So, I'm Maria-Christine Asenaro from the INSERM and University Paris Cité in the Genetics Department of the European Hospital Georges Pompidou, and this is in Paris in France. And I will, I'm really happy and thank the organizers for inviting me here to present some of our data on how we develop multi-omics biomarkers to diagnose endocrine hypertension. So, I have no conflict of interest to declare. So, before starting, actually, I need to make a little bit of a point here as we are in the endocrine society. So, why are we talking about arterial hypertension? What is the context? What is the background? And what are the requirements? So, you all know that arterial hypertension affects 40 to 50 percent of the population over 40 years, and it is the most important risk factor for stroke, heart disease, and kidney disease, and it is responsible for more than 10 million deaths per year. And although a large therapeutic arsenal exists targeting different systems, blood pressure is uncontrolled in up to two-thirds of patients. So, detection of secondary forms of hypertension is key to targeted management of the underlying disease and prevention of cardiovascular complications. And it is in this context that endocrine forms of hypertension represent major targets for stratified approaches of hypertension. And among the endocrine forms of hypertension, primary aldosteronism is the most common form of secondary hypertension, with a prevalence up to 25 percent in patients with hypertension resistant to treatment. So, what would be the diagnosis and the treatment of hypertension in an ideal world? In an ideal world, a hypertensive patient, or just a patient, whatever, would go to his GP, and there, if high blood pressure is measured, the GP would be able to distinguish between primary hypertension, what is also called essential hypertension, and secondary hypertension. And if secondary hypertension is identified, the GP would be able to address the patient to a referral center for diagnosis and management in order to get a targeted treatment. And in the case of primary hypertension, the GP would be able just to treat the patient with the appropriate treatment directly. But we all know that this is absolutely not what happens in real life. And so, there is still a lot of space for improved diagnosis and improved treatment options in hypertensive patients. So, detection of endocrine forms of hypertension is key for targeted management and prevention of cardiovascular and metabolic complications. So, when I speak today about endocrine forms of hypertension, I would rather focus on adrenal forms of hypertension. So, secondary forms of hypertension, which are due to excessive and autonomous production of adrenal hormones leading to high blood pressure, which are primary understerilism, pheochromocytoma and functional paraganglioma, and Cushing syndrome, which I will abbreviate PA, PPGL, and Cushing CS from now on. And so, those patients, we know, are at increased cardiovascular and metabolic risk, and they also have a diminished quality of life. And this is because of the diagnostic complexity, which delays the treatment by several years after onset of hypertension. And so, those patients are at increased risk of renal, cardiovascular, and metabolic complications. And they often have a lifelong to time antihypertensive treatment, and so they have a reduced quality of life. So, primary understerilism is not rare at all. It is the most frequent form of secondary hypertension. So, these are data from the literature, which have been replicated in many studies. This is the prevalence of primary understerilism in primary care in a recent study showing that in primary care practice, patients, 6% of patients with hypertension actually have primary understerilism. And in most cases, in these studies, the forms were bilateral rather than unilateral. And the prevalence of primary understerilism increases with the severity of hypertension, reaching approximately 20% in patients resistant to treatment. So, those are data from our hypertension department at the European Hospital Georges Pompidou in Paris, provided by Michel Azizi and Laurence Amal. And you can see those are data over two years, where more than 3,500 patients have been included. And 27% of them have secondary hypertension. Among them, 15% have endocrine hypertension, 11% primary understerilism, 3% PPGA, and 1% Cushing's syndrome. And this actually surge of endocrine hypertension is even more important in young patients. And actually, this is again, data from our hypertension department performed on more than 2,000 patients with confirmed hypertension aged 18 to 40 years. And actually, 30% of them have secondary hypertension. And among them, 50% have primary understerilism, which makes a prevalence among these patients of primary understerilism, which is 16%. So, primary understerilism is associated with an increased cardiovascular risk, which goes far beyond that of just blood pressure. And this is whether the risk is adjusted or unadjusted to the duration of hypertension. Again, a study, approximately 450 patients with primary understerilism, more than 1,200 controls with essential hypertension matched for sex, age, and blood pressure. And the blue ones are adjusted for the duration of hypertension. And you can see here the increased risk of primary understerilism to develop any coronary heart disease, nonfatal myocardial infarction, heart failure, or atrial fibrillation. So, what is primary understerilism? Primary understerilism has been discovered in the early 1950s by Jerome Kohn and described as the overproduction of aldosterone leading to hypertension with hypokalemia due