Risk stratification in preeclampsia using an AI-based approach

Welcome to the final session of the Hemostasis Learning Institute Master Class 2025. I'm Doctor John Mitzios, Medical Sciences Partner at Siemens Health and Ears, and it was my pleasure to be your host throughout this course. Thank you for your participation and in case you've missed previous sessions, you can always catch up. Last but certainly not least, today's speaker is Professor Patricia McGuire. Patricia McGuire is a professor of Biomedical sciences and a director of the interdisciplinary Research Institute, the UCD Institute for Discovery at the University College Dublin in Ireland. She's the author of over 60 peer reviewed publications in leading international journals and has won multiple accolades for her research. Her research focuses on platelets and she has developed the palladin or plated based diagnostics which can in theory be applied to any disease in which platelets have been implicated such as breast cancer, colon cancer, Alzheimer's disease, rheumatoid arthritis, preeclampsia and multiple sclerosis. Preeclampsia has a severe complication of pregnancy marked by high blood pressure, high levels of protein in urine, and a profound hypercoagulable state. With an incidence of approximately 2 to 8% of all pregnancies, it remains the leading cause for maternal death. This talk presents a new AI based approach that enables early diagnosis and treatment of preeclampsia, hopefully helping to save many lives. Professor McGuire, the floor is yours. So good afternoon everybody. Here are my disclosures and the learning objectives for today are the following three learning objectives. We're going to talk about the platelet cargo or as a a term I coined about 20 years ago, the platelet release date and how this is suitable for biomarker discovery. And we will also understand that this barcode of the platelet release date alters in physiologic and pathological conditions. So my journey started off about 20 years ago with this particular publication in Blood. And in this, we looked at the proteins that are released out of platelets following activation. And this was published in Blood in 2004. Since then, I've joined forces with a hematologist called Professor Furnula Neonia and together we hypothesise that platelets and their parent mega megacaricides may be educated by their environment. We are not the only group in the world that believes this, but we are the leaders in looking at the proteomics aspects of the platelet and megacaricides, which is very important as these are anucleate cells. And what we hypothesized was that the cargo that the platelet holds within itself or the release date as we've called it, is a barcode for the health status of an individual at a given time. So it's very interesting, I believe to look at and to catalog these proteins. But just to give you an example of some of the proteins that are contained within the platelet release date, these are pretty powerful factors such as pro angiogenic factors and also antiangiogenic factors such as platelet factor 4, which is a platelet specific protein. There's also a lot of chemokines and cytokines, a lot of coagulation factors and indeed a lot of different enzymes and adhesion molecules and growth factors actually in this release along with proteins found, found inside this release aid. There's actually a large number of small molecules found, including five HT or serotonin. And the platelets are the largest storage of this outside of the brain. So how do we get this information out? Well, we take resting platelets that are circulating in the blood and then we take the, we take whole blood, isolate the platelets and then use an agonist. And in this case, we've used trombin and then we've used trombin to take out all of the information that's stored within the platelet. We captured that information and we can biochemically separate it into different different sections basically. So we can use, we can take all the soluble proteins that are found within like platelet factor 4, like thrombosponin, like the chemokines that I showed on the last slide, Or we can take what's known as vehicles, these small little extracellular vehicles or EVs for short or large EVs which were traditionally shown, uh, and known as microparticles. And we're not the only lab in the world to do this. There are a number of very famous labs in the world doing such analysis. These small extracellular vesicles or exosomes can actually be found inside preformed granules inside the platelets. And here is an electron micrograph showing these tiny little extracellular vesicles or small E VS that. You can actually put about 1000 of these on one human hair. That's how small they are. And but these are found within the preform granules inside the platelets. And when we put an agonist like Tormenon, we can actually get these out and separate them out and find out what's in there. And just to put this into size into perspective, so these small EV's, like I said, you could fit or exosomes, traditionally known as exosomes, you couldn't fit about 1000 of these on one human hair. And these are interestingly the same size as as viruses, which platelets also take up in store. And these large extracellular vehicles, traditionally known as microparticles or micro vehicles are similar sized to bacteria. And then platelets themselves are between 1:00 to 5:00 microns. And this is a lot smaller than what a