Anthony Chang | Data and Machine Learning | Exponential Medicine 2016

One of the best things I like about Exponential Medicine is that it's really a special sanctuary for all of us who are visionaries, and maybe even considered a little bit of a unicorn at our institution — sort of a herd of unicorns, if you can say. Intellectually, I'm very grateful to Daniel for having started this in the first place and really spurring me on to chase my dream and vision of developing an expertise and experience with artificial intelligence and medicine. So, a little bit about my bio on one slide: basically, I'm a practicing pediatric cardiologist. I founded not just on evidence but data. I'm just curious how many people here are active clinicians. Okay, so this is an amazing number. So you can understand what I'm talking about. So February 14, 2011, the night that the supercomputer from IBM beat the human contestants on Jeopardy, was the night I realized artificial intelligence was here in a big way. And that was the night that I downloaded the application for the Stanford data science program. And, so that was quite an epiphany. Going back to school as a senior physician, with my classmates a third to half my age — what does that feel like? And you can... Russian hockey team on a daily basis. I'm on the ice, I got skates on, but that's about it. So for those of you who are thinking about such a career, it's absolutely worth the journey, because as hard as it is, it's amazing what you learn in a very short time.

So I'm a clinician, so I just want to talk about my day-to-day right now and where the pain points are. So when I see a patient — this happened last week — I look at the echocardiogram, since I'm a cardiologist, try to remember what the last echocardiogram showed, sometimes going back to the last study, get some data, encourage parents to give me data. So this is what I sometimes get. You can tell that some of the readings are taken during breakfast time. Simple text on my iPhone. So this is 2016, with all the technology that you've heard about today and will be hearing about in the next few days. Why are we not further along in a typical clinical setting? This is what we're about to launch next month at Children's Hospital Orange County, something I call an intelligent clinic. So we're trying to leverage the technology, artificial intelligence, and the analytics into our daily practice. You've heard great things about technology and what's available, but how do we distill that into a day-to-day situation is the hard part. So we're going to have wearable technology for vital signs. Ideally, we're going to look into having, for instance, pharmacogenomics, so that we know which medicine and what dosages would be best for that particular patient, which is not typically done as you know. And then there will be compliance measured with a sensor technology that's readily available. And lastly, we're going to have an AI-inspired educational avatar for all the questions and the nuances that come up between visits. So we want to build a continual experience for our patients and families.

So what is it going to take for this to happen in the near future? Three things I'd like to leave with all of us. First, we have to look at the cultures of clinicians and the computer and data scientists, and how the data scientists actually understand physician culture. So we're sort of educated on evidence-based medicine when, in fact, the world is changing into what I call intelligence-based medicine, where you have to dig data deep by writing algorithms and getting data from EHRs as well as other types of data. So the world's changed, and physicians are not always cognizant of that. The second aspect of the culture of physicians is, yes, we have a little bit of hubris. We simply can't accept the fact that a machine can perhaps read an MRI more accurately and faster than we can. No matter how many times I talk to radiologists about this, there's always a pushback. And we should accept the fact... mentioned machine learning. This is a review article or editorial that literally came out this week on machine learning for the first time. So if you Google or PubMed search 'machine learning' and JAMA or New England Journal, there are a variety of articles on this topic. So no, when physicians are not well read on this particular topic, and then the sense of competition — hospital to hospital, there's really little sense of sharing data that the data scientists have and have open-source software. So we're used to competing, not collaborating. What if we look at the flip side with the data scientists? Most data scientists would love big data, and sometimes I think just as important is to look at the... that's something that I think data scientists can hopefully in the future appreciate a little bit more. They're always focused, or mostly focused, on decision support, right? That's what they think helps physicians, when in fact sometimes what we need help with is just workflow efficiency and access — access to data in a timely fashion. And I'd like the data scientist sometimes to focus a little bit more on that also. Correlation doesn't mean causation, and that's something that the two cultures need to reconcile. And lastly, what I like about the data scientist world is there's wide open collaboration. This is the number of available... Neil talked about our software before. Most physicians have not heard of R or software, and probably not in a productive fashion. Work more closely with one another. It's rare that you have a data scientist team work very closely with physicians. It's also very unusual for physicians to have a data science background or work actively on a day-to-day basis with data scientists. So I think both cultures have a lot of ground to make up in terms of understanding each other and reconciling the differences between the two cultures.