to an aldosterone-secreting tumor or associated with suppressed reading. So, nowadays, the Endocrine Society guidelines have defined primary understerilism rather as a group of disorders in which aldosterone production is inappropriately high, relatively autonomous from the renin-angiotensin system, and non-suppressible by sodium loading. And it is due, in the majority of cases, to unilateral aldosterone-producing adenoma, APA, or ibilateral adrenal hyperplasia, BAH. And primary understerilism may induce severe, resistant, or complicated hypertension. It may have severe cardiovascular consequences, and this is due to the specific effects of aldosterone on target tissues, but it is amenable to specific surgical or drug treatment. So, when a patient is at risk for primary understerilism, it should undergo primary understerilism screening, subtype identification, and targeted treatment. So, which categories of patients are at increased risk? In 2008 and 2016, the Endocrine Society has released the guidelines for the management of primary understerilism and defined those categories of patients. So, those are patients with sustained hypertension, with resistant hypertension, with hypertension associated with hypokalemia, hypertension and adrenal incidentaloma, hypertension and sleep apnea, hypertension and a family history of early onset hypertension or cardiovascular accidents before age 40, and all hypertensives first-degree relatives of primary understerilism. So, those patients would undergo screening for primary understerilism, which is performed by measuring the aldosterone-to-renin ratio to detect the cases, and then they would go confirmatory testing and subtyping. So, confirmatory testing would be performed by any of the confirmatory tests, which is an infusion test, the most frequent one, or flutocortisone suppression test. Then the patients would undergo adrenal CT scanning to see whether there are nodules present in the adrenal, and if surgery is desired, and the patients would undergo adrenal vein sampling to identify unilateral form, which then will be treated with laparoscopic adrenalectomy or bilateral form treated with mineralocorticoid receptor antagonists. And if there is no desire for surgery, patients could just be treated with mineralocorticoid receptor antagonists, which are extremely efficient. So, there is a real, a little right part here, which is some exceptional cases where one can undergo directly to treatment with MRA or directly to laparoscopic adrenalectomy. So, despite the publication of these guidelines, there still remains some diagnostic and therapeutic challenges in primary allosterianism. First of all, because of this diagnostic complexity. So, the complexity of the workup, which I have just presented, but also because there is only weak consensus regarding biochemical methods, the thresholds for screening, the confirmatory testing, subtype differentiation. And so, less than 2% of high-risk patients with treatment-resistant hypertension are tested for primary allosterianism and much less than 1% are ever diagnosed or treated with mineralocorticoid receptor antagonists. And as I said, those patients remain at increased cardiovascular and metabolic risk and diminished quality of life. And so, there is really a need here to implement new biomarkers directed at patients with primary allosterianism to establish eventually surrogate biomarkers from the underlying genetic causes, because the genetics of the disease is now known, and to implement targeted treatments. What is also complicating the whole situation is that there is some evidence for common, not only pathogenic mechanisms, but also commonalities between unilateral allosterian producing adenoma and bilateral adrenal hyperplasia. So, adrenals from healthy individuals, they show what we call, and you can see here on the left, they showed aldosterone-producing cell clusters or aldosterone-producing micronodules, that's the same, which actually carry somatic mutations in genes that are also involved in primary allosterianism. And the number of these micronodules increases with age, and this is paralleled by an increase in autonomous allosterian production. So, this is normal population. And actually, the number of this APCCs or APN is increased in adrenals from patients with bilateral adrenal hyperplasia, indicating that there might be some common pathogenesis to unilateral and bilateral forms of adrenal hyperplasia, of adrenal, of primary allosterianism. Also, adrenals from patients not cured after adrenolectomy show bona fide allosterian-producing adenoma, carrying somatic mutation in the majority of cases. And so, that defines a category which we call bilateral adrenal hyperplasia with asymmetrical allosterian production. And finally, recently, different teams, and not only ours, but other teams too, have identified genetic risk loci for primary allosterianism, which increase the susceptibility of developing the disease in the general population, and which are in part shared between allosterian-producing adenoma and bilateral adrenal hyperplasia. On top of this, there is also evidence for a continuum between primary allosterianism and hypertension in the general population. So, there is a strong evidence for inappropriate allosterian production, playing a role in a larger subset of patients with primary hypertension. And we know, all of us, that the aldosterone to renin ratio, or even aldosterone, are continuously distributed. There are patients with hypertension who actually have low renin hypertension, but not yet primary allosterianism. And in population studies, particularly in the Framingham studies, but other populations, it