moved to the field of AI and rare diseases around four and a half years ago, because I found out that I can contribute to this field with my data analytical skills.

00:03:17: I work with Adela, also at the University of Bonn, and I'm mostly busy with working on radiographic images and retinal images addressing rare bone diseases and rare eye diseases.

00:03:28: How did the word gestalt, because I'm the language guy, find its entry into the English language?

00:03:34: Yeah, and this is a really, really interesting part.

00:03:36: Maybe you can say something about that.

00:03:38: it's actually not just face but gestalt, right?

00:03:42: Yeah, so gestalt is a term that is used in human genetics, in dysmorphology, yeah, in German and in English as well.

00:03:52: And it's a term that encompasses like not only the face but the facial features and patterns in the facial features.

00:04:03: So patterns that are recognizable not only in one individual but can be identified across multiple individuals.

00:04:11: And patterns is a magic word for today, I think.

00:04:13: Yeah,

00:04:13: I think so.

00:04:14: Patterns and AI's capacity to see and analyze those patterns in a way that humans cannot maybe to frame it a little bit.

00:04:22: I think what is a critical thing, and you introduced that in the beginning already, there are a lot of different rare disease estimates.

00:04:30: It's a little bit hard to pin down because some say six thousand, some say seven thousand, some say seven thousand five hundred.

00:04:37: Maybe it doesn't matter so much.

00:04:38: It's a lot.

00:04:38: It's a lot of different diseases that affect a very small number of people often.

00:04:43: And the chance that a physician who's supposed to diagnose a person with a rare disease has already seen someone who has the same symptoms or maybe the same gestalt is quite rare.

00:04:54: Because if you have maybe five or maybe fifty people all across Germany and you have your GP, the chance to get the right diagnosis is probably pretty low.

00:05:04: And this is something that for patients with rare diseases or people affected by rare diseases is a common problem.

00:05:09: I think the average time to diagnosis is around five years.

00:05:13: This is a long time, of course.

00:05:14: And in this time, you don't get the right care or if there's a treatment available, you don't get the right treatment.

00:05:20: So having a faster diagnosis can be really important on the one side for medical care.

00:05:26: On the other hand, also, I think just seeking a diagnosis for a long time and they call it the diagnostic odyssey can take really big tall on a person and often it's of course many rare diseases around eighty percent have a genetic component.

00:05:40: That means they often manifest already in children.

00:05:43: So it's the caregivers, the patients who seek this diagnosis for their children.

00:05:47: And you can imagine that five years on average for some people it's much longer is really long time where you don't know what's wrong.

00:05:54: And you feel like maybe you're doing something wrong or what could I do better?

00:05:58: What can I do to help my child?

00:06:00: And having a diagnosis.

00:06:02: helps you also, first of all, to know what's going on and to, for example, adjust the lifestyle in a way.

00:06:08: For example, if you would say it's a disease that affects also the eyes, but wasn't recognized, maybe just getting glasses could be already helpful.

00:06:16: So here, having a better way to find diagnosis is really important.

00:06:21: And this is something where Gestaltmetscher and bone to gene come in, because often this genetic component of a rare disease actually translates into a specific morphology into a specific Gestalt.

00:06:34: Exactly.

00:06:45: So maybe you can introduce a little bit what the concept is behind the tools that you introduced in the beginning Gestaltmetre and bone to gene and also eye to gene.

00:06:54: So how does that actually work and how can AI help finding better diagnosis?

00:06:59: Yeah, so Gestaltmetre is an AI tool that uses advanced computer vision methods or next generation phenotyping.

00:07:09: to recognize rare diseases in a simple portrait picture.

00:07:15: So the idea is that in the clinics, the pediatrician or a different doctor would just be able to take a picture with their phone, have an app, the Gestaltmetscher app on their phone, and within a few seconds get a list, a ranked list of potential differential diagnoses.

00:07:35: Wow,

00:07:36: amazing.

00:07:37: Pretty cool.

00:07:38: For my understanding as the non-scientist at the table, we knew that certain facial features would help to find a diagnosis before AI came in, of course, but that left it open to the doctor and his or her expertise at that point in time to make the right pick.

00:07:59: And the app helps to narrow the possible diagnosis down, right?

