#46 AI in Healthcare, an Insider's Account (with Arnaub Chatterjee)

In this episode of Data Framed, a Data Camp podcast, I'll be speaking with Arnab Chatterjee. Arnab is a senior expert and associate partner in the Pharmaceutical and Medical products group at McKinsey and Company. We'll be cutting through the hype about artificial intelligence and machine learning in healthcare by looking at practical applications and how they're helping the industry evolve. Stay tuned for an insider's account into what has worked in healthcare, from machine learning models being used to predict nearly everything in clinical settings, to imaging analytics for disease diagnosis to wound therapeutics. Will robots and AI replace disciplines such as radiology, ophthalmology, and dermatology? How have the moving parts of data science work evolved in healthcare? What does the future of data science, machine learning, and artificial intelligence in healthcare hold? Stick around to find out. This conversation with Arnab was part of a webinar we ran, so the audio quality will be slightly reduced compared with other episodes, but the strength of our conversation will make up for that. I'm Hugo Bauen Anderson, a Data Scientist at DataCamp, and this is Dataframe. Welcome to Data Framed, a weekly datacamp podcast exploring what data science looks like on the ground for working data scientists and what problems it can solve. I'm your host Hugo Bown Anderson. You can follow me on Twitter Hugo Bown and datacamp at datacamp. You can find all our episodes and show notes@datacamp.com community podcast.

So Ana, as we've discussed before, I'm really excited to have you here today to talk about the role of AI, data science and machine learning in healthcare. What has happened and hasn't worked. But before we get there, I'd love for you to set the scene of your own journey and let us know how you got into data science originally.

Yeah, no thanks Hugo, and thanks for having me today. The notion of how I got into data science is quite serendipitous and as is a lot of things in life fit about right time in the right place. I think there's also sort of been a concurrent movement within healthcare where data science has really taken off within the last 10 years. So as a lot of perfect storms work, all these factors kind of aligned. So I think my career has been a little bit of a zigzag in terms of the fact that I've held roles in consulting, worked for the previous administration in both of the technology and the policy worlds, and then in pharma, and now back in consulting. But the central theme or ethos around how I've kind of worked has always Been around data science and the common thread and some link to data science. So just to kind of give you the quick background, I started my career after grad school as a consultant, originally focusing on pharma M and A, but then helping AIDS build health data infrastructure in a pre ACA time portal Care act time. And that actually led me to go work for the Obama administration. I transitioned over to initially work on some data science efforts around healthcare fraud and abuse and thinking not only about the policy, but also around how do we think about the utilization of that data to predict who who may be more likely to commit fraud and fraud, the government and things sort of transitioned into more of a technology bench from that perspective. I then had the opportunity to work with the chief technology officers within hhs and at the time, Todd park and Brian Sivak were sort of creating a new movement around open Data and building APIs and platforms to access this tremendous amount of data that the government sat on. Very, very fortunate to have run into two kind of Silicon Valley guys who brought their DNA over to government and kind of start a number of initiatives like the Health Data Initiative. We're building platforms around fda.gov, clinicaltrials.gov, open APIs. A lot of that was also very right time and right place and being able to kind of learn from folks who've done it in the private sector. And with that technology mindset and then the pharma aspect of it, I actually ended up transitioning over to pharma because of government and followed some folks up. I spent some time working in government and started a team that was focused on data science at Merck. And what we were doing at Merck was very much around how do we utilize and identify novel data sets which could include rote data like claims, but could also include clinical and genomic and social media and sensors to really utilize and think about how we demonstrate the clinical and economic value of Merck products. And, and this was a whole new way of thinking about positioning the drug in a different way and thinking about sort of the advent of new methodologies within data science to kind of support and bolster the value of the drug. And I did that for a number of years and worked with a variety of academic institutions on different machine learning methodologies, different data science methods. All of that eventually led me to McKinsey, where I am right now. And in that capacity I do a lot of work with not only pharmacliants, but technology clients and how they are entering the healthcare sector. So in that capacity I've been very fortunate to, you know, sort of be at the front lines of how different companies are deploying machine learning, data science and in a variety of different settings. And hopefully we'll cover a lot of that today.