has been shown that subjects with normal allosterone levels, but in the highest quartile of the normal distribution, have an increased risk of developing incident hypertension or increasing their blood pressure over time. Furthermore, up to 22% of patients with primary hypertension have abnormal allosterone regulation, and the percentage of these patients rises with the severity of hypertension. So, there is this notion of a continuum of aldosterone dysregulation throughout the spectrum of blood pressure. And recently, risk loci that have been identified for primary allosterianism in genome-wide association studies have shown that these loci are shared between unilateral and bilateral primary allosterianism, but they partially overlap with risk loci for blood pressure and resistant hypertension. So, to summarize these last two or three slides, which are quite tricky, but really represent a paradigm shift in our understanding of what is primary allosterianism and why it is so difficult to diagnose, is the question here, how do we improve diagnosis of primary allosterianism in the continuum of dysregulated allosterone production? And I always borrow this slide from a very nice paper by Jennifer Brown here, where she depicts actually the blood pressure stages here from normal blood pressure to resistant or severe hypertension. Here, the biochemical stages of non-suppressible and really independent allosterone production, where even within the normal blood pressure category, there might be a certain degree of dysregulated allosterone production, which then goes further right to biochemical overt primary allosterianism. And this, of course, is associated with increasing cardiovascular and renal risk. And we have also discovered the genetic underpinnings of all of this with risk alleles, which can modify adrenocortical function and be involved in increasing the susceptibility of moving from this place here to this category here. So here is all the problem in developing better biomarkers for the detection of primary aldosteronism. And this is because we are not developing biomarkers or diagnostic procedures for a disease, but we have to develop biomarkers for a continuum. So how can we improve this diagnosis? So we have actually tried to approach this question within two consecutive European programs. These are called Horizon projects, Horizon EU programs. So the first one was NSAP-HT, which was aiming to develop and evaluate an omics-based stratified health promotion program for patients with endocrine forms of hypertension. And this program is now finished, and it has been followed up by a new one, which is HD-Advanced, whose aim is to improve treatment efficacy in hypertension by biomarker-guided personalized decision support. So the aim of those projects is to use Momics biomarkers. So Momics is multiomics biomarkers for improved diagnosis and treatment of hypertension. So I have set here the design of those studies and their interconnection. So in NSAP-HT, we have performed two different studies, a retrospective study and a prospective study. And with this, these studies were implemented within the NSAT network. So this is the European network for the study of adrenal tumors, and then within the NSAP-HT registry. And here in NSAP-HT, we have developed multiomics biomarkers for patients with primary aldosteronism, PPGL, Cushing syndrome, but also able to distinguish patients with primary hypertension. And then we have also used healthy volunteers, but I will describe this better. And we have measured different sets of biomarkers, which then were integrated through machine learning to generate multiomic signatures, which then have been retrained and validated on a prospective study. And then the aim of the next program was to validate those biomarkers within a randomized control trial. And that randomized control trial here is the NSAP-HT endotrial, which is then associated with two other studies, which I will explain later. So what is NSAP-HT? NSAP-HT is aiming to apply omics-based strategies for improved diagnosis and treatment of endocrine hypertension. And actually, we aimed in an exploratory phase to identify omics signatures of endocrine hypertension by measuring high throughput genomics, metabolomics, and phenomics characteristics and integrate them through bioinformatic modeling to generate biomarkers, which then would be validated as stratification biomarkers and then applied in a prospective phase of the study to the stratification of hypertensive patients and to measure diagnostic performances and a certain number of outcomes. And the whole program was evaluated for its validity, for stratifying patients, for improved diagnosis, treatment, and outcome. So this was a very large European program, and those are the partners. So we were quite several partners in France, in the UK, in the Netherlands, in Germany, in Italy, and also in Australia, and all patients. So the partners had different specificities and all contributed towards this project. So how did we proceed? So starting from the European Network for the Study of Adrenal Tumor Registry, which was a retrospective registry, we retrieved biosamples from 487 patients. So those were patients with primary aldo, PPGL, Cushing's with primary hypertension, and we had also in parallel assessed two peripheral volunteers. So again, as I told before, we measured different omics, which I will detail later, and then these different omics went into a machine learning pipeline to generate multi-omic signature. And then the idea here was actually just to have a very simple test, but very performance simple test to be able to identify each category of patients. So which