00:08:04: Yes, exactly.

00:08:05: So the dysmorphology is a specialty in human genetics that already exists for longer than AI does.

00:08:12: These dysmorphologists are really expert in seeing these patterns in the faces and then making the link to a possible disorder that could be behind that.

00:08:22: But of course, this specialty is very difficult to learn.

00:08:27: And you also need to have some talent in recognizing these patterns and facial features.

00:08:32: So these experts are also rare, at least compared to the pediatricians.

00:08:38: So the idea behind Gestaltmetcher is also to bring this expertise from the dysmorphologists to the pediatricians, because they are the first point of contact of these children in the health system.

00:08:51: So if they already can make use of this knowledge, we hope that this whole diagnostic odyssey can be shortened.

00:08:58: Which would be really important.

00:09:01: And how about bone to gene?

00:09:02: How does that work?

00:09:03: It's also a similar concept, similar to disorders that affect the face.

00:09:08: The disorders that affect the skeletal system also manifest in the, of course, some part of the skeletal system, which are visible mostly in radiographs or X-ray images.

00:09:18: And again, trained dysmorphologists who have experienced with these kind of disorders might be able to identify them.

00:09:26: But you mentioned it correctly that, for example, we have an Atlas or reference book of Professor Spranger, late Professor Spranger, showing different radiographs of these different rare bone diseases.

00:09:37: But as you mentioned in the beginning, there are a lot of them.

00:09:40: And sometimes they have similar features.

00:09:42: Sometimes they have different features.

00:09:44: And the experts are also rare.

00:09:46: So with the similar principles, we are trying to train an AI that would identify these different characteristic patterns.

00:09:54: And when an x-ray is presented from a patient, it would list, again, a like top five or top ten most probable disorders or gene mutations that could be associated with these over seven hundred rare bone diseases.

00:10:07: Yeah, that's a good point, but you're mentioning over seven hundred diseases.

00:10:10: So recognizing which one you might have in front of you, if you know, okay, you have maybe, you see there is something with the statue, maybe, maybe a mature statue or some features that would be a little bit different than what you're used to seeing, but you don't know how to classify, right?

00:10:25: And seven hundred is already a lot.

00:10:27: And I think that the overall gestalt, the overall facial morphology is affected even more often, right?

00:10:32: And they say up to forty percent, I think, of ready.

00:10:35: diseases, which if you count at no six thousand, then you're at, I'm not so good in math, two thousand something around for maybe three thousand.

00:10:46: three thousand dish.

00:10:48: We talked about it yesterday, that my doing math without a computer or without a phone, unfortunately,

00:10:54: has decreased.

00:10:55: Yeah, you're better at it still.

00:10:56: But

00:10:57: the main thing is that there are a lot of them.

00:10:59: Yeah.

00:11:00: And it's really hard to, yeah, to recognize by eye.

00:11:04: Let's put it like that.

00:11:05: Yeah.

00:11:05: And I mean, on a personal level, to make the five years a shorter period of time to find the right diagnosis, especially when we're talking about kids, of course, you know, That's a big thing.

00:11:17: And if I narrow this down and shorten this period of time, it's perfect.

00:11:22: It's wonderful.

00:11:23: Yeah.

00:11:24: And from my side, I work as a consultant for biotech and pharma.

00:11:27: So for me, bringing in the other perspective, of course, the people who make drugs in the rare disease space, they would also like to find their patients.

00:11:36: Because often you have a very rare syndrome, and you know there should be maybe, I don't know, hundred, hundred fifty people in Germany, but you need to find them.

00:11:45: And again, here, I think there is a lot of synergy, what it's good for the patient and it's good for the drug developers.

00:11:51: And it's good for the physicians also who get answers for their patients, because I think especially in the rare disease space, physicians are quite closely linked to the families often because they see the patients a lot and they of course care to give them the best possible diagnosis and best possible treatment.

00:12:18: One question, maybe why do you think is AI actually so good at that better than us?

00:12:24: The main thing is that with the increasing the amount of data that we have and then the improvement of the algorithms, it's easy to when we have the data to train algorithms that can do this search and matching in a few seconds.

00:12:42: So I'm not going to say that the AI is definitely better than a trained human.