Absolutely. And as you say, it took a lot of moving parts or it's almost a perfect storm of, you know, your interests and serendipity of technology in the emerging data science stack and all the data available that made this career path for you what it's become. And we'll see that that happen in this space as well. That it actually took a lot of moving parts, availability and generation of large scale data, computational power, statistical insights that kind of have allowed data science in health and otherwise to emerge as it has. I'm also interested in the fact that you started thinking about this type of stuff at grad school and that's when you started working with data. The skills you needed then and need now to work with data, are these skills that you learn on the job or were you trained specifically to do this, do this type of work?

So it's an interesting question. I think in grad school I spent some time in biostatistics and epidemiology. My background's in healthcare on the business side and then also health policy bent. The funny thing about data science is if you ask a lot of folks in health care, they'll tell you that a data scientist is a statistician in California and they'll basically say that that notion and term has changed a lot as methods and kind of, you know, the advent of machine learning has really warped, you know, what that definition means. But in some capacity, I think people who have worked with traditional claims data and have skill sets like epidemiology, like biostatistics, you know, have, are in their own right, you know, data scientists. And I think what's changed now is like you mentioned, greater volumes and different types of data. There are different ways of processing and understanding how we use it. So, you know, my training started there and then now I think like a lot of people having to evolve and learn just based on the fact that a lot of disciplines are converging. So computer scientists and the fact that there are actual data science degrees now, graduate global programs that are kind of teaching different ways of using data, those are converging with a lot of old school ways that people have used healthcare data. But that's kind of where I find myself right now.

Great. And I love that you mentioned different types of data because I suppose heterogeneity of Data is something which is so abundant and actually when people ask me, where does this occur? The clinical setting is one that I mentioned initially that you can have tabular data from experiments and controls in addition to imaging data from scans, which we'll get to, in addition to natural language from doctor's notes on patient files and that type of stuff. So I think it's a great time to move into this conversation about the use of AI in healthcare. So as we know, there's a great deal of hype around the use of machine learning and AI in healthcare. So I'm wondering from your point of view, what has actually worked in this space?

Yeah, it's a really important question and I think when we talk about what's actually worked, I think it's important to note that this is a very much evolving space. So in some cases the jury's still out. In other cases we're starting to see very promising signs. So just to kind of acknowledge the hype, I guess in my personal opinion, we're sort of in a golden era of funding for AI in healthcare. I think the stat that I saw most recently was that since 2016, there have been 300 startups in AI in health care alone that have emerged. And these companies span the gamut from reducing administrative scut work within insurance and billing companies to actually creating new drugs and empowering compound development. And I think what's important about that is we're starting to now see where a lot of the attention for AI is being driven towards and where the media attention is going towards. The venture dollars are flowing to companies that are also streamlining a lot of operational and efficiency tasks, administrative efficiency tasks in healthcare. They're also flowing towards companies that have these bold aspirations of upending processes that have been happening for decades now. So what we're trying to discern is what a successful outcome looks like. And then how do we think about what the bigger aspiration is? Where do we see tangible improvements in healthcare? How do we think about the advancement of patient outcomes? So this conversation isn't meant to put, it's not meant to splash cold water or put a wet blanket on the great work that's being done. It's just meant to try and understand, like you mentioned at the very beginning of the conversation, a lot of talk about this has been taking place. Where are we seeing the promise? So let me take on a few different examples and spaces where I think this has worked and we can kind of have a deeper conversation about this. You know, the very first thing you mentioned was around imaging, and particularly within diagnostic imaging in healthcare, that's sort of a bedrock. I think it's really important to remember that pioneering use cases of AI in other industries actually started with the ability to read pictures and look at faces and patterns and objects and photographs. That is very much similar in healthcare, where a lot of our lighthouse use cases where we've seen great success are starting to take place. To give you a little bit of context on like the growth of this market, AI facilitated diagnostic imaging is supposed to be a $2 billion industry by 2023. That is just a fraction of the whole medical imaging market, which is $57 billion. And $57 billion includes equipment, it includes software services. So it's a massive, massive marketplace that has been in healthcare for quite some time. And I think what we're seeing now is that the consensus from a number of parties, whether it's hospitals and technology companies, is that AI is going to transfer the diagnostic imaging industry, whether it's enhanced productivity, whether it's improved accuracy, personalized treatment planning, or all of these functions are up for grabs. So just to build a little bit more on this, why imaging as sort of the first place we're seeing improvement. For one thing, hospitals are producing around 50 petabytes of data per year, and 90% of that data is from medical imaging. So we're talking about MRI scans, PET scans, CT scans, and all of these are also embedded within EHRs. But I think that's kind of one reason is the availability and the ubiquity of this kind of data. I think the second reason is that there are a number of compelling use cases that are actually there now within healthcare. So to kind of pick on some great work that Google has done, Google Brain has published a really powerful paper within JAMA where they worked with 130,000 patients, you know, from an eye institute, and they looked at retinal fundus images, and what they were able to do was come up with a more sophisticated convolutional neural network that was able to predict diabetic retinopathy, which is one of the leading causes of blindness. Globally, around 400 million people have this disorder. So what they effectively did was come up with a more refined version using a subset of that 130,000 images. And they outperformed a panel of eight ophthalmologists in terms of understanding where the retinopathy took place. And how do you actually characterize what are sort of the contextual clues? Their f score was 0.95. The fact that they had a fantastic. Auc, the fact that it was in jama, there's quite a strong clinical argument to be made that, you know, if we get access to more of this type of data and we're able to build it into different tools and processes and how ophthalmologists see their patients, that's just the beginning, I think. You know, DeepMind has a very similar study within the retinal space. So just that those two examples alone I think are pretty compelling. And not only are you seeing this in ophthalmology, but you're also starting to see it within dermatology and pathology. Sort of your next set of lighthouse use cases.