omics and why omics? So of course, if we start from the biological molecules that can be measured either in plasma or in urines, we would, there is DNA, and then there is RNA, and then there are proteins, and there are metabolites. And if we look at the omics cascade, which is the collective characterization and quantifications of pools of biological molecules, then this would correspond to the genomics, to the transcriptomics, to the proteomics, and to the metabolomics. And of course, this is now feasible because all the high throughput technologies which are required to perform these measures are available. So high throughput DNA sequences, which yield genomes. High throughput RNA sequencing and micro or RNA-seq approaches and microarrays, which leads to, which provides transcriptome and microarray norms. Mass spec or 2D gel is not really high throughput, so rather mass spec technology allowing to have proteomes and mass spec technologies in the majority of cases to obtain metabolomes. So what we did actually is to retrieve urine and plasma samples from different centres from the patients, and each patient had a complete set of urine or plasma samples in which we could measure all this single omics here, which were small metabolites. So we measured 189 small metabolites by mass spectrometry, micro RNA on a panel, 173 micro RNAs, 16 plasma steroids, 27 urinary steroids, and plasma nephrons. And the idea was to put them all together into a funnel and to generate multi-omics biomarkers. So again, this is real, this is complete multi-omics, integrated multi-omics in which every single type of omics goes together to generate this multi-omics biomarkers. So this is the, this is all the characteristics of our retrospective patient samples. So of course, patients were included by each reference centres according to the procedures following the endocrine society guidelines. The core NZHT samples were 487. Diagnosis was based on guidelines for each disease in each expert centre. 408 patients had a total of five omics after quality control and data cleaning. And you can see here the characteristics of the patients. We had 100 patients with primary aldosteronism, 69 with PPGL, 30 with Cushing's syndrome, 108 patients with primary hypertension, and we had also 101 healthy volunteers. So you can see that this distribution between males and females was quite equivalent balance. We had, of course, a little bit less number of patients with Cushing's and pheochromocytoma. And you will see that this is all over our studies. There is always this class imbalance which we have to deal with because those diseases are, of course, much rarer than primary aldosteronism or primary hypertension. So the age of the patients was quite young. And of course, her normal intensive volunteers were younger. So the omics, multi-omics measures were, once they were measured by the omics centres, they entered a machine learning pipeline. And multi-omics data integration, which was done at the University of Dundee, first by preprocessing the data, then training and evaluating the data, and then generating models, and then discussions with collaborators. And that was several circles until we got a final signature. So how did we proceed with the biomarker discovery and supervised machine learning? So we selected different disease combinations. We performed, of course, outlier detection. We choose different supervised machine learning classifiers. Then we configured experimental parameters, and we considered different evaluation scenarios. And I walk you through a little bit in all of that. So of course, we had non-overlapping training and validation cohorts, 80% was a training validation, 80% always with 80-20 split. And then we had an independence test data set, which was 20%, which was never used for any training. So the training data set was data sample used to fit the model. The validation data set was data sample used to validate the trained models using 100 random repeats. And then there was the test data set, which was completely independent, unseen data used to test the final trained model. So there were five disease combinations we looked at. We looked at the main disease combination we approached was all versus all. All versus all means we want to distinguish in a pool of patients, we want to distinguish patients who have a PPGL from patients who have primary ALDO, from patients who have a Cushing syndrome, from patients who have a primary hypertension. So this would be the most relevant category probably at the first enter of hypertensive patients into a clinical path. Then we address also the comparisons endocrine hypertension versus primary hypertension. So endocrine hypertension meaning PPGL plus primary ALDO plus Cushing syndrome versus primary hypertension. And then we did also binary comparisons, Cushing's versus primary hypertension, primary ALDO versus primary hypertension and PPGL versus primary hypertension. And actually the omics data from the healthy volunteers were used just to compare individual biomarkers with patients of different hypertensive hypertension types, but were not used to generate the model. So we used eight classifier algorithms, which are classic machine learning algorithms, which are decision trees, naive Bayes, key nearest neighbors, logical logic model trees, simple logistic, random forest, and sequential minimal optimization. And for the feature selection, then we compared different methods, wrapper and filter methods. And then we generated a performance metrics, which was balanced accuracy. So we used balanced accuracy because we had to adjust for this class imbalance problem with between primary ALDO and