00:12:47: So maybe an experienced physician who has been working with these kind of patients for forty years is as good as AI, I would say.

00:12:55: But the point is that, as Adele mentioned, not all the physicians are that experienced and they're also not everywhere.

00:13:02: They are few at some specialized centers.

00:13:05: So that's how AI is going to help.

00:13:08: And that's how AI is going to make this big parameter search much faster and then save the time for clinicians to make better connections with the families, as you mentioned, and do the diagnostic workup.

00:13:23: You say when you have the data, you need to have that, of course, in the first place, and you need to have qualified data in order to make it not hallucinating, for instance.

00:13:36: So it's got to be reliable.

00:13:38: Where are you in this process right now, both?

00:13:42: Yeah, so for Gestaltmetsche, we have a process in place to acquire the data, because for rare diseases, it's even more difficult to get data because the diseases are rare.

00:13:54: So the data is also rare.

00:13:56: But for GestaltMatcher, we have the accompanying GestaltMatcher database, which is a curated database of medical images.

00:14:05: So mostly portrait pictures.

00:14:07: It also includes x-rays or pictures of other body parts.

00:14:12: And they are contributed by clinicians from all over the world.

00:14:16: So the database is open to clinicians and researchers.

00:14:20: They have to register.

00:14:21: We do a quick background check and then they can access the database because it's also really sensitive data.

00:14:27: We don't want it to be open to everyone.

00:14:30: And then they can upload their cases from their patients who gave consent.

00:14:34: And next to the images, we also have metadata.

00:14:37: So this can be date of birth, ancestry, of course, the molecularly confirmed diagnosis.

00:14:44: That's really important for training the AI and other data.

00:14:48: This is where Or this data then is used to train the Gestaltmetscher AI.

00:14:54: Is this database growing as we speak?

00:14:57: Yes, it's growing.

00:14:59: Can you recognize only or can the AI only make predictions about indications about diseases that it has seen from this database or can it go beyond that?

00:15:10: So the Gestaltmetcher algorithm, which was updated in twenty twenty three is particularly suitable for ultra rare diseases because the way the algorithm works and I'm not a bioinformatician.

00:15:23: So I will explain it in my terms and you can add to it if you like.

00:15:27: Each patient or the facial features are transformed into a high dimensional vector.

00:15:34: And this vector you can imagine as one dot.

00:15:38: in a clinical feature space.

00:15:41: And then if there is a very similar patient in the database, these dots will be close together.

00:15:48: So you only need one other patient in the database that is close to make the match.

00:15:55: And the closer two patients or two dots are, the higher is also the probability that the underlying pathogenic mutation change is the same.

00:16:06: So it could be either the same gene or a same downstream pathway of the gene that is affected.

00:16:11: Really interesting because I mean, as you mentioned, there are many ultra rare diseases where you really have very few patients or sometimes you only started to identify the first ones.

00:16:21: And here it's of course much more difficult than to find the right diagnosis.

00:16:26: Can I ask how you say devil's advocate question?

00:16:29: Why don't you just sequence the whole genomes of every newborn to find out if they potentially have some rare disease?

00:16:36: I mean, there are projects that are trying to do this.

00:16:39: I think the genome project in the UK is already quite far with this.

00:16:45: There are also some efforts in Germany.

00:16:47: I think at the University Heidelberg that are doing this.

00:16:51: And I think this opens a lot of other questions as well, ethical, but also organizational.

00:16:58: Do we only include diseases that are treatable?

00:17:02: Do we include all of them?

00:17:04: How do we explain it to the parents who are like with a newborn, exciting, lots of stress as well.

00:17:12: But there are efforts that are trying to do this.

00:17:15: The genomic newborn screening, yeah.

00:17:18: And if I may add, that's an excellent question.

00:17:20: However, sequencing data or DNA sequencing data alone is not enough.

00:17:25: That's the key thing.

00:17:27: It's just one of the many pieces of information that goes into the puzzle of solving the diagnosis.

00:17:34: And one of the main reasons for that is that when you sequence the DNA of all of us, there will be lots and lots of variants everywhere.

00:17:43: Of course, many of these variants are fine.

00:17:45: That's what makes us different.

00:17:47: But then a few of them are pathogenic.