The second aspect of this is the data aspect. And one — we're at the presidential election time, and one of the very famous bylines in the previous election was 'It's the economy, stupid.' You probably remember that. And I like to say it's the data. It's all about the data. The AI part is... getting into the classical period in medicine; we're still sort of in the medieval period. You know, we're still struggling with use. And of course you know that to have good intelligence or AI application, you need really, really good data. What's happening in healthcare and data? Again, most physicians I talk to think about AI in healthcare as sort of the robot, and I know we had some great talks on the robot, but it's not only about the robot, right? So it's also about AI in the sense of a special resource or energy that's going to be illuminating all the dark areas in medicine. So it's not only about the robot; the robot has some applications in AI... because I had a lot of homework assignments to do in the Stanford program. And as good as their database is at Stanford, often times at least 50% of the data cells are missing when we do a project. At least 50% of the data sometimes are missing. Unstructured data is our problem in healthcare as well. These are beach stones north of here, and it reminds me of healthcare data. This is a sign-out note from one of our nurse practitioners, and you can see there's a lot of information there, but it's very, very unstructured. Now, thankfully with natural language processing, or NLP, we're going to be able to solve some of this problem with unstructured notes. But wouldn't it be wonderful if the medical field and the nurses and the doctors can have a... recorded and plotted the pulse oximetry readings for me, that actually affected my decision that day to proceed with surgery for that particular child. So we really can't think of this as a silent event in terms of using data. We need to get the patients and the families involved as well. And organized data is escalating. Just reading material alone, the number of articles in the healthcare arena is doubling now every two or three years — doubling. So it's at a point where no physician is able to keep up. There's also the issue of future data, what I call data tsunamis coming. This is wearable technology. So the data tsunamis are actually fragmentation. So think of AI as a major, big rocket that's going to be able to pursue the moonshot projects we hear about in medicine, and think of data as the fuel to help launch these missions. What's happening is that each hospital, each patient is sort of hoarding their little gallon container of rocket fuel, and we're not able to fuel the rocket to go to great distances. So one thing we need to do is really get this data fueled together. And lastly, what's been really in the news recently is cybersecurity in healthcare. So a lot of the phobia or the fear is appropriate in terms of a problem. So the conundrum in healthcare is lots of data. It's not that we don't have enough data; it's how is the data going to be better structured, better formatted, more complete, and more understood and shared. So the second takeaway I'd like to leave with you is: data need to be improved and shared amongst all stakeholders, and that will make the intelligence part far easier and far more productive.

This is something that we're working on at Children's Hospital Orange County, which is a... you probably have heard of graph databases, which is a special type of database that we can... sites use. So you can say that healthcare is literally a decade behind a dating website, which is pretty sad to me. So this is a hypergraph database format that will be particularly well suited for healthcare because it's — think of it as a three-dimensional structure for a database, which is something that we're not used to seeing in healthcare. The reason why this is important is because deep learning, as you heard from the previous talk, is really now in a very robust form. And the way the computer thinks about playing the game Go that you heard about is, to me, very similar to how a clinician or nurse would be making a decision with a patient. So the dawn of deep learning is really going to make a tremendous difference in healthcare. And this is basically how healthcare will be eventually thought of: multiple layers of information and data, going from imaging to genomic information, all the omics, etc. So that is the future of AI in healthcare, which is essentially building a medical brain that would be the smartest and the fastest clinician on the planet in any subspecialty that you can name. So that's our sort of passion and goal: to build this medical brain so that sometime in the near future we will have most of the answers, if not all, since we're at an exponential conference in the last year or two.

