Transcription:
Harnessing Artificial Intelligence to Improve Cardiovascular Care

Melanie Cole (Host): Welcome to Better Edge, a Northwestern Medicine podcast for physicians. I'm Melanie Cole. And I want you to join us as we explore harnessing artificial intelligence to improve cardiovascular care. Joining me in this fascinating interview today, is Dr. Ramsey Wehbe, he's a former Artificial Intelligence Fellow and a current Heart Failure Fellow at Northwestern Medicine.

Dr. Wehbe, as I said in my intro, I'm really psyched for this topic because wow, it certainly is a wave of the future in cardiovascular care with heart disease, being just the number one killer of men and women and you know, something we really need to tackle. Before we get into this topic, I'd love it if you told us a little bit about yourself and your interest in this topic as a cardiologist and an AI engineer.

Ramsey Wehbe, MD (Guest): Well, thank you so much. And thanks for having me on today, Melanie. This is a topic I'm very passionate about and I'd say, you know, my entire career I've always been what I would call a data science geek that dates back to medical school. And very quickly I realized the power of data and the ability to unlock insights from data and use those insights to improve care.

And I spent a year at the National Institutes of Health and learned some data analytic techniques from a statistician, staff statistician there. And that's where really my love of this field blossomed. But I realized over time that, that the data that we were collecting in medicine started to change. And the majority of data we collect today is unstructured data like imaging and clinical notes, free text and sensor data. And so I started to see that there was this second revolution in the, in the data science world of modern artificial intelligence. And I became very excited about these technologies and it just so happened that the stars aligned and at the very same time, we were developing this artificial intelligence program at Northwestern which I took full advantage of.

Host: Well, then give us an overview of your AI Fellowship at Northwestern University that you completed last year.

Dr. Wehbe: Absolutely. So, the Bluhm Cardiovascular Institute's Fellowship in Artificial Intelligence is a first of its kind program. It's designed to train future leaders in artificial intelligence and cardiovascular medicine. I think leadership at the BCVI quickly realized that in order to develop truly impactful digital innovations it isn't enough to have clinicians and engineers working in their respective silos, but rather we need physicians who are trained as cross domain experts to bridge the clinical and engineering scientists and bring these technologies all the way from conception to the bedside.

So, this is a one-year program. It's designed for clinical cardiology or cardiac surgery fellows, and it involves intensive immersion and formal instruction in computer science, data science, and the theory, design and implementation of AI and machine learning systems, which culminates in a Master of Science in Artificial Intelligence Degree from the Northwestern McCormick School of Engineering.

I should say, this is not a program that is designed to simply expose fellows to the concept of AI. But rather it empowers fellows with the skills to actually build artificial intelligence systems alongside other computer scientists. Additionally, during the course of the year, AI fellows are engaged in research alongside senior faculty in order to prepare them for successful careers at the cutting edge of applied AI in cardiovascular medicine. Completing the fellowship, like I said was one of the smartest things I think I've ever done. In fact, I intend to build my career off of it, and it's really given me the tools and the foundation to excel as a physician scientist in this space.

Host: Excellent information. And thank you for sharing that. So, why don't we tell the listeners what is artificial intelligence? Because we, we really hear these terms really, we've heard these terms since like the 1960s, and we think of these robots and wandering around and we don't really know what it is. Tell us what machine learning and deep learning are and how these technologies can be applied to improving medical care overall. And then we'll get into cardiovascular care specifically.

Dr. Wehbe: Yeah, these are all really great questions. And I think these terms cause a lot of confusion. So, to try to clear it up, artificial intelligence is simply the study of machines or computers designed to perform tasks that we normally reserve for human intelligence. And you can see right away that this is a moving target. As our computers get smarter over time, so does the bar for what's generally considered artificial intelligence. And AI is not a new field, as you mentioned, it's it has its origins in the 1950s. Originally AI systems were rules-based systems, which essentially are computer programs designed to perform certain tasks. But the current explosion of interest in AI is primarily being driven by breakthroughs in machine learning and particularly deep learning.

So, machine learning involves computer algorithms that are designed to learn without explicit programming instead relying on patterns and inference from data to learn to complete a particular task. We should be completely clear that although machine learning seems to evoke these images of a sentient machine, a computer learns by simply solving a number of mathematical problems in the context of some predefined algorithm. Deep learning is a subset of machine learning that relies on a specific type of algorithm called an artificial neural network, which is inspired by the human nervous system.

