Okay. Well, I hope you can all hear me, so we'll start this presentation. My name is Nikita Pazdeev, and I am an endocrinologist. I practice medicine at the Lodz Academic Center, and my practice is focused mostly on diagnosing and treating thyroid cancer. So, that is why this topic is very close to my heart. And, well, I don't think I will reach the high bar that we're all set in the previous talk of making it interactive and entertaining, but I will try my best. Okay, so I will talk about a very specific area of endocrinology, which is the management of thyroid nodules and diagnosis of thyroid cancer. As you know, thyroid nodules is an abnormal growth of thyroid cells, and two big types of thyroid nodules, they are non-neoplastic. Some of them are non-neoplastic, which means there is no underlying genetic alteration, and usually we get a diagnosis on surgical histopathology of adenomatoid or hyperplastic nodule in this case. And then we have neoplastic thyroid nodules, which are tumors, and they could be adenomas, follicular adenomas, oncocytic adenomas. So, those are genetically driven tumors, but they are benign, and we could have cancers, thyroid cancers. And clinically, our task is to separate the two categories of thyroid nodules on the left, which are benign from those which are cancerous and need an intervention, surgical intervention usually. So, I'll talk a lot about our clinical practice and also the gaps in the clinical practice, that's not just because I like to be critical. I think this is important for us to understand where the deficiencies are in the current clinical care, and this will serve as a motivator for developing computer-assisted systems for this use case. So, once thyroid nodule is discovered, in most cases that's actually incidental discovery on some kind of imaging, which is done for another reason, we start by doing thyroid ultrasound, and then identify thyroid nodules in the ultrasound and decide if biopsy is required to exclude thyroid cancer. And we do a lot of biopsies, about half of the million biopsies each year in the United States. And thyroid nodules are very common, and even despite that high number of final aspirations, we can't really biopsy all of those nodules, so we have to stratify somehow, and go just for those which are more likely to be cancerous. And that's why we have this clinical ultrasound-based risk stratification schemas, and probably the most commonly used one is from American College of Radiology, which is called THARADS. And the radiologist or endocrinologist is assessing each thyroid nodule using these five features, composition, echogenicity, shape, margin, and the presence of echogenic foci. And then, depending on the presence of suspicious features, the points are added together, and then we end up classifying this nodule into one of those THARADS categories, THARADS 1 to THARADS 5. And depending on the category, follow recommendations, whether that nodule requires a biopsy, or we can watch it, or we don't need to do anything. So this works, and that system is very useful, but there are some deficiencies in how we use it. So on this slide, I show some performance metrics from ACR THARADS performance in the research studies. The meta-analysis has been published that show a good sensitivity, we only miss about 10% of cancers, and good specificity and area of the receiver operating characteristic curve. The problem is this data is our research studies, and usually those are done by radiologists who are highly experienced in interpreting ultrasounds. And in some studies, even multiple radiologists evaluated the nodule and reconciled their recommendations to make it particularly accurate, and that's not how we practice medicine, of course. So that's why we went into trouble of looking of how exactly it works in the clinical practice, and basically took all ultrasound reports from just the routine clinical practice and large academic center, and matched to the diagnosis, and we found that sensitivity continues to be good. We don't miss a lot of cancers, but specificity was very low, which basically means the radiologists overcall risk features, and we recommend to biopsy many nodules, and then the overall, our ability to distinguish benign from malignant cancers is actually quite modest. And this is just a plot showing tarot points on the x-axis, and benign and malignant nodule distribution based on how those tarot points were assessed by the radiologist in the routine clinical practice. And as you can see, there is a major overlap between the two categories. And trying to understand why our performance is suboptimal, we found that one of the reasons is the human interpretations of tarot ultrasounds are inconsistent, and this table shows us the results from the study from Frontin-Tessler group, where several radiologists from private practice were asked to read ultrasounds, same ultrasounds, and then the agreement in the interpretation of these different features was assessed. And as you can see, despite the shape of the nodule, which is relatively easy to do, the radiologist only agreed in maybe 30-40 percent on the presence of these highest features, and the overall tarot's recommendations of whether to proceed with the biopsy was agreed only in 44 percent of cases, so that's very low. And that is one of the motivations for us to develop an AI system to help with our radiology images interpretation, is because we can make a system which is deterministic, which means once we give it an image, it will always produce the same risk assessment. So this is some statistics which we took from the