normal cell would look like. And this is just for your information to put these all into perspective in terms of their size. So how do we profile all of these? Well, we've designed a, a, a protocol that we take these blood samples from patients and it's all down to really logistics. So we need to get that blood from the patient's arm and into the lab within one hour. And the lab means the lab in our Research Institute. So when we do get that in, within one hour of blood draw, we can isolate this plate, platelet release, age. We can umm, take all of the proteins that are in there and digest them up and separate them out by mass spectrometry or subject them to mass spectrometry. And then we do a lot of protein mass spectrometry in the lab, but we don't analyse this in standard ways. We actually use machine learning or artificial intelligence techniques to actually analyse all of the information that the mass spectrometer gives to us. And in this way we can look for diagnostic differences. And this particular protocol is our Paladin protocol or platelet based diagnostic protocol. And we have done this particular analysis, this Pagan protocol, we have looked at 32 healthy different donors using this. So why is this interesting? Because most people only look at very small number of donors using proteomic analysis. But here we compared the plate release date of 32 healthy different donors and we published this back in 2018. And what we found was that when you look at the barcode of each of these 32 different healthy people and each line, if you like, on this heat map represents one of those 32 healthy people. And when we look at the barcode of their platelet release date and the proteins that are expressed there, we see that they are, it's remarkably reducible. In fact, so reproducible that when we look at the coefficient of variation in these particular 32 healthy people, the the coefficient variation is between 2.6 and 3.5%. So it's a tiny amount of variation. And we've published this back in 2018. So what we find is that the majority of the plate release aid proteins are vesicle derived. If we look at sea here, what we've done is we have taken the small little extracellular vesicles from all of these 32 healthy people and looked at the amounts of them across different size ranges. And what we see is also reproducible between. So we look between 50 to 200 nanometers and we see that they look very similar and we can actually look at pictures of these underneath the microscope and see these particular vehicles and measure the different sizes and also measure the amounts. We've also looked as well as in 32 healthy people, we've also looked at 18 pregnant women giving the proof of principle in human pregnancy that the platelet release date is actually altered in healthy pregnancy and we published this in 2019. What we found was that when you compare healthy pregnancy versus healthy women who are not pregnant, we actually see a difference in the barcode. So if you like, this is a very expensive pregnancy test, you can basically look at someone's plate release date and see if they are pregnant or not. But we can separate these out by principal component analysis and also we find 11 proteins that completely separate out the two groups. And this was our proof of concept in 18 healthy donors, healthy pregnant donors. Then we went on to look at the platelet release date in preeclampsia. What we did was we recruited women over a 21 month period. We were specifically looking for women over this time period that were diagnosed with early onset preeclampsia. We identified 40 women over these time periods and recruited 26. So why look at the complication of preeclampsia? Well, every 7 minutes a woman loses her life because of preeclampsia, making about 76,000 women dying from preeclampsia every year. Every 45 to 60 seconds a baby dies from preeclampsia, giving statistics of over 500,000 babies dying from preeclampsia complications every year. The symptoms can really vary among the one in 12 women whose pregnancies are complicated by preeclampsia. Some women have no symptoms at all, some women can have very mild symptoms, and some women's symptoms can be severe and get progressively severe very fast. Right now there are no quick diagnostic procedure to to diagnose preeclampsia or also prognos preeclampsia and prognos future outcome. We've looked in this group that we recruited, we've even looked at mean platelet volume and actually a time of booking mean platelet volume was the same. So you can't really predict preeclampsia looking at mean plate volume. But actually at at time of onset of early onset preeclampsia, these women did have a change in mean Pate volume saying that there is something different about their platelets. So we isolated platelets for these women and we subjected to these platelets and they're platelet release 8 to our pardon platform and identified a number of peptides and proteins that were different between these particular women. We actually could separate out the two groups with this early onset preeclampsia versus pregnancy. We could separate out these two groups based on 26 proteins or 33 peptides. And we actually found that there was significantly increased number of proteins of of great interest and we patented these particular proteins. And this is the subject of this proposal AI Preemie that I'm going to talk about in a minute. We also in our analysis uncovered some markers that could