I'm just wondering if it strikes me as interesting. And I'm wondering if it's interesting to you that it's companies such as Google in this case that aren't traditionally known in the healthcare space that are making such advancements.

Yeah, I think you're starting to see a lot of interesting collaborations between companies that have world class machine learning and healthcare institutions. So kind of on the other side of Silicon Valley, you have Facebook. Facebook just announced a collaboration with the NYU School of Medicine where they're utilizing their AI to speed up recognition of MRI scans. And the project for those of you who are Interested is called Fast MRI. It's looking initially at around 3 million images of the knee and the brain and the liver and looking around 10,000 different patient cases. So that was just recently announced. We'll see kind of what the fruits of that labor are, But I don't think it's all that surprising. You know, I think the computational power, it's now kind of incumbent upon Google to figure out how do they think about where the applications are, where the use cases are. And I think this is where you're starting to see imaging as sort of that first initial lighthouse for them because they can compellingly, they've done this in other industries and now they're compellingly able to do it with healthcare data as well.

So you were about to go in and tell us a bit more about use cases in ophthalmology and dermatology, for example.

Yeah, like I think we're starting to see similar, you know, the ophthalmology is obviously the retinal disease example. We started to see different cases with breast cancer. There is a great example of a partnership between Kaggle and Intel and a company called Mobile ODT where they developed an algorithm that accurately identified a woman's cervix and how do we better screen and treat women for cervical cancer. So the data was consisting of around 10,000 labeled cervix images and it had type 1, 2 and 3 cervical cancer. And this was a 50 layer convolutional neural network deep learning model that accurately segmented different parts of cervix type identification. So just another example where this algorithm, just by leveraging the power of the crowd, it wasn't even academically trained or clinically trained folks, they were able to capture and accurately identify cervix type 75% of the time. And so I think something to note here that's really important is that these CNNs are actually reproducible. You don't have to rebuild and reinvent the wheel every single time. I think that's where you're going to start to see great refinement and you're going to start to see a lot of enhancement in terms of how we do imaging recognition and reproducing these algorithms. I think the second thing is these major partnerships where you're starting to see tech companies partner with eye institutes and large corporations that have the imaging data that's going to be very compelling and powerful.

And so when you say reproducible, do you also mean reusable in the sense, in the transfer learning sense?