literature of how well we do in choosing tarot nodules for biopsy from major U.S. and international centers, and we found that only 8 percent of biopsies produce the cytologic diagnosis, which leads to the cytologic diagnosis, which leads to referral to a surgery. So those are the test categories, five and six, suspicious for PTC or popular tarot cancer. And we also found that 74 percent cytologic diagnosis are benign, which means if we would know in advance that this is going to be the outcome from the biopsy, we could have avoided referral for the procedure. And I think the biggest problem comes from this two categories in the middle, three and four, because that's where the cytology is not definitive, the diagnosis is indeterminate, and we start doing molecular tests, which have lots of deficiencies too, and frequently we end up without the diagnosis. We don't know whether that's cancerous or benign tarot nodule, and then we have to refer those people to a surgeon, and lots of these nodules come back benign on the final surgical histopathology, and then we have this awkward discussion with a patient on the post-op follow-up. So you got the imaging for unrelated reason, maybe CT scan of the neck because of the car accident, and then we ended up doing tarot ultrasound, biopsy, molecular testing, diagnostic surgery, which was unnecessary because you've got a benign nodule at the end. And that is an imperfection in how we evaluate tarot nodules, and big motivation for us to improve the clinical workflow here. So here's another problem with our human-based risk stratification of tarot nodules. So here on one side of the image, I have benign nodules, and on the other side, I have cancers. And it's a little difficult to do this in this online mode, but I've presented in person this picture to experts in endocrine society and American Tarot Association, and asked them to guess which one are malignant and which one are benign. Generally, the split is close to being 50-50, suggesting that it's difficult for us to decide which of these are cancerous, and that is because some of the tarot cancers, they just don't look suspicious on the ultrasound. They don't have all those classic features like microcalcifications and hypoecogenicity, which we have in tarots. So that's the answer. Those which are on the left are carcinomas, and those on the right are adenomas. So to conclude this first part of this talk, we have several motivations to use artificial intelligence for tarot cancer diagnosis. First is the ultrasound-based clinical risk stratification schemas, such as tarots, they're difficult to use, and generally, the agreement between providers is poor. And those of you who had the privilege to listen to Susan Mandel at the American Tarot Association conference last week, she presented on the new guidelines, which are just coming. That risk assessment is going to get even more complicated, possibly more accurate, but it will be even harder to use. So I don't expect that radiologists and endocrinologists will have a better agreement in the future. We also know that 17% or so of biopsies producing conclusive results, which result in an additional workup, and sometimes unnecessary diagnostic surgery. And the best way to avoid those unnecessary medical care is not the biopsy in the first place, that model, which is benign. And most of our biopsies that we do, they are done on benign tarot noodles, and they are also potentially avoidable. Finally, some tarot noodles don't have suspicious characteristics on the ultrasound, and they would be missed by our current clinical practice. So just one slide briefly to illustrate what we mean by tarot noodles. So just one slide briefly on what kind of aspect of artificial intelligence we will be using. So AI is a big field, and it has lots of different fields, as I've shown in this slide. The one that we are particularly concerned in is machine learning, which is defined as the ability, the field of computer science that allows computers to learn without being explicitly instructed on what to do, which is very powerful. And two tools which are most commonly used is regression analysis and endocrinology. We use FRAX tool, full cohort equation. So we are effectively using routinely regression machine learning in clinical decision making. What AI is mostly referred to as deep learning. This is neural networks and all those big large language models and computer vision and self-driving cars. And we'll use both to help diagnose tarot cancer. So what we did, we used the traditional supervised machine learning approach. We developed convolutional neural network and trained it on about 32,000 tarot nodal images, and then we assessed it using cross-validation, which is a machine learning technique which makes sure that the performance is true and not just because the neural network memorized those images. And we also tested on a separate image set, which was derived from a different health care system. And we found that it works reasonably well. We have a good sensitivity, don't want to miss cancer. Specificity was modest, but okay. And the area of receiver, and the receiver operating characteristic curves about the same as experienced radiology would achieve in a research study. So what does it mean? We're good to go. We can now just start, you know, deploy the system and start using it. As Raul cautioned us, we should be careful before we pull the trigger and assess AI tool more closely to make sure it's actually have a value added, and it's going to help us patient and not hurt it. So we started looking at how the decisions are actually