primary hypertension on one hand, and then PPGL and Cushing syndrome on the other hand, we generated metrics on sensitivity, specificity, AUS IRA under the curve F1 and Kappes-Bohr. We also evaluated different scenarios because we wanted to study the possible biases due to age or sex in our patients. And so we compared different scenarios here, for instance, in the scenario number one, we compared for a set of data containing all omics plus age plus sex versus only omics features. And that was to study the impact of age and sex as discriminating features. Then we compared only the males versus the females. And this was to study the influence of sex by comparing classification accuracy and to find sex specific discriminating features. And we made a third comparison, which is an age based comparison patients above 50 versus patient age below 50. And that was to investigate how omics are affected by age. And of course, hormonal status in women after menopause. And then the top features from the set A where everything was included, were selected with a cutoff for the feature frequency of 50. And this was selected arbitrarily for the final training and testing stage. So, here are the results of this study. So, here are the omics that were investigated. So, this is a total of more than 400 features. So, of course, age and gender, the micro RNA, plasma micro RNAs, plasma metanephrines, plasma steroids, urinary steroids, and plasma small metabolites. And you can see how plasma small metabolites and plasma micro RNAs actually build up quite a huge part of all this of the features. And on the right panel, you can see the selected multi omics features in each one of the comparisons, all versus all, EHT versus PHT, primary aldol versus primary hypertension, PPGL versus primary hypertension, Cushing versus primary hypertension, with the different colors reflecting each type of features. And you can see that in most comparisons, actually, small metabolites were the most, in four of the five comparisons, small metabolites were the most represented features. So, this is a circular heat map showing the top features selected for the classification of the five disease combinations using the multi omics, which is here on the top circle compared to the mono omics. So, just micro RNAs, just metanephrines, just steroids, just urinary steroids, just small metabolites. And you can see here are the names of each features. It's difficult to see, but they are aligned here. And you can see how the actually multi omics signatures takes much less features than each one of the single omics when it combines each one of those omics. So, actually, the results were that the random forest classifier provided approximately 92% balanced accuracy, and this was 11% improvement on the best mono omics classifier, with 96% specificity and 095 AUC for multi class all versus all comparison on an unseen test set using the 57 mono omics features. And for discrimination of endocrine hypertension versus primary hypertension, the simple logistic classifier achieved 096 AUC with 90% sensitivity and approximately 86% specificity using 37 mono omics features. And this is really extremely important, extremely high quality results. So, one plus micro RNA and two small metabolites features were found to be most discriminating for all disease combination, whatever the disease combination, those features were present. And so we have filed a patent on this and we have also published this results. And so, and recently received additional funding to go ahead with this. So this left panel here presents the performance metrics of our multi-omics biomarkers, which you can see here on the top line here and the top line here for the all versus all comparison or the primary aldosteronism versus primary hypertension comparison. And you can see how actually the multi-omics outperforms all the each one of the single omics, whether we compare this all versus all, or we compare the primary aldo versus primary hypertension category. And this is also seen here when you see the area under the curve of the multi-omics compared to the mono-omics. So in order to retrain the signature and to validate it, we performed also a prospective study within the NZHT. And in this prospective study, seven European Society of Hypertension Excellence and Insights Centers recruited nearly 2,500 patients prospectively during the lifetime of the project. When we filtered them out for valid consents or for valid disease, actually we ended up with approximately 1,700 patients falling into a valid category of primary aldosteronism, primary hypertension, PPGL, or Cushing syndrome. And for 1,093 patients, we had valid momics measurements. And so the core NZHT sample here was 1,093. And for 961, we had a total of five omics after quality control and data cleaning. And those are the numbers. And as you can see, again, we had quite equivalent numbers of primary aldosteronism and primary hypertension, but we have quite low numbers of PPGL and Cushing syndrome. And this of course reflects what we see in real life in clinical practice. And actually we applied exactly the same strategy that I have just shown to you before to retrain the signature and to validate the signature. And we ended up with improved performances for prediction of endocrine hypertension. And so we generated efficient momics endo biomarkers. So I would now like to switch and go further and explain why we asked for a new project and why to go actually to more studies to validate this approach in clinical practice. So there is a very nice review, which has been published recently in the Lancet Digital Health, which raises actually all the questions which