Yeah, I think we'll get to this, I think, later on, hopefully. But sort of one of the big challenges with AI is just making it reproducible within healthcare. So the big hurdle there is that data is different in many different parts of the healthcare system. So the patient that you're seeing in California will be very different than the one that you're seeing in Texas or South Carolina or Boston. And I think what we're trying to better understand is how do you create a generalizability on an algorithm that may have been used within a certain subsection of the U.S. population or the global population, and then being able to consistently come up with those networks, I think is what. Sorry, with those algorithms is what is a challenge because there are also different ways of characterizing this. And I'll spend some time talking about this later. But for radiology specifically, the outcomes that you're looking for are different. So you might look at the probability of a lesion, you might look at the feature of a tumor, you might look at the location of a tumor. You have to consistently do that exercise with different types of imaging data over and over again in order to say the algorithm is reproducible. So I think that's where we're starting to see. We have to continuously be able to prove that this algorithm is accurate and identifiable with other data settings.

Look, in all honesty, having this conversation makes me realize even more how important it is to demystify these things because particularly because there are so many touch points, right, of where AI can have impact in health, as you've mentioned, everything from administrative tasks to the scut work, the insurance industry to all of these diagnostics. And before we move on, I'm wondering if there are any other examples or super interesting use cases to your mind.

Yeah, absolutely. So the second, I think bedrock lighthouse that we're seeing a lot of is around diagnosis prediction. So how do you think about new variables that you haven't unearthed within data that might contribute to the progression of treatment? So we're actually, you know, working with several clients in this space right now on coming up with novel prognostic variables that could lead to the progression of a disease, perhaps predicting earlier onset of the disease. And I think, you know, what makes this compelling is, you know, there's still a great amount of misunderstanding. There's still a great amount of unmet need that we're not characterizing within our patient population. If we're able to better understand using machine learning methods on who those patients might be, we might be able to do incredible things in how we're getting them in and out of the hospital, seeing a provider quicker. A great example of this is actually Emory just released a study on sepsis where they looked at 42,000 patients and they looked at 65 different measures that might be predictive on the onset of sepsis. And they looked at it within different time intervals. So within 4, 6, 8 and 12 hours. And what was cool about this is that they were able to come up with the same model and the same level of accuracy as a physician in predicting sepsis. The validation cohort between the doctor and the tool, the algorithm was basically indistinguishable. So this is an example of not machine versus physician. It was more that we have the ability to not only confirm and corroborate what physicians are finding. If we continuously refine this, we might find more measures that are more predictive on sepsis. The one other example I wanted to share with you actually just came out last week and this was in the Journal of the American Medical Association, a pretty top tier publication. This was looking at a randomized trial of 500 some patients with staph infections. And they looked at patients over a six year period and they found that an algorithm did just as well as doctors in suggesting how to treat them with antibiotics. What makes this really compelling is that they were able to say that patients who were on certain antibiotic protocols, you know, were on a drug for a certain number of days, they might have been on a certain drug for a fewer number of days. You basically are looking at how we can think about antibiotic protocols and what is the best practice for keeping patients in and out of the hospital. So I think that's where you're starting to see a lot of compelling evidence. And given that this is now appearing in top tier medical journals, it's not, you know, something that's sort of in the future. These are things that are happening right now.

Foreign.

We'll jump right back into our interview with Arnab Chatterjee after a short segment. Are you inspired by what you're hearing on today's episode to share your data science expertise? Here at datacamp, we're currently looking for freelance part time instructors to build new courses and projects in data engineering, Python, SQL and spreadsheets. It's a great opportunity to build your brand and broaden your reach to the millions of Data Camp learners worldwide. On a more personal note, I myself have built five data camp courses and worked on about 20 others with external instructors. And to be honest, these have been some of the most fun and rewarding journeys of my professional life so far. If you want to partner with Data Camp to create the future of data science education, go to datacamp.com create or email createdatacamp.com with the subject line Podcast. I'm here for another segment on multitask learning. This time I'll be exploring ethical considerations around multitask systems with Manny Moss, an ethnographer of data science and machine learning who is currently a doctoral candidate in cultural anthropology at the Cooney Graduate center and research analyst at Data and Society. Manny has recently worked with Cloudera Fast Forward Labs as a writing fellow and contributed to their most recent research report, Multitask Learning.

Hi Hugo, it's great to be here. Thanks so much for having me.

It's a pleasure to have you here, Manny. So before we get into multitask systems, why are ethics important for thinking about data science and machine learning?