00:17:49: And when you just get the sequencing information and you find thousands of variants, you don't know which ones are really pathogenic and which ones are causing the disorder.

00:17:59: And those are called variants of unknown significance.

00:18:02: So matching those genotype information with phenotype information and other clinical data is really essential for establishing a diagnosis.

00:18:10: That's why even before AI, clinicians were looking at the Gestalt, at the radiographs, at the whatever other frontotopic information that they could get.

00:18:19: And this matching is really essential.

00:18:21: Yeah, I think in the recent years, there was this paradigm shift.

00:18:24: So it used to be that you take the phenotype and then Make the sequencing or genetic testing and nowadays you need the phenotype to make sense of the.

00:18:37: genetic information.

00:18:39: And there is more genetic information coming, right?

00:18:41: We had like an episode about the dark genome.

00:18:44: So the other ninety eight percent of the genome that we're only scratching the surface off, which probably also have a lot of disease relevant information, right?

00:18:54: This is really interesting.

00:18:56: I think one thing you just mentioned, Adela, in context of what to do with the parents.

00:19:01: So how do you explain it?

00:19:03: the tools that you're developing are physician only, right?

00:19:06: So this is not something that a private person or a parent could use.

00:19:10: I think I saw that, for example, in Australia, there's a tool that also laypeople can use.

00:19:16: Maybe I'm wrong about that.

00:19:18: But how would you feel in general about having this kind of information?

00:19:23: I mean, in which hands should this information be?

00:19:26: Yeah, so now Gestaltmetcher is for clinician use only, but we are actually also exploring ways to make it accessible to the parents, maybe by integrating it into the appointment scheduling software like Dr.

00:19:40: Lipp, for example, that they can immediately, when they make the appointment for the routine screening at the pediatrician, can take a picture.

00:19:48: And even before the child and the parents come to the pediatrician, they already know, oh, there might be something.

00:19:54: But what is really important with this approach for us is that the result is not communicated to the parents, but communicated to the pediatrician who puts it into context and then communicates it to the parents.

00:20:09: Because it might be that there is a developmental delay, but the pediatrician knows the child usually from birth.

00:20:18: So maybe for all of the milestones, the baby was just a bit slower.

00:20:23: then it doesn't need to be meaningful.

00:20:26: So we still need the pediatrician in this equation.

00:20:30: I would agree strongly to that.

00:20:31: I think medical information in the hands of laypeople can be really dangerous type of information, because it's having the, as you say, the context.

00:20:40: Also, what does it mean?

00:20:41: It's a bit like this we had.

00:20:43: our previous episode was about.

00:20:45: dealing with a cancer diagnosis.

00:20:46: And I mean, cancer isn't cancer, right?

00:20:48: So it doesn't, it can mean that you have an overall survival rate of eight months or of eighteen years.

00:20:54: And this could be a very different outcome.

00:20:56: And I think the same, of course, is true for knowing you have a rare disease.

00:21:01: I mean, just the context of it, what does it mean?

00:21:04: Can you do something about it?

00:21:05: How much will it affect your life?

00:21:07: I think this is if you're not worst in the field and especially if it's something where the clinicians themselves don't have a lot of information is really hard to handle.

00:21:18: I think if you just throw it at someone without the filter of the physician's eye.

00:21:23: And the parents always have the right to know what was the result of the analysis.

00:21:28: So I think it's also important to support the pediatricians because they also take on a new role.

00:21:33: They now have to explain results from an AI with an algorithm.

00:21:37: How should they communicate this to help them and support also the communication of these results?

00:21:55: What is the type of feedback you get from physicians on these softwares?

00:21:59: Because I often, when I'm speaking at events, I get from physicians, AI doesn't work like this.

00:22:06: AI doesn't work.

00:22:07: I

00:22:07: don't use apps.

00:22:09: So, I mean, it also follows the previous discussion.

00:22:13: Although we try a lot to make sure about data quality that you mentioned and verifying all those algorithms, the point is that AI can still make error like any other software.

00:22:25: And that also makes it very important.

00:22:27: And then there is a pediatrician in place who can connect these different pieces of information and say, I know here, I think this AI is not making sense to what it's outputting.