are associated with implementing artificial intelligence in clinical practice. So most models of AI for health are tested only retrospectively using surrogate endpoints and outside of real world clinical settings. So actually we compensated for that by doing also a prospective study. So that was also a first step. However, it is not uncommon for AI to perform worse when deployed prospectively. And so there is a scarcity of real world evaluation of AI system. And this contributes to substantial uncertainty, including in terms of the possibility of many meaningful risks to patients and clinicians. And so, and also there might be undetected AI biases which lead to disparities in both outcome. So that's why we run a second project and which is HD-Advanced, which is the follow-up of NZHT and which is aimed at improving treatment efficacy in endocrine hypertension by biomarker guided personalized decision support. So the aims of HD-Advanced are to validate the previously identified MOMICS endobiomarkers in a pragmatic outcome-based approach and to go beyond the identification of endocrine hypertension and address also the response to treatment in primary hypertension by developing and validating MOMICS treat biomarkers for treatment response. So together with NZHT, HD-Advanced would provide an all-in-all approach for the diagnosis and the treatment of hypertensive patients. So the consortium is essentially the same we had for NZHT with a few changes. And so again, this is a European project. There are 14 partners in France, in the UK, in the Netherlands, in Belgium, in Germany, in Switzerland, and in Italy. We have seven MOMICS teams. So this means teams making, generating the MOMICS. We have one team for IT and machine learning. We have six clinical teams. And this time we have also three teams for methodology, ethics, and health economics. And I would like to share this with you because actually the use of AI in clinical practice is associated with several methodological and ethical challenges, which we have to address. And for those, we really need specialists in those domains. And then we have one team who is taking care of management, translation, and exploitation to bring this forward to a real use in clinics. So we have implemented three studies in HT-Advanced to improve treatment of hypertension in the general population. So the first study is HT-ENDO. So HT-ENDO is the outcome-based randomized control trial, which will validate the previously developed MOMICS biomarkers for endocrine hypertension, the MOMICS endobiomarkers, for the identification of patients with primary understeranism, PPGL, and Cushion syndrome. We have also another trial, which is currently ongoing, which is AT-PREDICT. So HT-PREDICT is using all the pipeline and all the strategies we have developed in NSAD-HT to extend this biomarker, MOMICS biomarker approach, to identify biomarkers for treatment response in patients with primary hypertension to improve blood pressure control and guide personalized treatment strategy in daily clinical practice. And we will test the combination of MOMICS-ENDO and this MOMICS-TREAT biomarkers for treatment response in a second randomized control trial, which is the HT-TREAT trial, in which we will perform, this will be again an outcome-based, pragmatic randomized control trial, where we will first identify patients with endocrine hypertension and patients with primary hypertension, and then apply appropriate and standard treatment to patients with primary hypertension for a global solution to hypertension management. So if I depict this in the current, so in a patient journey, so this would be the current patient journey of a hypertensive patient. A hypertensive patient goes to the general, to the GP, and here the GP eventually from current data we have, in one to 2% of cases, will address the patient for primary aldo screening. Let's say if this is, if the patient is young, if there is hypokalemia, if there are some, if there is a large pretest probability. In this case, the patient, but this is a minority, if detected or even if primary aldo is not detected, will be able to get a targeted treatment. And in case primary aldo is detected, the patient will be sent to a reference, tertiary care center for confirmation, subtype identification, and then targeted treatment, either medical or surgical target treatment. For the rest of the patients, and this is probably the majority, there will be this one fits all treatment approach, which in some cases will have no effect. And in the case it has no effect, but there is some time here before the no effect is really clearly established, then the patient will go back maybe to a primary aldo screening, which is a delayed screening. The patient has already medications, so we have to switch medications because there are interfering medications that then render the diagnosis more difficult. So it is best to switch those medications and then eventually enter this virtual cycle. And in some cases, the patients may be symptomatic or there will be a suspicion for PPGL or Cushing syndrome. And in this case, again, this will, there will be also time here. Again, those patients will enter the specific diagnostic path and go for confirmation and subtype identification and then get targeted treatment. So this is the current patient journey, but what is the result of this patient journey is that patients with endocrine forms of hypertension have a five to 10 years delay until they get a proper diagnosis. So this would be the new patient journey using the multi-omics biomarkers. Patient would go to the GP and test, an easy test would be available just on blood and urine sample. And