made. And those colorful heat maps, it's one of the methods of interpretable AI. That's called saliency heat maps. And basically what we are highlighting, the area of the nodule which was used by a computer model to make a decision whether it's cancerous or benign. And then first we look at nodules which are confidently classified by artificial intelligence as cancerous or benign. And we found that the malignant nodules with very high probability of malignancy by AI were those classic populated cancers, which are hypoechoic with microcalcifications, irregular borders. And that's what the model picked up when it assigned that high probability of malignancy. So that was reassuring. And the same thing with benign nodules. We saw a lot of sponge-form nodules, purely cystic nodules, those that we don't worry in practice and don't need to biopsy many of them. That was good. What we also found that the AI used some features which we don't use as human providers, such for example as small cysts. And posterior enhancement was one of the cool ones. The posterior enhancement, it's an artifact from the ultrasound, which occurs when the sound waves go through a benign toroid nodule, which is transparent for the sound waves. And then after going through it, it hits the normal toroid tissue and gets reflected back to the probe. And that's why we see this brighter area posterior to the nodule. And it's difficult for radiologists to reliably assess it. That's why we don't use it in clinical risk stratification schemas. But it was not a problem for AI, and it learned it with one of the features. So the bottom line here is there is a benefit in machine learning which does not restrict the computer to specific features which are designed for the human use. The AI is potentially more powerful in assessing the risk. And this slide shows yet another feature which was picked up by the computer, which is anterior invasive growth, an extension of the nodule into subcutaneous tissues, which is an indication of invasive growth and malignancy, of course. Again, it's difficult for radiologists to identify it. In our highly specialized clinic, we use it, but there's a lot of guessing happening in here. But the computer was able to pick up many, which were later confirmed on histopathology after toroidectomy. Well, the system was not perfect when we started doing a phalenmon analysis and started looking at the nodules which were felt negative. So they are called nodules which are felt negative. So they are called as low risk by the computer, but we know that they are cancerous based on the results from the surgery. Here is a few examples. As you can see, those nodules, they are not obviously concerning for us. They don't have those classic suspicious features to it. We also found that most of the nodules which are felt negatives, they belong not to classic papillary thyroid cancer, but to a more difficult to diagnose subtypes of thyroid cancer, such as follicular variant of papillary thyroid cancer, follicular thyroid cancers, and oncocytic thyroid cancers, which makes sense. We know that those cancers, they are difficult to diagnose even when we have a tissue sample such as from biopsy. From biopsy. And it's even harder to do it based on the radiology images. So we call it the impossible problem of thyroid cancer diagnosis. We know that some thyroid cancers, follicular thyroid cancers, follicular variants of papillary and oncocytic thyroid cancers, they frequently don't have sonographic risk features, which are obvious to the human eye. So those are tarot's three category nodules, or American Thyroid Association low-risk nodules. And most of the nodules overall, benign and malignant, belong to this category. So not to miss these cancers, we are forced to biopsy many of these low-risk thyroid nodules just to capture a few thyroid cancers. And that's why we have this outcome, which I showed on one of the previous slides, where we only have 8% of biopsies resulting in malignant serologic diagnosis, and most of the biopsies produce benign. And I'd argue that further adjusting clinical risk stratification algorithms, which are relying on features which are recognizable to the human readers, will only produce maybe a small improvement, but will not be a decisive improvement, which we would like to have in this field. And that's what we hope that the AI is going to bring, this decisive improvement, using possibly some features which we can't use because of the limitations of our brain. And Caitlin Bell is one of my collaborators who is working on solving this impossible problem by collecting lots of this hard-to-diagnose images from follicular adenomas and follicular carcinomas. And then we thought, well, let's test our model and see how it performs on this. And as expected, it did not perform well, better than the random chance. So we got the accuracy of 72% for our model, which we call virtual thyroid biopsy. But the sensitivity was poor, and the area of the receiver operating characteristic curve decreased significantly from when we look at the all nodules overall. We are not sure if this is just a deficiency of our tool or if it's a problem overall in this AI tools. So we took two others which were available to us through collaborations online, AI Thyroid and AI BX, and they performed even worse. So at this stage, we don't have a reliable AI tool which would diagnose these follicular and oncocytic thyroid cancers with the degree of accuracy that we want for implementation into clinical practice. And one way to deal with it is what Caitlin is doing, just by collecting a large bulk of