00:22:38: But the other sort of feedback, and it's also very important, it seems, that we receive from pediatricians and clinicians is that they want to understand how the AI comes up with that decision that is outputting.

00:22:51: So they want to be able to interpret or explain the output as Adele also explained to themselves and also to the families.

00:22:58: So that's, I think, one of the main feedbacks we receive from the clinicians.

00:23:02: And that's one of the fields that is very active now to try to make sense or try to find additional ways to explain how the AI come up with this decision.

00:23:12: And then the more the clinicians understand that, the more they can trust it.

00:23:17: Because if it's a complete black box, they won't like it.

00:23:21: Yeah, understandably.

00:23:22: Understandably.

00:23:23: That's

00:23:23: a case for further education.

00:23:25: I mean, you know, for professionals, you know, all together, I think.

00:23:29: But

00:23:30: per se, the algorithm is a black box, or is it an explainable algorithm?

00:23:34: It's difficult to explain these algorithms, really, because they are... They are a mixture of millions, if not billions of parameters.

00:23:45: So if there are four or five, like some physical models, you can say, okay, if you increase that, that would be like this.

00:23:51: But when it's millions of parameters, you don't know how this combination is basically outputting some results.

00:23:58: It's an active field of research.

00:23:59: There are ways that we can shed some light on how this is done or.

00:24:05: what part of the image, for example, the model is paying more attention to when it's outputting the result.

00:24:11: But one cannot say that, okay, now we know how exactly these millions of parameters together are outputting this result.

00:24:18: Therefore, we have to, in parallel to trying to find ways to explain them, we have to ensure data quality, do a lot of external tests.

00:24:27: We have to make sure that the AI is performing correctly via these additional.

00:24:33: spot standard approaches of testing the AI on external cohorts, on different ethnicities, on different centers and hospitals, on different conditions, different ages, etc.

00:24:41: to make sure that at least maybe ninety or ninety-five percent of the time it's working correctly.

00:24:49: How would you define both of you the pathway for further adoption of the method in real life?

00:24:56: Yeah, I would say vor Gestaltmärsche.

00:24:58: The long-term goal is to implement it into the routine pediatric screening appointments.

00:25:06: So in German, ist das U-Untersuchungen?

00:25:08: Yeah, I remember.

00:25:11: From yourself?

00:25:12: Not from myself, from you.

00:25:15: Yeah, and then make it a part of it so that this expertise that I mentioned is really in every pediatric.

00:25:25: appointment included in every pediatric office, so to say.

00:25:29: All right.

00:25:30: Is there a time frame you would come up with?

00:25:33: Should this be implemented in like three years, five years, ten years?

00:25:37: Well, I would say the sooner the better.

00:25:40: But getting something into the health system as a routine thing that takes a long time.

00:25:47: Okay.

00:25:47: Would you say are the biggest hurdles?

00:25:49: Is it the physicians or is it really just the regulatory?

00:25:52: I would say the regulatory.

00:25:54: part and also the administrative part.

00:25:57: Yeah, I can imagine.

00:25:59: Especially in Germany, but I think also in general.

00:26:16: How about bone to gene?

00:26:20: I guess there are fewer skeletal affecting diseases.

00:26:24: Is this correct?

00:26:26: Yes, in total there are fewer specifically skeletal diseases.

00:26:30: but also many of the other disorders like the ones that affect the face also have some effects on the skeletal system.

00:26:36: So there's also some overlap between the different body parts that these different disorders influence.

00:26:42: In terms of development, bone to gene is much younger than Gestaltmatcher.

00:26:45: We are now at the phase of collecting data for training our algorithm.

00:26:51: We have focused at the moment on a list of disorders that are associated with short stature.

00:26:57: and a little bit more common among those rare groups.

00:27:02: And also for some of them, there are treatments and therapies available.

00:27:06: And it's very important.

00:27:08: if we identify these patients as early as possible, as you mentioned, their children, and then they're growing and the treatment window is short and it might be closed.

00:27:18: So our goal is to make the first AI model, AI product.

00:27:24: addressing this small number of disorders around nine or ten, depending on how much data we will be collecting in the next few months.

00:27:32: And then hopefully show that it works and it can help.

00:27:36: And then we will have similar challenges as Gestalt matter.

