 Great well, why don't we go ahead and get started and move the afternoon on just a bit? I'm Paul Underwood a cardiologist. It's here in Phoenix, and it's my pleasure to Keep this discussion on amyloidosis going. I think it's been great the way that Sandish April the steering committee and certainly the our presenters have been able to lay out an approach to amyloidosis that involves pretty much all specialties pretty all the Specialties that are involved here. We look very closely at the the pathology the physiology the treatments and then some of the clinical Presentations and now I think we're going to go a look at a little bit more in-depth on how we can actually Identify some of these patients and so these next sessions will be on How we identify the patients and what can we do to make sure they're not none the patients are falling through the net? I'd first like to introduce Dr. Vikram Singh From the College of Medicine in Tucson He's an internal medicine resident with the University of Arizona and College of Medicine in Tucson, so please come forward It's oh excuse me Seymart. No, no, no Okay, well nope. Nope, so this is dr. Seem is dr. Seymar seeing Right who is who is a medicine resident in the University of Arizona here? So thank you very much and and we look forward to hearing from you. Thank you everyone And it's an honor to be speaking here, thank you doctor Dave for the for the invitation and It's a privilege to be speaking right before my brother Vic who was just introduced Let's let's talk a little bit about clinical decision support, so what is that well if you ask chat GPT What this is it'll give you something like this and I think we all hope that it would be something on the left But it's probably right now something more like what's on the right if you ask the government And the office of health information technology what CDS is they would say it's a set of tools that are there to provide knowledge person specific information at the right time at the right place and They really taken data from all these different sorts of modalities and Synthesize that for us to make better decisions in health care To me that sounds a lot like a resident so and if you ask chat GPT what a resident on rounds looks like This is this is what it comes up with So why do we need decision support in amyloidosis? Well, I think this has been extensively covered here You know these a lot of these signs and symptoms are not specific and so physicians from a variety variety of backgrounds will end up seeing these patients and That oftentimes if they're not communicating will delay the the diagnosis tremendously and we know that Delayed diagnosis of amyloidosis is incredibly morbid The mean diagnostic delay has been reported to be 39 months and 42 percent of patients will actually not have a diagnosis in four years after After the onset of cardiac symptoms so it's a significant challenge that we're all working to address and The hope is that clinical decision support will Eventually with the help of newer technologies such as artificial intelligence get patients on the Diagnostic and therapeutic pathways much much quicker and so a In example of of how we can build small rules to help us make decisions is you know the Davies score Which we're all quite familiar with now And so by looking at in the literature to see what we thought was associated with cardiac amyloidosis And then creating univariate regression models piling that into a simple score We can look and see that patients with heart failure who have a Normal EF, you know and a high-risk should probably get a PYP scan and be sent down that out that diagnostic pathway What I found interesting about this is that if you take that and apply it to the previous Heft-peft cohort trial cohorts, you realize that there's significant proportion of those patients probably We're at very high risk for amyloid and we probably miss them So if you take this one step further like who then colleagues did they essentially mined the EHR and trial claims data The claims databases to identify ICD codes that were associated with With cardiac amyloid and what they did was they created a they trained a machine learning model So we'll talk a little bit about what machine learning is, but essentially they created a random force model that Would give a give you a prediction of whether or not these patients were at high risk or low risk for cardiac amyloid And so what they found and this is you know something that actually Dr. Dave and and our group has been looking at as well is to really quantify what predictive Comorbidities are out there and so for cardiac comorbidities. They they found atrial flutter pericardial fusion pericarditis conduction defects and Abnormal serum enzymes some of the strongest predictors of cardiac amyloidosis For non cardiac comorbidities They found carpal tunnel syndrome synavitis and tenon synavitis is some of the strongest and also asides All right, so What the same group ended up doing was trying to deploy this within the electronic health record? and so they actually had two publications where they they presented its implementation in four large health systems, I believe Cedar Sinai University Utah MedStar and Wash you where some of this work originated from and What they found was that yes, we can we can implement this system So we can use a machine learning algorithm that will fire and tell us when a patient is at high risk for for ATTR and Unfortunately, that's as far as they got so they really didn't give us any summary statistics They didn't tell us whether or not these alerts were actually valid or they false positives Or they false negative so I think that's the type of work that really needs to be done kind of validating some of these machine learning models Interestingly so something that you know, we were thinking about recently was you know What if we become so reliant upon these these alerts though? Let's say it doesn't fire. Do we now have a lower suspicion that this patient does not have cardiac amyloidosis? Do we rub that from our from our thought process? You know that those are some of the things I'd like to start thinking about and So I kind of transition into artificial intelligence machine learning and deep learning So artificial intelligence, you know, this is a general term to really describe creating intelligence human-like intelligence and The amount of interest in artificial intelligence has not escaped cardiology Right, so just like every other field we have an exponential rise in the number of publications that that have been published actually in cardiology in the field of artificial intelligence and Some of the terms that I think is are important to think about is how artificial intelligence is kind of an umbrella term within which Machine learning is a subset. So there we take data, you know, very well curated data He said to train algorithms that can give us a result something important Whereas deep learning is actually a subset within that where we can actually provide no a priori information and have a data set Basically analyzed by by the computer to give us some sort of Output oftentimes is a black box. We don't know what we're looking So how is this implemented? So a group out of Mayo has created an Algorithm a deep neural network. So this is deep learning, right? So they've provided they've trained this data with almost 3000 EKG recordings of patients with and without cardiac amyloidosis now they use both light chain and ATTR and They found it actually has a very good sensitivity in predicting cardiac amyloidosis. In fact 56% of the time they're able to have a diagnosis With EKG before the actual clinical diagnosis is made So What's also interesting in it, you know, we actually had This was actually effective on a single lead So maybe someday this can be implemented in your in your Apple watch And I won't talk much about imaging them because Dr. Shaw did a great job this morning So I will say that you know integrating that sort of image analysis with clinical data Is going to be very very important and I think that might be the future So what are some of the challenges in AI application? Well reproducibility in general generalizability is very important You know a lot of us know how to read a clinical trial You know, we know how to understand what a good trial is not a lot of us and understand what makes a good AI Publication what goes into that? You know, we did a systematic review of the AI literature and we found that you know Only 25% reported ethnic information and 0% reported any any socio-economic information so the external validity of some of these studies are you know quite thrown into concern actually and Then you know very few actually externally validated 25% and 55% have Made the data available now Only 25 made the percent made the code available. So you see these models are only as good as their code They're only as good as their data. And so without that, you know, it's very difficult to really analyze and interpret how effective they really are So here's an AI generated picture of cardiac amyloidosis. So thank you very much for your time