images, and then we will fine-tune our model to recognize those cancers better. I think it will work to some degree. But because those nodules don't have, may not have suspicious features at all, which would distinguish them from benign or anomalous, there is an inherent limitation of this approach. And that's when we thought, okay, well, if we cannot do a perfect diagnosis, so good enough diagnosis from the images only, why don't we bring another risk assessments to our decision making here? And that's when we thought about thyroid cancer as a genetic disease. And when I was in training 12 years ago, and many patients asked me, well, why I developed thyroid cancer, I eat broccoli and I don't smoke, I basically said, I don't know. But right now, we have pretty good understanding that at least half of that risk is coming from parents, it's inherited genetically. And if we could calculate that risk, and combine it with our image based assessment, plausibly, we would have a much better system to decide which nodules require biopsy, or which don't, and we can just watch them. So this is a complicated figure. But basically, on the x axis, it shows allele frequency, so how specific genetic variants are common in the population. And on the y axis is the effect size of how, if you have that mutation or variant, how likely you are to develop a thyroid cancer. And in the left upper corner, we have this thyroid cancer associated syndromes. So those are rare variants, such as multiple endocrine neoplasia, or pitta and hamartoma syndromes, which we know increasing our risk of developing thyroid cancer. And my argument is that if we detect this on the genetic testing, that would indicate that this nodules in this patients needs to be biopsied because they are much more likely to be cancers. So we thought, okay, well, we have now large population based data sets where we could look for these mutations, these variants, and just to see how common they are. And the United States now runs this very large population based biobank, which is called All of Us, which at this time has about quarter million of whole genome sequences in there. So that's a massive amount of genetic data linked to the clinical data, such as data extracted from electronic health records, which is a real treasure data set for questions like ours. So we queried it. It's a massive data set, as I've mentioned. So we have to use some distributed computing tools and Google Cloud computing resources. And we only looked at known pathogenic or likely pathogenic variants, which are deposited in the ClinVar database. And then we applied machine learning to see which genes and which syndromes are associated with increased thyroid cancer risk. And then we tested a lot of genetic syndromes, adjusted for multiple comparisons, so did all the statistics properly. And the results are quite surprising, because we found that the population prevalence of those genetic syndromes, which makes it much more likely to develop thyroid cancer, is actually quite high. In the literature, for example, the population prevalence of multiple endocrine neoplasia type 2A, so those are mutations in the RET gene, is estimated as 1 in 30,000 to 1 in 50,000. And we found that 1 in 2,200 enrollees, participants, and all of us have that mutation. So they, by definition, have multiple endocrine neoplasia type 2A. So initial thought was there must be something wrong in this data, or maybe we are calling some variants which are not pathogenic. So we restricted our variants just to those which are described as pathogenic, and they are actionable in 2015 medlery thyroid cancer guidelines. And we still get a quite high prevalence, 1 in 2,700 overall population of the United States who have this genetic syndrome. Fortunately, most of those mutations were moderate risk, and most common mutations was RET V804M. But what surprised us only that a small, only a small percentage of these patients with this syndrome already diagnosed with a thyroid cancer. So we have a large pool of people walking out there who don't know that they are at risk of developing medlery thyroid cancer, because the system is just not designed to detect them. And we only see them when we do this kind of a population-based blanket screening by just genotyping everybody, effectively. And then the second syndrome increasing the risk of thyroid cancer is Peter and Hammer-Thomas syndrome. And we found the same thing, that the true prevalence of the population is about 1 in 10,000. Published prevalence is 1 in 200,000, so it's very much underestimated. And we only have 1 in 5 people who have been diagnosed with thyroid cancer. So the other 80%, they probably need an ultrasound and a biopsy if thyroid nodule is found. These are just some other interesting facts which we found in this study. I'll just focus on the last two bullet points. For those of us who practice medicine and evaluate people with thyroid cancer, 1 in 70 patients with thyroid cancer will have a genetic thyroid cancer. So they will have mutations in RAD, PTEN, APC, or DNA mismatch repair deficit genes. And 1 in 39, if you send them for a genetic testing, you'll have at least one actionable mutation for cancer-related thyroid syndrome, which means you find it, you might save potentially somebody's life. So here's the conclusion for this section of the talk. We found that thyroid cancer-associated syndromes are much more common than previously thought. And it is important to collect family history and refer patients at risk for genetic testing and genetic counseling. It could be very impactful. And by looking and