00:27:39: Absolutely.

00:27:40: Let's talk ethics a little bit, because maybe it has a smell even for those who are not in science.

00:27:49: Yeah, for Gestaltmetscher we also thought about this of course.

00:27:52: Of

00:27:52: course, you need to.

00:27:54: Which is why we implemented this registration process and it's not open to everyone.

00:28:00: With physicians it's quite clear that it's for a medical purpose, but of course if you give it to in the hands of everyone it's a different thing because you could think that also a picture of somebody else could be analyzed, not only your photo.

00:28:18: And yeah, this is a whole other path that we would take.

00:28:23: Are there any worries?

00:28:25: I mean, privacy, but also this question of someone takes a picture of my face and diagnosis me with a disease without me wanting this.

00:28:34: kind of situation.

00:28:35: Is that something that you have come across as a worry or is it because it's really physician limited, no challenge?

00:28:44: I think mainly that's a very important point.

00:28:47: And as Adela mentioned, if this is an app that everyone can use, then we'll be worrying because yeah, these things can happen.

00:28:55: But at the moment, Gestaltmatcher and similar tools are meant to only be used by clinicians and then Of course, we cannot control all the clinicians around the world, but we hope and we know that the majority only care about their patients and therefore probably they won't use the app on random people.

00:29:14: But yes, that's something that probably needs to be known and thought about and emphasized when releasing these kind of tools.

00:29:22: Yeah, and I think it's because I think facial features are something that is very explainable to the public.

00:29:29: I think it's something that is accessible to all of us, whereas looking at an x-ray or radiograph picture or retina picture, no one would be able to make any sense of, right?

00:29:38: Who is not, I mean, if I see a picture, if the bone is not clearly broken, I would probably not know what I'm looking at.

00:29:44: But face is something that we see all the time that we also use, for example, to open our phones based on image recognition.

00:29:53: So all these kind of things I think play into potential misconceptions around this technology, right?

00:29:58: Which I could imagine needs some explaining to make sure that people understand that this is a safe tool that is really only used for diagnosis.

00:30:08: Yeah, but as you said, I don't think it's necessarily an entirely new problem because you can already open your phone with your face.

00:30:15: So somewhere there must be data about your face stored somewhere.

00:30:20: Nobody, nobody worries about it or not that many people.

00:30:24: We worry, but we, as we always say, we go along.

00:30:27: I mean, we very often have this debate about privacy and then about all of us or most of us clicking all the boxes on the social media, whatever, Apple updates, whatever you do, you agree to many things of data usage that probably are not in your best interest.

00:30:43: And for medical data, we become like very concerned suddenly about certain things that we don't worry so much about in other contexts.

00:30:51: Yeah,

00:30:52: I guess it's also a question about does the risk of this equal the benefit that you could have for the rare disease patients?

00:31:01: And what do we need to implement to ensure that this balance is okay?

00:31:06: And I think that's a pretty clear answer you came up with in this show.

00:31:10: Izzy, what do you think about posing a visionary question?

00:31:14: A visionary question is something good, but I want to discuss very briefly, which might be visionary, but also something slightly scary.

00:31:25: This Chinese study from the Chinese Academy of Science, where they did the reverse.

00:31:31: So they used a genetic sample to reassemble facial features.

00:31:35: I have to say, like from the images they showed in the paper, I would not be sure that I recognize this person on the street, but it's similar maybe to this.

00:31:43: What do you call that in English?

00:31:45: There's phantom pictures that, you know, like the police draws.

00:31:49: Like

00:31:49: almost forensic work, right?

00:31:50: Exactly.

00:31:51: Yeah, well, it is.

00:31:52: It could be forensic.

00:31:53: Of course,

00:31:53: could be used for that.

00:31:55: So this is, I mean, has nothing to do really with Gestaltmatter, but it goes in the other way, right?

00:32:00: It's the question that if you can assume a genetic component based on the morphology of a person, you can of course also go the other way around and AI seems to be.

00:32:12: this was a study from this year, where which I didn't introduce properly, but where the scientists apparently were able to use genetic profile to reconstruct faces, which of course, leaves us with a massive amount of ethical questions, legal questions, and just shows that I mean AI and the link between genetics and morphology is really moving in a new direction.