That's a great place to start. Ethics are important because what we do affects other people, which means we have a set of responsibilities to them. Doctors have an ethical responsibility to treat their patients as best they can without causing additional harm. Civil engineers have a responsibility to build bridges that don't collapse and injure members of society who have put their trust in those engineers by driving across those bridges or walking underneath them. And data Scientists have a responsibility to build algorithms that treat people fairly, respect their privacy and autonomy, and don't take advantage of them or their data.

Manny, I agree completely. So what does this mean for working data scientists?

For data science and machine learning, this means thinking carefully about who is affected by the models that get built, as well as what the unintended consequences of those models might be. In recent months, several helpful documents have been published to help think through the effects and consequences of machine learning. Integrate AI has put out a very helpful document for engineers to incorporate ethical thinking into the agile development method. The Institute for the Future has put out an ethical OS for anticipating and mitigating possible risks, and O'Reilly has published an ethical checklist for guiding an ethical development process as well. There's also very important work being done around fairness, accountability and transparency and machine learning to make algorithmic systems more fair and less biased. You can check out the FATML workshops and FAT STAR conference for how that work is moving forward.

Great. So how does multitask learning fit into these ethical concerns?

Well, multitask learning is susceptible to some of the same ethical risks as any other form of machine learning, because it's basically a way of learning to accomplish more than one task at a time. With a single machine learning approach, if it learns an unfair representation from training data, it can replicate that in decisions it makes about who is likely to repay a loan or show up for a trial or do well in a job. It can also reveal personally identifiable information to malicious attackers about individuals in the training data if it's not properly designed and secured. But because of how multitask systems can be architected, there are some very interesting ways to use multitask learning to address tricky ethical problems as well. And that raises its own ethical challenge if you think about it, because using an algorithmic tool to address ethical concerns can run the risk of having you think that you've solved ethics or used an ethical algorithm, when in fact ethical approaches to machine learning require thinking carefully about every step of the design build, deployment process and paying close attention to both the technical and social components of any system. Multitask learning is also susceptible to being thought of as a master algorithm that, because it can solve many things at once, should be used to solve everything all at once. And there are lots of good reasons to keep models siloed.

Thanks Manny, for giving us this survey of ethics and machine learning. Next time we'll talk about how multitask learning can actually be used to address ethical challenges that machine learning raises. After that interlude, it's time to jump back into our chat with Arnab.

And something you've hinted at several times already is that we're not necessarily. There's a false dichotomy between humans and machines, essentially. Right. And what is more interesting, I think, is the idea of human algorithmic interaction and the idea of having artificial intelligence and machine learning models with humans in the loop. And do you see this being part of the future of AI in healthcare?

So I guess this is your robots versus physicians conversation. Yeah, I think, you know, there are a few. I'll give you two funny anecdotes that are sort of perpetuating this theory of, you know, whether we, you know, we hear a lot of statements around whether physicians are going to be replaced by doctors. So one example is, you know, medical students are actually reportedly not specializing in radiology because they fear that the job market won't even exist anymore, you know, within 10 years. The other example is there's a really interesting company in China called iFlytech that is a pretty substantial Chinese AI company, and it was the first machine to pass a medical exam, and it scored substantially higher than the student body population. When you hear these types of statements and then you see all the JAMA evidence or the New England Journal evidence coming out of doctors being at the same level as a machine, there's going to be a lot of conversation. It also demonstrates how far machine learning has actually come. I think there's a few things that make me believe that we're not at a place where physicians are anywhere close to being replaced. You know, one is that a lot of these AI systems, like if you take the radiology example, they perform what's called narrow AI. These are single tasks and they're being programmed, and the deep learning models are being set for specific image recognition tasks. So like detecting a nodule or looking at a chest CT and looking for a hemorrhage. These are kind of n of 1 tasks, and they're sort of binary and either yes or no. And I think if we keep tasks in this sort of narrow detection focus, we're going to find a lot of interesting things. But what that means is that these are going to be augmenting tools. They're going to help physicians improve diagnostic accuracy. But as we all know, like physicians do quite a variety of tasks. There is a substantial amount of mental labor that goes into how physicians diagnose patients. In the short term, I think we're looking for AI to power a lot of solutions that can reduce cost and improve accuracy and augment physicians, decision making. I don't see it replacing the tremendous work that docs do or our providers do anytime soon.