predict the future severity of outcome. So we actually uncovered particular markers that could predict a mother who would give get severely sick from a mother who would remain stable. And this was the basis of the AI Premium project, what AI Preemie is, it's a project now utilising our patented biomarkers. And we're testing these across 3 Dublin maternity hospitals, which covers about 50% of all the births in Ireland. We are working, Fennula and I are working with obstetricians and gynaecologists in each of the three Dublin maternity hospitals. And we're also working with midwives and research labs based in these hospitals. We have a lot a number of staff in in in our labs, but we're also working with clinical research centres to process and look for the biomarkers in each of these women's samples. We're working very closely with data analytics and machine learning, both scientists and also companies, namely SAS and Microsoft. We're also looking at care pathways and value care pathways in both in Ireland and in multiple jurisdictions that we're aiming to launch AI Premium working with health economists in that area. So what is AI preemie? Well, AI Preemie, we believe is going to be a tool that will provide clinicians with a decision support system that we believe will remove diagnostic uncertainty and ensure inspected mothers and their babies are provided with the best informed clinical care at that time. We hope this will save lives. We hope this will reduce mortality, reduce the length of stay and reduce other costly tests and follow up. What we're doing is we are taking our patented biomarkers and we are combining this with lots of other information about the women and bringing this in to the clout. So I'm going to explain this on the next slide. So I've talked you through our diagnostics, where they've come from, they've come from mass spectrometry. But what we know from our platelet studies is that if you find a a particular marker in a platelet, you can also find it in the plasma of that woman. So this is if you like uncovering a needle in the haystack. So what we've done is we've converted our tests from mass spectrometry to a simple ELISA based test that you can do in any hospital lab across the world. Still we've completely converted this in to we, we think actionable diagnostics from very small amounts of blood. As I said, we are in the three Dublin maternity hospitals, the Coombe, the Rotunda Hospital and also the National Maternity Hospital with ethical approval granted in each of these hospitals. We are in the middle of this. We want to recruit 1000 women into our study. We are at at this point 327 women and we have tested our diagnostics, our patented diagnostics and prognostics in each of these women. And we've also combined this with all the information available on that woman throughout her pregnancy journey. So what do I mean by all the information available? So taken our AI preemie test, our patented diagnostics that I just showed you how we found over the last number of years. We've also taken patient history. We've also taken in a clinical assessment of that woman and the ground truth at that moment in time. And then all her investigative data, including her hematological data from her time of booking in that hospital. And we combine all of that information together and bring it in to a a analytics platform known as SAS via. And we run this in the Microsoft Azure cloud. And the reason behind all of this is that we can do this in any hospital across the world that has access to the Internet. Basically. Where are we right now? And this current model is based on 300 women and we are right now with an area under the rock curve of .866. So in comparison to other available solutions in the third trimester. So this particular solution is for women in the third trimester or late 2nd trimester, somewhere between 20 and 40 weeks of gestation. That's who we're recruiting to our study. So we are now at an area under the rock curve of .866, as I said, but this is a small number of patients. There may be some overfitting here. We want to get to 1000 patients and we will in 2024. That's the goal. This idea is that if you combine actionable diagnostics that you get and that you uncover through excellent biochemistry together with all of that information available on the patient, combine it together into a solution that you can create. This unified by signature that we hope will give effective, efficient clinical decision making. And actually, we believe that you could do this really for any dis, any complication or any disease. And the vision is that we want to get our solution to every person who needs it across the world because we really do believe it will save lives. So in conclusion, we have shown that the platelet cargo or this platelet release date is a barcode for the health status of an individual at a given time. And I'm very privileged to lead this team, but it's a very multidisciplinary team from across multiple research institutes, must multiple hospitals and also our two partner companies on this particular project, which is our SAS and Microsoft. And we're very grateful to the National Irish Funder Science Foundation Ireland for supporting this since 2020. And just to bring to your, umm, bring to your information that we've just published a, a chapter in a new book from Wiley called AI in Clinical Medicine and our chapters in. Titled AI in Hematology. Thank you very much.