identifying these people with genetic high-risk syndromes, we can improve our decision to proceed with a biopsy because these people need biopsy unquestionably. Okay, so now let's move to a different type of variants which are common, which have a much higher population prevalence, those that are on the right bottom part of this figure. So they are common, and those of us who carry this mutation, the risk of cancer includes only a little bit. The problem comes when you have multiple of these variants contributing this little risk. And in the end, if you're unlucky, you might have a combination which increases the risk of thyroid cancer quite substantially. This is useful for us to decide which nodules to biopsy and to calculate the inherited risk for thyroid cancer. So we looked into this, and genomics is very big, and identifying these common variances is not a simple task, requiring a lot of data. So we built this virtual thyroid biopsy consortium, which includes 14 biobanks, at this time from many countries in the world, and collected data for actually five thyroid diseases, not just thyroid cancer, but also benign node locator and the others, and looked at 1.8 million genome-wide genotypes, and we applied machine learning to detect the sites in the genome which are associated with increased risk. So this is what geneticists call Manhattan plot. You can see those peaks. Those peaks actually consist of dots, and each dot represents a variant, which increases your risk of developing a thyroid cancer. And those peaks, they resemble skyscrapers in Manhattan and New York, so that's why they call it Manhattan plots. And the variants which are shown here in red, these are newly discovered variants which were previously unknown to increase the risk, but because we looked at such a high, large dataset, we were able to discover them. And what's plausibly, many of these germline variants, they're actually in the genes which we know increasing the aggressiveness of thyroid cancer or leading to the development of thyroid cancer when you have a somatic mutation, when you develop that mutation just in the tumor, like p53 and ATM mutations. And we have tert and POT1 genes which are increasing your risk of cancer by lengthening the telomeres and effectively immortalizing the thyroid cells so that they turn into cancer. And then, because we also collected the same data and performed the same analysis of benign non-maleguators, we thought, well, this gives us an opportunity to compare genetic predisposition to cancer to genetic predisposition to a benign disease and then find what makes, what the differences are in these two genetic risks. And then we can use it to decide what type of thyroid nodule the patient is more likely to develop. So this is illustrated on this, what geneticists call the Miami plot. So at the top, we see the Manhattan plot for thyroid cancer, so those skyscrapers indicating genes which increases your risk for thyroid cancer. And then think of it as Miami skyline reflected in the waters of the Atlantic Ocean. On the bottom, it's the same thing for benign goiter. And that type of visualization is used to compare genetic risk for two diseases. And you can see that some of the peaks, they are reflected quite perfectly, but the others are unique to either thyroid cancer or benign nodular goiter. And that's what we are most interested in because it allows us first to think of the mechanisms of why some people develop cancer and the others develop benign disease. But it also allows us to make predictions and say you are at a greater risk of developing cancer, so we biopsy you. And somebody, the other patients may say, well, you are at a great risk of developing benign nodular goiter, but not so much cancer, so maybe we can just watch you and improve overall how we manage those thyroid nodules in clinic. So this is what drives our future research. It's a hypothesis at this time, so I don't really know if this truly works as it is shown on this figure. But what we believe is because of the genetic differences, we could use this genetic predisposition to effectively predict of what type of thyroid nodule that patient is going to develop depending on what kind of genes they have affected. And those genes shown in black on the very right, we call them benign genes. They will lead to those non-neoplastic, hyperplastic, and adenomatoid nodules. And I would argue that those nodules, we can safely ignore them. They are not even tumors. They will never turn into cancer. Those in the middle, these are the genes which we see both in benign nodular goiter people and also in thyroid cancer people. So we think those will cause thyroid neoplasms. But those could be adenomas, so some of them could be benign, and the others could be just a low-risk thyroid cancer and maybe even precancerous lesions like NIFT-P. And those lesions maybe can be monitored with active surveillance instead of going straight for the surgery. And then we have genes which are unique to thyroid cancer. And those potentially could lead to the aggressive disease. And this is where we need to send them to the endocrine surgeons to get it removed. So if it truly works like this, this is a hypothesis, but I think a rather plausible hypothesis, then effectively we will have a thyroid disease fortune teller because the genetic architecture of individual is defined when the egg gets fertilized by a sperm cell, so very early and at birth when we can genotype that person and determine what kind of thyroid future we should expect for that particular individual. So effectively we'll be predicting the future but