then with the detection of omics and application of machine learning, there would be a diagnosis, at least the probability of having primary aldo, PPGL, Cushing, or PHT, primary hypertension, which would be go back to the GP who then can treat the patient according to the disease with targeted treatment or refer the patient to tertiary care if required. So I would like to end my talk just with a few thoughts about how difficult it is actually to implement AI in clinical practice. So this is maybe you might think more of a European problem, not a problem elsewhere, but it might be a problem coming in other countries, in other places too. So it's not a problem actually, it's a challenge, it's not a problem, it's important. So the Europe, as part of its digital strategy, Europe, the EU wants to regulate artificial intelligence to ensure better conditions for the development and use of this innovative technology. And so the Europe, the European Union has released what we call the AI Act. So it's the EU Artificial Intelligence Act, and this is a European regulation on artificial intelligence. And it is the first comprehensive regulation on AI by a major regulator anywhere. And the act actually assigns applications of AI to different risk categories. And AI for healthcare is high risk because if not well done, it could be of high risk for patients. So it is strictly regulated. And also, of course, the AI Act liaises with ethical principle, such as safety, non-discrimination, transparency, among others. So this is something really important that we have to integrate in our clinical studies and all the regulatory issues that are related to clinical studies to bring this AI devices into clinical practice. And there is, again, from the same review in Lancet Digital Health in 2024, there are several issues and research priorities which have been listed and which are important to implement AI-based diagnostic in clinics. And I would really like to share this with you because we are completely into them and trying to go further and trying to make this really working. And so what the issues are is that most study actually are single country study, national study. They may lack representativeness for broader populations. So there is a real need for international collaboration, multi-center trials. And here actually in our study, and I have just put here HDNO, which is our randomized control trial for our biomimics endobiomarkers. We fit into this. We have a multi-center European trial with six clinical centers from five different countries. Also, there is very infrequent citation of reporting guidelines. So this AI guidelines are consolidated standards of reporting of the trial. And so there is a real need for greater transparency in the trial methods and to prioritize comprehensive reporting and participant diversity. So we would address this too in HDNO. We will use international AI reporting standards. Most trials also evaluated interventions or endpoints related to diagnostic yield or performance that might not accurately reflect the overall effect of AI system on patient's care. And so it is crucial for real world evidence to focus on clinically meaningful endpoints, such as symptoms, need for treatment, or longer term outcomes such as survival. So in HDNO, what we will be doing, we will performing a pragmatic outcome-based randomized control trial where we will evaluate blood pressure control at six months following diagnosis of endocrine hypertension using momics biomarkers as the endpoints. That's why I said it was pragmatic and outcome-based. Also AI systems may either streamline or complicate clinical workflows depending on the specific application or the context. So successful adoption of AI tools will depend on factors such as operational efficacy, cost effectiveness, level of training, the performance. And so all of this needs to be addressed. And actually we are addressing this by having procedures which are readily translatable to clinical practice and certification. And so this will also be very strictly monitored in our consortium, and we will also perform an economic evaluation. And finally, there is absolutely a need to address ethical, legal, and social issues for the use of AI in health. And the AI Act has defined this use as critical. And so implementation of AI in clinics requires compliance to well-defined ethical rules. And so that is one of the reasons why in our HD Advanced Program, we have an entire work package which is dedicated to ethics issues. So I would like to stop here. I would like to thank all the participants of the NZHD and NZ Advanced Studies that you can see are listed on this slide. And I thank you for your attention and don't hesitate to ask a question or look up our websites for those programs. Thank you very much. Thank you so much. That was an incredible presentation. We will be taking questions now, so feel free to put them in the chat. We'll give everyone a moment to do that. Let's see, maybe no questions, maybe your presentation was so perfect. I will give you a couple of more minutes, just in case typing takes a bit of time. Let's see, we have one coming in. How do you think you can improve the testing at the primary care and general practitioner level? How can AI and integrated health systems help with this aspect? So that is exactly the important question, because I think the place where we miss most patients for diagnosis is at the primary care. And so it is really, if you look how it is difficult to diagnose this endocrine forms of hypertension properly, and it would be really the most important place where to implement this strategy, and that is actually our aim. The aim