These algorithms contain nodes or neurons, which are stacked deeply in successive layers, which is where the term deep learning comes from. The remarkable thing about deep learning algorithms is that they are able to automatically model complex and unstructured data like images or free text data with very little human input. So, sort of the same way a child learns to distinguish a cat from a dog by being shown many examples and learning from mistakes; we can train a deep learning model simply by exposing it to many examples so that the model can learn from mistakes and adjust parameters to build a better performing model over time.

And deep learning has been revolutionary for many industries, underpinning breakthrough technologies, like self-driving cars and smart voice assistants. And we're starting to see remarkable applications across medicine to automate mundane or time consuming tasks, assist in the detection or diagnosis of disease, or even to predict prognosis.

We're in this area of big data, but as I mentioned in medicine, the majority of the data that we actually collect is unstructured data. And so modern machine learning allows us to model this data in a high throughput fashion in ways that simply were not previously possible. So, just as an example, as an AI fellow, I helped lead a large multidisciplinary effort last year to develop, train, and evaluate a computer vision deep learning system designed to detect COVID-19 on chest radiographs, using PCR results as the reference standrd.

And we published the results of this system in the Journal of Radiology showing performance similar to that of a consensus of experience, thoracic radiologists with results in a fraction of the time. So, this type of system could be important for triage say in an emergency department setting while waiting definitive results from a PCR test to come back and it shows how deep learning technologies can allow us to respond to problems in the clinical domain in novel and really agile ways.

Host: So many applications. Really, it is absolutely fascinating. So, what are some of the implications of using it in the clinical care of patients with cardiovascular disease? Where do you see its best uses or even future uses?

Dr. Wehbe: Absolutely. So, well machine learning in cardiovascular medicine is still in its infancy; research in this space is accelerating at a really a breathtaking pace and particularly in the field of cardiovascular imaging. And we're, we already have several mature commercial products in that space designed to automate processes and aid and detection are diagnosed of disease.

One particularly interesting application from a company called Caption Health, which we partnered with, allows novice ultrasound operators to acquire essentially diagnostic quality bedside point of care ultrasound images of the heart. And so that's just one example of the many across the spectrum of cardiovascular disease, we're seeing applications in ECGs and home monitoring systems.

But as these technologies get rolled out, I think it's imperative that clinicians are familiar with the potential risks and benefits they bring and how to really evaluate these systems. One of the largest barriers to widespread clinical implementation of these systems and widespread adoption of these technologies is often they have poor performance outside of the patient cohorts in which they were developed.

And this is a problem called overfitting. And it's akin to memorizing the answers to a test rather than learning the underlying concepts. Concerningly models can even overfit to bias in datasets, which can lead to systemic bias or prejudice in their predictions. So, it's very important that any system we trust clinically is repeatedly validated in diverse cohorts outside of the institution or dataset on which it was trained.

And the last thing I'll say is to be successfully implemented, there needs to be successful interaction and partnership between the human and the machine or what we call the human computer interface. I want to be clear that these systems are not designed to replace clinicians nor could they, they're not that good.

A successful system is one that assimilates into the clinician's workflow and has some degree of interpretability and explainability. And when implemented effectively you know, these systems working in concert with the machine are far superior than then a machine or human working alone. And that's a concept known as augmented intelligence, which I think is a very important one.

Host: Well, while you're bringing a bunch of this up, Dr. Wehbe, potential risks and challenges in implementing these technologies. You've just mentioned a few of them. Expand a little bit more for us about some of the barriers to bringing these systems directly to the bedside and improve outcome. I'm sure that you've heard myths surrounding it. I'm sure that you've had questions from other providers. Answer some of those now, if you would.

Dr. Wehbe: Yeah, I think one of the myths that I touched on briefly was the idea that you know, modern machine learning or modern AI is coming for your jobs and that couldn't be further from the truth. I think what I see is that these technologies are going to allow us to be more efficient and free us from burdens of sort of the mundane tasks and things that, that really take us away from the bedside with patients and will allow us to be more compassionate physicians.