Yeah, and I love that you mentioned narrow AI, which as you said, is algorithms, AI models to solve specific tasks. Because I think in the cultural consciousness, when people hear AI, they don't think of narrow or weak AI, they think of a strong AI, which reflects human cognition in some sense. And that isn't even necessarily what we want in most spaces or what we're working towards.

Right, yeah, that's right, exactly. It has to be more expansive. I think the other thing that's worth mentioning is that there has to be. We talked about this already, but the consistency and the portability and the models has to happen. So we're still a long way from integrating this into position decision making. I think different vendors are focused on different deep learning algorithms and on a variety of different use cases. So even certain things, we'll get to this. They're being approved by the fda, but they have completely different focal points. And until we can start to standardize a lot of that, you know, it's going to take some time. So, you know, to your point of narrow versus much more expansive thinking of AI, that's also part of the equation. And then how do we actually make this reproducible?

Something that you've mentioned several times is that a lot of the power we see now is from deep learning. And you mentioned the use of convolutional neural networks. I'll just step back a bit and for those people who want a bit of demystification of deep learning, deep learning is a sub, and correct me if I'm wrong at any point, is a subclass of machine learning, and mostly in supervised learning, where you're trying to predict something as a particular type of supervised learning model, which is inspired loosely from neural networks in our. In our physiological systems, in our brains. Convolutional neural networks are ones that are quite good at picking out patterns in images, essentially, and uses convolutional technique to do so. But of course, an AI predates convolutional neural networks. And although they're very, very strong at the moment, I'm sure you've seen trends emerge and dissolve. And I'm just wondering if you could speak to how the moving parts of data science, ML and AI have evolved in the healthcare space since you've been working in it.

Yeah, and I think, you know, if we're going to generalize it to healthcare, I mean, there's quite a bit. There's A whole variety of different models and sophistication of those models that are being thrown at different problems. So I think at its very basic level, a lot of early applications of AI in healthcare were focused on the relationships between diagnoses and medications. And some of the more basic techniques such as association rule mining or supervised learning, those were meant to come up and find and extract important associations. There were a lot of limitations to those methods. So I think if you look at ARM methods, they were only looking at item level co occurrence, they weren't really getting to abstractions at a higher level. You know, there wasn't a lot of usefulness for data exploration or for clinical decision support. And I think if you look at like supervised learning techniques, they are tackling these problems from a prediction perspective. So, you know, if we have the right level of data, we can come up with more non predictive applications. So things like disease group identification or patient stratification, these are things that can happen as data becomes much more usable, I guess, for lack of a better word. And then I think that's where we'll actually be able to see supervised learning become much more applicable. You know, going from small data sets with few observations into much more massive examples. So, you know, there's great work, for example, being done at Stanford and UCSF with Tool Buttes group where they've looked at hundreds of thousands of patients over 10 years of longitudinality, billions of observations and come up with sophisticated deep learning neural networks. I think that's where you're starting to see the far reaching applications of AI. In other cases we're still working on the data problem, which is that we're getting enough data to make this interesting, but the sophistication of certain models or methods may not be there because the data is quite frankly not that good.

All that having been said, what does the future of data science, ML and AI in healthcare look like to you?