00:32:39: I would say that if this would be something where I would say data protection for me would be very important and I would really like those tools not to be implemented.

00:32:50: because that could be quite scary to just collect DNA somewhere and say like, this is the person who left the chewing gum at the corner.

00:32:59: Now combine this with the genomic newborn screening.

00:33:01: Exactly.

00:33:03: Yeah.

00:33:03: And we have like a perfect big brother society.

00:33:07: Or a movie right now.

00:33:08: Well, I mean, it depends from which viewpoint you're going at.

00:33:12: But I would say maybe going back to the question of visionary now that I brought up this very interesting, but a little bit, as you say, like nightmare scenario, where would you say realistically will be with these tools in the next maybe five or ten years?

00:33:27: How not with, let's say, the bad side of it, but with a good side of it?

00:33:32: What do you think will be the impact on rare diseases?

00:33:35: Five years still is short, but we hope, really hope that tools that are already developed like eschatolmatcher can be part of the screening and diagnostic routine.

00:33:46: Also with bone to gene, we're trying to do that to make it run under the hood when an x-ray is taken.

00:33:53: And then if there's some anomaly detected, help with the clinicians doing the referrals and diagnosis much faster.

00:34:00: But I think we will have a lot of challenges really integrating this into the current existing workflows.

00:34:07: And I think that's going to be one of the main challenges because if these tools are going to be an extra hurdle for the clinicians, the adoption will not be that smooth and it will be delayed.

00:34:19: So I think one of the main things in addition to or in parallel to these developments that we're working on is to find ways how to smoothly include them in the hospital systems so that they actually saves time from the clinicians, not make it worse.

00:34:35: So I think that would be one of the main challenges.

00:34:38: But I really hope that we come together, we add these different models and pieces and AIs together, and together we find a way to make it easier and streamline it, hopefully, then maybe in the next ten years.

00:34:51: Yeah, I think for Gestaltmetscher, a visionary plan is also to combine it with maybe the smart glasses.

00:35:01: So that pediatricians have the smart glass and don't need to take out their phone, maybe even their personal phone.

00:35:09: Maybe they are concerned with that.

00:35:11: Also the children, they are young, they are moving a lot.

00:35:15: They don't like sitting still for an image.

00:35:18: So if they have a smart glass, maybe even we can make it work with a video.

00:35:24: And I think that would be also easier for the clinicians.

00:35:28: Really interesting.

00:35:30: really cool scenario.

00:35:31: May I add maybe from the industry perspective, of course, having better diagnosis and better disease understanding, which you get by linking the genetics to the phenotype, you have a better angle to develop new drugs.

00:35:44: And knowing that you have a better understanding of your patient group, of your patient pool, how many are there, who are the treating physicians is also a motivation, of course.

00:35:55: to bring new treatments to patients and especially in the rare disease space, we have a lot of options with gene therapies, with other new modalities.

00:36:03: So I think in a sense, I mean, we started this podcast three years ago by a revolution, like new technologies coming together to cure diseases.

00:36:12: And I think this is one of the examples where I would say this is really happening.

00:36:15: There is a lot of.

00:36:16: I mean, we just see the scenario of having your smart glasses to diagnose your patient and then you have a very specialized therapy that can be tailored to this patient.

00:36:27: And this is really future.

00:36:28: Now we're talking not ten years, but I'm saying twelve point five years.

00:36:33: Okay, twelve point five, but then a few seconds instead of five years for diagnosing.

00:36:38: That's quite a promise, I think.

00:36:40: That is a great promise, yeah.

00:36:41: Thank you very, very much for being with us today and explaining your Bonne worldview, so to say.

00:36:50: Very much appreciated.

00:36:51: Thank you.

00:36:52: Thanks for having us.

00:36:52: It was a pleasure.

00:36:54: Thanks a lot.

00:36:58: This was the by revolution podcast for this time easy as always.

00:37:03: we invite you to subscribe to our podcast to Put your comments on Spotify for instance or YouTube if you want to and let us know what we should have as a topic in the show.

00:37:17: and you find of course as always further information at

00:37:21: science-tales.com.

00:37:29: It's a story and not a wagging tail.

00:37:32: That's it.

00:37:33: Thank you.