20+08 10 15 40 35 25 30 100 75 50 252 6394 300 200 20 15- Hemostasis Learning Institute presents Risk stratification in preeclampsia using an Al-based approach Prof. Patricia Maguire Powered by hemostasis experts John V. Mitsios PhD, Medical Sciences Partner Siemens Healthineers Disclosures Mallinckrodt pharmaceuticals, Microsoft, Google, Intel, Dell and SAS Institute Employee No relevant conflicts of interest to declare Consultant Major Stockholder Speakers Bureau Honoraria Scientific Advisory Board UCD SIEMENS Healthineers Learning objectives 1. The platelet cargo or the releasate (PR) is remarkably reproducible & suitable for biomarker discovery 2. The "barcode' of the PR alters in healthy pregnancy 3. The PR can be used to uncover differential markers in Pre-eclampsia blood Volume 103, Issue 6, 15 March 2004, Pages 2096-2104 Genomics Characterization of the proteins released from activated platelets leads to localization of novel platelet proteins in human atherosclerotic lesions Judith A. Coppinger, Gerard Cagney, Sinead Toomey, Thomas Kislinger, Orina Belton, James P. McRedmond, Dolores J. Cahill, Andrew Emili, Desmond J. Fitzgerald, Patricia B. Maguire CAN S Hypothesis Evidence increasing that platelets and MKs may be "educated" by their environment Flaumenhaft, 2012 Best et al, 2016 Koupenova et al, 2018 Becker et al, 2018 Davizon-Castillo et al, 2019 Cunin et al, 2019 We hypothesised that Consultant Haematologist Rotunda & Mater Misericordiae Hospitals, Dublin. "The platelet cargo or the releasate (PR) is a barcode of the health status of an individual at a given time" Platelets play a broader role: Platelet Releasate (PR) Pro-angiogenic factors Enzymes VEGF, SDF-1a, CXCL12 Acid phosphatases, heparinise, hexosaminidase Anti-angiogenic factors Angiostatin, endostatin, PF-4, Small molecules thrombospondin-1 5-HT, ADP, ATP, Mg2 *, Ca2+, polyP Chemo- and cytokines Adhesion molecules PF-4, RANTES, NAP-2, P-selectin, vWf, Fibrinogen, Fibronectin TG, CD40L Coagulation factors Apoptotic factors plasminogen, protein S CD95, Apo2-L and Apo3-L Anti-coagulation factors Antithrombin, Tissue factor Soluble Proteins () Growth factors Pathway Inhibitor, APP Exosomes () PDGF, TGFß1, EGF, IGF-1, 9protein nexin II) Microparticles () VEGF Heijnen et al, 1999 Coppinger et al, 2004 Garcia et al, 2005 Piersma et al, 2009 Wijten et al, 2013 van Holten et al, 2014 Aatonen et al, 2014 PR contains two types of EVs & soluble proteins Small EVs (Exosomes) Resting Platelets Activated Platelets Soluble Proteins Agonist Biochemical Separation Large EVs (microparticles) Parsons et al, 2018 Szklanna et al, 2018 ble Proteins Exo Total particles Wild type 500nm et al, 1999 er et al, 2004 et al, 2009 en et al, 2014 Putting EV size into perspective soluble proteins Protein Platelets aggregates Viruses Bacteria Exosomes Microparticles Apoptotic Body Cell (Small EVs) (Large EVs) Proteomic profiling of the PR: 100 PREPs project Blood samples Platelet releasate Sample lysed and Proteins enzymatically isolated proteins denatured digested Diagnostic differences Proprietary ML algorithms to Peptides separated by HPLC ionised identify differential and analysed by MS/MS in a Q-Exactive peptides Orbitrap mass spectrometer LOGISTICS solated 100 Platelet Releasates (100 PREPs)REPs project Proof of principle in 32 healthy donors RESEARCH ARTICLE Proteomics Label-Free Quantification www.proteomics-journal.com Platelet Releasate Proteome Profiling Reveals a Core Set of Proteins with Low Variance between Healthy Adults Martin E. M. Parsons, Paulina B. Szklanna, Jose A. Guererro, Kieran Wynne, Feidhlim Dervin, Karen O'Connell, Seamus Allen, Karl Egan, Cavan Bennett, Christopher McGuigan, Cedric Gheveart, Fionnuala Ní Áinle, and Patricia B. Maguire* nostIcs IRISH RESEARCH COUNCIL HRB proteins Core releasate Other releasate Coeffiecient of variation (%) ₩ Control 32 Control 29 PR is remarkably reproducible & suitable for The majority of PR proteins are vesicle-derived Percentage of genes Lysosome Cytoskeleton Extracellular Extracellular region Core Vesiclepedia releasate Parsons et al, 2017; 2018 Particles per ml Size (nm) 100 Platelet Releasates (100 PREPs): Proof of principle in human pregnancy Clinical Applications DATASET BRIEF The platelet releasate is altered in human pregnancy Paulina B. Szklanna, Martin E. Parsons, Kieran Wynne, Hugh O'Connor, Karl Egan, Seamus Allen, First published: 14 October 2018 | https://doi.org/10.1002/prca.201800162 THE ROTUNDA HOSPITAL SPITAL EARCH COUNCIL The "barcode' of the PR