So where is this money going and how are we going to use it in healthcare? This is sort of how I divide up all of the AI influence in healthcare conveniently into six different areas. I'm just going to summarize what's happening in each of those areas. Decision support is the traditional use of AI in healthcare, including hospital monitoring. So imagine all the ICUs in the hospitals now have access to all the ICU data from everywhere else in the country or around the world to make a better decision for the ICU physician. And that's already happening to some degree. And again, approach to AI and learning: every mistake will be recorded and analyzed and hopefully corrected when you're a patient in the ICU setting or even in a hospital setting. So this is AI in real use. We started something called the Rothman Index, which is an adult application of ICU data to predict readmissions as well as mortality. And this is live use of that data. So Dr. Annis, the Pediatric ICU director, can actually think about using that as part of his decision-making process. A huge area in AI and healthcare is medical imaging. If you think about medical imaging, most of the images that... in the near future, and as you know, the AI or machine-learned algorithms have now repeatedly superseded the capability of even groups of radiologists. I would be the first one to tell you I don't think radiologists will ever be totally replaced, but it would be a smart radiologist to accept the fact that AI is an ally rather than the enemy. So this is a cute little project I did at Stanford. I'm a cardiologist, so as some of you know, MRI images are great for quality of images, but echo is great for physiology. And in my mind, I always thought, wouldn't it be great if the two images could be sort of magically converge and with a programming language and help the clinicians to make a decision? Another big area that you heard about already is precision medicine, drug discovery. How do we individualize any therapy for any particular patient? Not just cancer, but even if you have an infection. Most of us are lucky because the antibiotics we get happen to cover the bacteria, but wouldn't it be nice to have a very individual recipe — pharmacological recipe — for your infection? Even a fourth area is the entire big area of cloud computing, big data. So when we think about AI in healthcare and why this is the perfect time for this area to really blossom and yield dividends, it's what I call the ABCD of AI in healthcare: A is for analytics and... elements have really contributed greatly to how AI is going to be an amazing force in healthcare. The most exciting 25 years in medicine are coming up. So I think you sort of feel that, that's why you're here. Digital medicine, wearable technology is pretty much starting. The amount of data that's going to be generated from all of the wearable technology devices is going to be very daunting without some sort of embedded analytics right into the device. So there's Internet of Everything on top of Internet of Things. Internet of Everything implies that there is primitive AI built into the device, and not just the connectivity that you hear... think an exciting area is not just robotic technology in healthcare, but the aspect of virtual assistance, smart assistance. So wouldn't it be wonderful if all of us have a healthcare avatar that's going to guide us day-to-day in our healthcare and make best decisions and remind us of things we need to do to maintain healthcare? The scary thought is for insurance companies to impose this on all of us, some sort of AI-driven inspired guidance system that if you don't follow, then potentially that can have repercussions on your premium payment, for instance. So I think it's better for us to take control of this.

All afternoon, these areas are sort of converging and growing exponentially. So the impact of AI in healthcare is going to be, I think, still being underestimated, just like AI in general has been underestimated all these years. The bottom line about AI in healthcare is you really need humans to drive this process. This is a wonderful group portrait of around 25 children's hospitals coming together to collaborate. And we should never forget the impact of eye-to-eye contact in this digital and AI world, how important that is to drive the process. So intelligence needs to be learned and orchestrated by humans. I just want to leave you with one... world. So one of the residents always has commented that, you know, 'Dr. Chang, you're trying to replace us with robots.' Well, that's not the point. It's to try to make us more human. So I say AI is great because it makes the visible invisible. So the point is not to have the robot, the point is not to have the computer on wheels, but to have the AI embedded so that we don't have to see something that's not conducive to patient care. And the opposite is true too: it's going to make the invisible visible, because we're going to be able to data mine and find signals in the noise. So if you're curious about AI in medicine, Daniel is going to be one of the keynot speakers for an upcoming meeting on December 12th to... I'd like to see you join us. I want to thank the Walt Disney Company Foundation for giving me this opportunity to pursue AI in healthcare, and my amazing colleagues at CHOC, as well as my very patient mentors at Stanford — probably glad I finally graduated — and my fellow computer scientists at Chapman University, and the many, many, many special people I met along this amazing journey of AI in healthcare. So thank you very much.