But as I mentioned, that there are risks to, to these technologies and we've seen very high profile risks of systemic bias outside of medicine, and even within medicine already for systems that have been deployed particularly in, in a system that may be racially biased based off the data it's trained on. And so we need to be very careful about making sure that we rigorously evaluate these systems. And I think the future of this field is prospective randomized clinical trials, where we can really engender trust in the technologies and show that, you know, they not only work on retrospective data sets, but prospectively, it actually improves patient care and helps us improve the quality of care we give to our patients.

Host: Well, then tell us about any AI projects that you're working on right now that you'd like to mention for other providers.

Dr. Wehbe: So I think one interesting project is we're leveraging deep learning for the detection of ATTR cardiac amyloidosis. So ATTR cardiac amyloidosis is a form of infiltrative cardiomyopathy. It's caused by inappropriate folding of the transthyretin molecule and deposition of amyloid fibrils in the myocardium.

It remains an underdiagnosed cause of heart failure, especially in older patients, hospitalized with heart failure and patients with severe low flow, low gradient aortic stenosis, and it carries with it a significant morbidity and mortality. So, early diagnosis is key, especially given we now have an effective therapy that improves survival and quality of life for these patients, the TTR stabilizer tafamidis.

And thankfully, we also have had remarkable advances in the diagnostic tools available for ATTR cardiac amyloidosis particularly cardiac scintigraphy with bone radio tracers, like the technetium pyrophosphate radiopharmaceutical also known as PYP imaging. Cardiac scintigraphy has become a key component of the noninvasive diagnosis of ATTR cardiac amyloidosis, and it's reported performance in the literature similar to that of the invasive gold standard of endomyocardial biopsy. Unfortunately we and others have observed that the real-world performance of cardiac scintigraphy for detecting cardiac amyloidosis outside of expert centers may actually be much lower. And there are many potential pitfalls in the interpretation of this imaging for the inexperienced clinician. And so this limits the widespread clinical utility of this important diagnostic modality may lead to unnecessary invasive follow-up testing or misdiagnoses. So one potential solution we thought lies in leveraging advances in artificial intelligence for the task. So I was fortunate to receive grant funding from the American Society of Nuclear Cardiology and Pfizer for the development, validation and implementation of a deep learning system for the detection of ATTR cardiac amyloidosis on cardiac scintigraphy. And we've already built it, a pilot system that works on planar imaging alone that exhibits performance on par with experts at our institution.

And we're currently refining the system further to include some SPECT imaging and full 3D volumes. And our hope is that this system will you know, allow maybe more inexperienced clinicians to make an earlier and more accurate diagnosis of ATTR cardiac amyloid.

Host: What an interesting episode we are having today on the Northwestern Medicine podcast. As we wrap up, Dr. Wehbe, please summarize for us what you would like to tell other providers about translating artificial intelligence in cardiovascular medicine, from the bench to the bedside to improve clinical outcomes and clinical care.

Dr. Wehbe: Yeah, I think, one thing I'll say is that these technologies are with us. The future is now. And you know, I, I think if we are responsible about the way that we develop these technologies and use these technologies, it can do wonders for improving outcomes for our patients, improving our own quality of life and our own efficiency.

And I think it's just important that, that we're all sort of educated about these technologies and to some degree can evaluate how, how well a technology might perform for your patient. But I do think that this is a transformative age for artificial intelligence in, in cardiovascular medicine and I'll borrow a quote from a Professor of Radiology at Stanford, Curtis Langlotz.

And I'll adapt it a little bit to say "cardiovascular clinicians will not be replaced by artificial intelligence technologies, but cardiovascular clinicians who use and understand artificial intelligence technologies will replace those who don't." So, you know, I think it behooves us all to really learn about these technologies and make sure that they're used for good.

Host: Beautifully said, and it certainly will augment your armamentarium of options and therapeutic availability. So I, it's just really amazing. Thank you so much for joining us today, Dr. Wehbe. To refer your patient or for more information, please visit our website at breakthroughsforphysicians.nm.org/cardio to get connected with one of our providers. That concludes this episode of Better Edge, a Northwestern Medicine podcast for physicians. For more updates on the latest medical advancements, like you heard here today and breakthroughs, please follow us on your social channels. I'm Melanie Cole.