using science instead of magic. So that's what drives me to further inquiring to this question because I think it's pretty cool. Okay, well, going back to our use case, remember all this big trouble of building a consortium and collecting the data for almost two million people was for the one reason, because we want to estimate the risk for thyroid cancer and use it in combination with our image-based classifier and see if we can see an improvement. And this is what this slide shows. So on this area, the receiver operator characteristic curves, the blue curve is just an AI. So it's a neural network which looks at the ultrasound images. And in the red, when we combined AI with genetic risk assessment, we see some improvement. And when we add another covariance such as genetic ancestry, then we see even more improvement. As we built up our genetic research, and the goal is to expand it to three million people in the next two years, this added improvement in the performance is going to grow. And we believe at some point we'll have a tool which will perform much, much better than what we can do as clinical providers just interpreting those ultrasounds ourselves. Well, that's one of the final slides I have in my talk. And that's another interesting and potentially practice-changing finding that we got from our genetic analysis. I've mentioned that we did not just look at the thyroid cancer. We also looked at multiple other thyroid problems. And the question we thought, well, does the presence of the other thyroid disease affect your risk of cancer? For example, if you have Graves' disease, are you more likely to develop thyroid cancer? Which also can help us both in deciding how to manage those Graves' disease patients, but also in assessing if a thyroid nodule is cancerous or benign in somebody who also has a concomitant Graves' disease. And one way to answer that question using this bulk genetic data that we have is to do a genetic correlation analysis. Basically, this table shows if there is a correlation between genetic risk for developing one thyroid disease versus genetic risk of developing another thyroid disease. And here on this table, you see that the correlation between hypothyroidism and lymphocytic thyroiditis, when we did this genetic correlation analysis, is almost perfect, which is plausible because lymphocytic thyroiditis is the main, most common reason for hypothyroidism. We also can see that the correlation between a risk of thyroid cancer and benign oligoidar is about 50%, which also makes sense because we see that there is an overlap in those Miami plots between our thyroid cancer associations and our benign oligoidar associations. But there are also variants which are unique to one of those conditions. So the joint risk is partial. It's not complete as with hypothyroidism's lymphocytic thyroiditis case. We did not see any correlation of hypothyroidism with thyroid cancer or benign oligoidar. And there is a lot of literature out there that say if you have Hashimoto thyroiditis, you're at a higher risk of developing thyroid cancer. And the other studies show there's no association. Well, we clearly see that there is no association. So if you have a hypothyroid patient with lymphocytic thyroiditis, they are not at a high risk of developing thyroid cancer. And that is different for Graves' disease, where we see a quite strong association. It's about 30% risk is shared between Graves' disease and thyroid cancer and also benign oligoidar, which is mechanistically plausible because we know that activation of TSH receptor by TSH can make thyroid cancer grow. That's why we suppress it in people with high-risk thyroid cancer. We suppress the TSH. And plausibly, it works the same in Graves'. And the current practice is we don't ultrasound thyroid glands for people with Graves' disease because we think of it as a separate disease. They're not immune disease. But I think this data shows that those individuals are at increased risk of developing thyroid cancer, and maybe they deserve at least once to have an ultrasound. And if they have thyroid nodules, maybe they deserve to be biopsied. Okay, so closing on this presentation, this is the genetic analysis summary. We found that all thyroid diseases, including thyroid cancers and benign oligoidar, they have strong genetic predisposition, that patients inherited that risk from their parents. We also hypothesize that genetic architecture, genetic risk, defines what type of thyroid nodule the patient will develop. Could be non-neoplastic hyperplasia, could be benign tumor or low-risk thyroid cancer, or could be aggressive thyroid cancer. We could use that information to calculate the genetic risk, and that will improve the risk certification of thyroid nodules in clinic. And the final somewhat incidental finding is the Graves' disease, but not hypothyroidism, increases the risk of thyroid cancer, and we might consider doing an ultrasound in these patients. Now, there's a lot of data in the stock, and it's a large group of people from different specialties which contributed to it, which I want to acknowledge on this site. And especially the genetic side of the project would not be possible without several high-impact initiatives. One is the Global Biobank Meta-Analysis Initiative. That's how we were able to connect to all those biobanks contributing data to our study. And the All of Us participants who volunteered to donate their genetic data to the All of Us research program, which made it possible our study of syndromic thyroid cancer. And we are also thankful to Regeneron Genetic Center and Colorado