alters in healthy pregnancy PCA Pregnancy v Control 11 predictive proteins Up in Pregnancy Down in pregnancy AGT HRG C9 IGKVD-28 FBLN1 HPS PLG Component 2 SERPINA1 High VTN LOw Proof of concept: Healthy pregnancy in 18 donors Szklanna et al 2019 LOW Uncovering differential markers in Pre-eclampsia · Patients were recruited over a 21-month period. · 15,299 deliveries >24 completed weeks. · 334 (2.2%) were complicated by pre-eclampsia · Majority (294) late-onset (>34 weeks gestation) · 40 were classified as early onset pre-eclampsia (<34 weeks gestation) · 2 impossible to recruit (very short interval to delivery) · 10 had exclusion criteria · 26/28 remaining consented to recruitment Problem: Preeclampsia World Pre-eclampsia Day 2022 isuog. Every 7 minutes a woman loses her life due to pre-eclampsia (PE) associated complications Globally, 76,000 women die each year from PE 500,000 babies die each year from PE through premature delivery - the only cure for the condition Symptoms of Pre-Eclampsia Severe headache that Swelling of the face Weight gain of more won't go away even and hands than 2 pounds / 1 kg with medication in one week Difficulty breathing Nausea after Changes in vision Upper right mid-pregnancy (spots, light flashes, belly pain / or vision loss) shoulder pain International Society of Ultrasound in Obstetrics and Gynaecology (ISUOG) gasping, or panting mid-pregnancy gasping, or panting Uncovering differential markers in Pre-eclampsia: Mean platelet volume MPV at booking MPV at week 28-33 MPV at week 34-40 Cases Controls Monteith et al, 2017 15 - Coefficient of variation (%) Pregnancy Significantly increased EOP releasate proteins Log2 Total protein amount Unpublished Differentiating based on PET severity Predicting a mother who will get severely sick from a mother who will remain stable Health Economics Prof Jennifer Donnelly Dr Brian Mac Namee Dr John O'Loughlin Prof Gerardine Doyle Dr Zara Molphy Ms Deirdre Clissmann Dr Kate Cullen Lab/Research Commercial Infrastructure PREMIe Prof Patricia Maguire Ms Ella Fouhy Ms Ana Le Chevillier Project Management Mr Andrew Warrington UCD Clinical Research Centre Mr John Curran The National Maternity Hospital ENTERPRISE IRELAND Microsoft Clinical Lab Management Data Analytics Assoc Prof Nell O'Gorman Mr Edward Simons Ms Cora McCann Dr Martin Kenny Ssas Coombe ombe RELAND > AI_PREMie provides clinicians with a support tool that: / Removes diagnostic uncertainty Ensures expectant mothers and babies are provided with appropriate and better-informed clinical care / Will reduce mortality, length of stay, costly tests and follow-up / Early collaboration with Microsoft and SAS Institute, primed for 'in clinic' model deployment > We want to relate all available data (from sick expectant mothers) to models (about sick expectant mothers) to augment decision making in clinical care. > A unified biosignature of Pre-Eclampsia Diagnostics underlying AI_PREMie PET Screen Orbitrap Mass Spectrometry Building the Library from DDA Data MaxQuant Skyline Statistical Analysis of Skyline Output Perseus ELISA > Transition from bench to bedside / From mass spec to antibody based From platelet releasate to plasma Actionable diagnostics from small amts of blood > Protocol design enables translation to any hospital AI_PREMie study design Ethical approval granted from 3 tertiary maternity hospitals in 327 patient recruited and Generation of prototype ML Dublin (capturing 50% of all births in Ireland) biomarkers in plasma Goal: 1000 patients Prototype test for Pre-eclampsia covering 50% of all the births in Ireland AI_PREMie: A unified biosignature UNIFIED ADAPTIVE Microsoft Azure SAS Viya OPEN PET TEST Patented Patient biomarker Investigative quantification History Assessment /Ground truth data AI_PREMie: Where are we right now? Measure Value [95% CI] AUROC 0.866 [0.759 - 0.973] Sensitivity 0.7 [0.504 - 0.896] Specificity Accuracy 0.852 [0.756 - 0.941] The current prototype champion model is gradient boosting with a misclassification rate of 14.75%. Effective, Efficient Clinical Decision Making Vision Our team wants to get AI_PREMie to every person who needs it across the world Conclusion The platelet cargo or the platelet releasate (PR) is a barcode of the health status of an individual at a given time Acknowledgements AI IN CLINICAL MEDICINE DeLL Siemens Healthcare Diagnostics Inc ., 2025 The products and features mentioned here are not commercially available in all countries. Their future availability cannot be guaranteed. SLS-24-3803-76 QR700016156