is to develop simple tests, I mean simple, they are not simple for the measures of or for the machine learning pipeline, but they should be simple for the patients and for the clinicians, in the sense that they should be, it should be possible to implement those tests, just in general laboratories, where a given set of features is measured, and then integrated into a machine learning pipeline, which gives them an output as a probability of having one of the four conditions of primary aldo PPGL, Cushing's syndromes or primary hypertension. And that is really where we would like to focus all our attention. And on top of that, what we are now developing is trying to develop similar biomarkers, also for treatment response for patients with primary hypertension. But it is exactly at this level of healthcare, where we would like to focus, because once the patients come into tertiary care centers, they are well taken charge. That's not the problem is not there, the problem is to come where the problem is to screen those patients. So I think, yes, I really believe that should be working. And a follow up question to that, is it simple to interpret and act on? So the results, the results are quite so the sensitivity and specificity are quite good, they are really high. And so if I understand the question correctly, is it simple? Can you repeat the question? Is it simple to act on? Is that the question? Yes, simple and to interpret and act on maybe. So the interpretation would come from the machine learning algorithm, the machine learning algorithm would have risk categories or disease categories, I'm sorry, not risk category, disease category. And in the algorithm, the result provided to the clinician would be the disease with the associated probability to make the clinician being able to interpret the results like when glycated hemoglobin is measured or when whatever is measured in clinical practice, but it would be with disease on the other end. It's not to the clinician to make all these calculations or to launch the machine learning machine. It would be a clinical test with a result, which then the clinician can use. Thank you. Another question. Are the algorithms also working with missing data? Yes. So they are. Yeah, absolutely. Because when we train the algorithms, we had sometimes missing data. And so we had to make imputation for missing data. So they are. Perfect. And do you think it's time to integrate these concepts in medical education or having a formal AI curriculum? I think that would be really important. Also because it's so one needs to understand how AI works to be able to use it better. And also, I think as a doctor, we need to also be able to interpret AI driven results in our clinical practice. Just use them as we use other types of support and as either companion diagnostics or that. But it's absolutely important to do so, yes. Thank you. We have another one here. Are there any devices that currently exist or in development that can be used now even in beta testing? So not for this type for our multi-omics, because before we put them into a real device, we have to validate them by a randomized control trial. I guess that is really fundamental. But there are so there are several studies who are trying to develop this type of devices. But this is just going on if we are talking about primary and about endocrine hypertension. But this is all this is all currently going on right now, all the randomized control trials which are needed and required then to put this forward in the clinics. And also, actually, it's also a matter of certifying all the devices of really of them moving them into industrial production and proper production. It's not just so moving really from the bench to the bedside, but with everything which goes with it, with all the proper processes and operating procedures. Perfect. We also have some individuals saying amazing work, excellent presentation. Another question, what would be the cost of this test? Will it be possible to afford in developing countries? So actually, there has been so we have an economic evaluation of this test that will actually will be generated is implemented in the randomized control trial with it some sort of evaluation. But of course, it was not within an RCT when we did our prospective study. And so this is really depends where people are. It depends even within Europe, where we have very similar system, the costs are very different. What is a real cost if you want to do this properly following all the steps which are in the guidelines, that could be quite costly because there is this medication switch and then there is all is the ABS and eventually the surgery. But there have been some studies that show that early diagnosis, even by and targeted treatment, even with surgery, save money. And of course, you have also to consider the long life, the long time costs of morbidity and mortality of those patients when they are not appropriately cured. So we are trying to consider all of this. And after that, we will calculate whether there is a cost benefit. I would bet, I would argue that there might really be a cost benefit in using this type of devices. Perfect. That's all the questions. And that also brings us to time. So thank you again for your excellent presentation. Thank you very much. We appreciate your time. And for everyone watching, you will be automatically pushed back to the homepage now to access upcoming sessions. So thank you again. And we'll see you soon. Bye.

multi-omics biomarkers

endocrine hypertension

primary aldosteronism

arterial hypertension

machine learning

diagnostic accuracy

personalized medicine

Horizon Europe

AI in healthcare