Center for Personalized Medicine to help with genotyping and the data access. And I have also listed some of my finding sources. And with this, I'm going to stop talking and happy to answer questions that you might have. I don't think we can hear you, it might be on mute. Dr. Potseyev, thank you very much for this very enlightening presentation. So I'm Marcelo Correia from the University of Iowa, where I practice endocrinology and bariatric medicine. So the session is now open to questions. And I think I have Maria asking the first question. And the question is, in a real world setting, I expect the cost of AI nodules analysis plus gene study to be much more expensive than FNA. So is this true? Well, the cost question is a good one, because yes, if we start doing all those complex tests on everybody, are we going to drive the healthcare cost through the roof? My argument here though, is if we do the FNA, which is also not cheap, it's a couple of thousand dollars in the United States, we are running the risk of needing an additional testing, which would be molecular testing, which also expensive. And then potentially doing a diagnostic or therapeutic surgery, which is very expensive. So the artificial intelligence development is expensive. Running it on the images is cheap. The genetic part, if we have a combined system, in our setting, at least, it comes for free, because everybody who volunteers to participate in the biobank in our healthcare system gets genotyped. And all we need is just to use that data to develop a risk assessment. So it's a good question. I don't have a definitive answer. It depends on how that system is going to get implemented and how the patients or insurance companies will get built. But I think it has a pretty decent chance of being a better choice than doing a diagnostic surgery. Amira is asking, I think genetic stratification means that we need to do genetic testing, and the result of which can take some time. How can we address this issue? Yeah, if I order the genetic test at a time I see the patient, at a time I see the patient, it usually takes time. The area is rapidly developing, and we are finding greater and greater users for this genome-wide or whole genome sequencing data. What I'm predicting, in the near future, we will have a genetic passport as a test. So that's a test that's done at birth or at some point in life. Which we use as needed when we treat our patients. Because in contrast to the TSH or the other diagnostic test, or even ultrasound, your genetic data does not change through life. The way you're born with it, that's how it's going to be at death. So we only need to do it once. And once that data is available to us, we could use it for thyroid cancer, but also for many, many other clinical users. And once we learn how to use it effectively, that's what will make this whole genotyping practice cheap and cost-effective and impactful. I presume the Manhattan plots were based on blood tests, right, not nodule biopsies, correct? Okay. Because we are looking at the inherited risk, we don't want to detect those somatic mutations within the nodule. So that's blood or saliva. Oh, saliva. That's pretty practical. So Lizelle is asking, should we be doing genetic analysis on all patients with thyroid nodules, considering the use of GLP-1 receptor agonists? Oh, yeah. So the GLP-1 story, keep going. It started when they were developing the early GLP-1 agonists. In the animal studies, they saw C-cell hyperplasia in rats, not in human. They only saw it in animals. But out of caution, there was a black box warning, and I think there is still a black box warning on these drugs, suggested that it can increase your risk of medullary thyroid cancer. And then there was an additional concern that if it causes medullary, would it be also increasing your risk of differentiated thyroid cancer, which is a different type of cancer. It originates from a different type of cells in the thyroid. So if you have somebody with a family history of medullary thyroid cancer, my answer is yes. I think it might be worse of assessing their thyroid or possibly even genotyping them to see if they are at risk of developing medullary thyroid cancer. Those with differentiated thyroid cancer, such as papillary, follicular, and non-cathetic, there is no concern with GLP-1 agonists. And the new guidelines will specifically state that, that if you have follicular cell-derived cancer, you don't worry, you don't need to do anything before you start them on GLP-1 agonists. So Yusa is asking, we need to know the prevalence of thyroid cancer on patients with or without GLP-1 treatment. I think you partly addressed this question already. I think it's probably yes, right? Well, again, I don't think the data is strong enough to implement population screening for MEM, for multiple anachroneoplasia and medullary thyroid cancer, before you start summoning a GLP-1 agonist. But if you're concerned because they're saying, well, my dad had some kind of weird thyroid cancer and somebody said it could be medullary, for sure, I suggest that you investigate it before you start the medicine. Yeah. What you have found changes... Okay, so let me understand this better here. So what you have found changes a paradigm of only medullary thyroid cancer being genetic. Will guidelines change regarding familial papillary thyroid cancer? And there's an observation here, more than two first degree relative with thyroid cancer to consider the case to be genetic. Basically, the question is medullary thyroid cancer the concerns and familial papillary thyroid cancer in the context of genetic traits. Yeah, thanks, Marcelo. Those are great questions. Our research on syndromic thyroid cancer will be mostly changing management for medullary thyroid cancer, because that's where the dramatic increase in prevalence that we've discovered. I think we might also consider changing our management for some of the types of the red mutations like that one, V804M that I've showed. Most of those peoples, they are not diagnosed with thyroid cancer, at least in the all of us population, suggesting that's probably it's a low risk or low penetrance mutation. Maybe we don't need to do prophylactic thyroidectomy on all of those. And that's not an official recommendation, that's still an area of active research, but we'll be publishing this. The differentiated thyroid cancer, yes, there will be changes in the guidelines. And again, I'm not part of those guidelines team, but I was fortunate to listen to presenters from that team at the recent conferences. And there will be changes on germline genetic testing, even for follicular cell derived thyroid cancers. For example, if you diagnose on surgical histopathology, CREB4, Morula, some type of thyroid cancer, then you better test the PTEN gene because that's some type of cancer is frequently associated with either Hamartoma syndrome or Gartner syndrome. And it's not thyroid cancer which is going to put those people in trouble, it's a colon cancer. But the thyroid cancer diagnosis will provide an insight that they're at risk of developing the other malignancies. And there are others. There's a lot of discussion going about Dyser 1 syndrome. I'm not really sure we need to screen people for Dyser 1 syndrome because it's a very rare one and it's mostly because of benign onlookers, not so much malignancy. So yes, there will be changes. Again, saying that we should screen everybody for cancer-associated syndrome who have public thyroid cancer, no. But collect the family history. It might change how you approach the patient. The user is asking, thyroid nodules or goiter prevalence appear to be high in patients living with end-stage kidney disease. For ESKD patients sign up for kidney transplant, we were asked to evaluate or rule out a thyroid cancer. What should we do? A very common scenario in our practice as well because obviously the transplant folks don't want to transplant kidney into somebody who has an active thyroid cancer. So yes, we are asked a lot to evaluate thyroid for possibility of thyroid cancer. I'm not aware of any strong data suggesting that we should manage those people differently. We usually just use a traditional approach when we do ultrasound-based risk assessment. And if there is a high-risk nodule, we biopsy that nodule and then hope to reassure the transplant team that the risk of cancer is low. It gets more complicated if you detect thyroid cancer in these people because it disqualifies them from a kidney transplant frequently. And then what we do, we try to operate on them if the overall surgical risk is manageable and then hope that this will be a low-risk tumor and then we reassure the transplant team that the chances that person is going to get in trouble from thyroid cancer is minuscule and that should not affect on their management of end-stage kidney disease. So I think we have a couple more minutes and I have a question. So you covered imaging and genetic profile as information, food for AI thought. Are there other factors like geolocalization, geolocalization, exposure to environmental factors, pollution, things that can be actually measured with all the technology, portability of technology we have today, cell phones and all that. Things like we learned to do during the pandemic, exposure to COVID in certain communities at certain areas, right? With the alerts to the people potentially exposed. So is there any other type of information we can add to this models to basically make a perfect prediction of the risk of a nodule being malignant and again, not a bad nodule, right? Those that carry a risk of aggressiveness. Yes, Marcelo, very valid point. If you look at the overall model of the risk of developing thyroid cancer, half of it's genetic, but the other half, we don't really have a good way of calculating that. And that could be environmental factors or it could be just the bad luck. Like somebody developed the mutation because of the bad luck and it propagates. There is absolutely data on environmental risk in thyroid cancer. The strongest is radiation exposure. So if somebody's coming from Ukraine and was exposed to the fallout from the Chernobyl incident or same thing in Japan and Fukushima, then they're at a greater risk of developing thyroid cancer. There's a lot of research on endocrine disruptive chemicals. My colleague, Dr. Goldner is researching, for example, whether the exposure to pesticides used in farming could be a risk factor. And she has some data suggesting that it is. And if we find that that association is strong enough, yes, we could use geospatial data, for example. And if you live in a zip code where a particular highly pesticides has been used extensively, then we could factor it in in our clinical decision model. It's not going to be easy. It's a very challenging task. Possibly that's the future now. Great. So thank you very much, Dr. Posnayev. Thank you very much for your groundbreaking research. I think this session now adjourned. Thank you very much for the audience as well.

thyroid cancer

endocrinology

thyroid nodules

THARADS

artificial intelligence

genetic risk assessment

genomic data

personalized medicine

diagnostic accuracy

patient care