Yeah, so I think there's quite, you know, there's, there's a lot of applications that we haven't talked about yet. I think we picked on the two easy ones in terms of what's happened and what's been working. Disease diagnosis, prediction and then on the imaging, I think there's quite a bit of work in terms of what's happening in drug development. So the fact that we are looking at companies now and there's exciting startups that are doing this that are focused on things like drug repurposing, where they're using real world data and machine learning algorithms to explore the relationships between drug molecules and disease that is extremely compelling. That's where you're starting to see a lot of funding go into, especially from biotech and pharma. You know, there are companies like Benevolent, AI and Numerase and others that are using deep learning to mine quite vast amount of data to look at everything from scientific papers, clinical trials. They're effectively just trying to understand which compounds can be more efficient at targeting disease. You know, these are the types of things that I think are getting quite a bit of investment, you know, but we haven't seen the fruits of the labor yet. I mentioned Benevolent. You know, they started identifying hypotheses around ALS treatments and you know, this is just a start, but you know, it's starting to narrow down which drug targets or which compounds to go after. It not only saves a tremendous amount of time for biotech and pharma, it also expedites the drug development process. I think that's one example. There are really interesting and powerful examples of genomic data that we haven't talked about yet. So, you know, Deep Variant, if I go back to Google for one sec, Deep Variant is an open source tool that was about two years of work between Google Brain and Verily, which is Google's life science arm. And what they effectively are able to do is come up with a more sophisticated statistical approach to pop mutations and filter out errors. And what Deep Variant does, it changes the whole task of variant calling, which is kind of figuring out which base pairs are part of you and they're not part of some kind of processing artifact and it turns it into an image classification problem. And the Variant is starting to replace and outperform these basic biology tools like GATK and SAM tools, and reducing the amount of error by up to 10 times. So this is just, I think in the beginning stages, even companies like Google will tell you that their genomics work is a couple years out, considering this one took two years to build. But I'm extremely excited about that type of potential. You know, there are other examples around physician burnout and the advent of voice technology within healthcare, where we're starting to understand that doctors spend an enormous amount of time on ehr, electronic health record data entry. And if we're able to use, you know, machine learning, natural language processing and voice technology in the future, then we start to auto populate structured fields within records, make the doctor's job less burdensome, reduce physician documentation burden. So those are three use cases that I think are sort of on the frontier. They're kind of areas that There's a lot of hype and interest and really amazing work happening. But that's kind of a short list of where I see the future going.

Great. And so before I go on, there are actually a few interesting questions from the crowd. And Gamal has asked the question, what about liability? Because. And I actually want to frame that in terms of thinking about the future of data science and ML and AI in healthcare. And in particular the fact that a lot of the algorithms we've discussed are essentially black box algorithms in terms of. It's difficult to see why they make the predictions that they do. So in terms of interpretability versus black box, maybe you can discuss that with respect to, I suppose, liability for the models that we as data scientists build.

Yeah, I think that's an incredibly important question. So one thing I want to talk about is the policy space and what's, you know, in terms of where the future is. This notion of FDA approved algorithms is actually starting to happen. So what we're seeing right now is this lack of consistency transferability in the current models because they focus on different endpoints. They are done in sort of a black box setting where it's data in. We're not really sure what comes out. And I think what that means is that regulatory bodies are going to intervene, albeit in a positive way. So as an example, the American College of Radiology is actually helping to provide methodologies for vendors to verify the effectiveness of the algorithm before they're taken to market. So I think that's one example. The other example is about supporting algorithms and approving them as part of diagnosis. So they gave a positive vote to a decision support tool that uses an algorithm for neurovascular disease. They did the same thing for diabetic retinopathy in April, and then they did something for a computer aided tool that helps with wrist fractures in adult patients. So these are all the FDA's permitting the marketplace to be open and they are allowing algorithms to actually help providers in a regulated way. There's actually really cool stuff happening within the White House and the House Oversight Information Technology Committee. If you guys are extremely bored, you should read this really ominous report called Rise of the Machines that the House Oversight Committee just put out. And it's basically, you know, how is the NIH going to ensure that there's standardization within algorithms? Same thing with the White House. They put out a really interesting plan from the government to build up AI in an ethical way. I think, you know, the black box problem is going to continue to happen. We've already seen it kind of be problematic for some large companies. We need to be able to address that. And although we don't love government intervention, I think this is one instance where we're actually seeing a lot of positive things come out of it.

So following up on this, we've actually got a great question from someone in the audience called Daniel about ethical questions in general. And the question is that ethical questions in data science have a much bigger impact on AI in healthcare. And are there moves towards developing ethical guidelines for researchers in this space? You've spoken to that with respect to top down. I'm also wondering about practice from within the data science community. What type of stakeholders will hold data scientists accountable? And also, I mean, the fact that, you know, in marketing, for example, or advertising, perhaps interpretive, like if you show someone the wrong ad, that doesn't have such an important impact as giving someone the wrong diagnosis. Right. So are there things that are, you know, particularly valuable in the health space? Yeah.

So I think what we're seeing is how do we standardize an ontology for a disease? And that's an evolving question. There are academic consortiums that are focused on reproducing these phenotypes. So a phenotype is basically how are we kind of characterizing a patient and their respective disease. If academic groups and organizations come together to say, this is a generally accepted algorithm, this is how we can avoid erroneous cancer treatment advice, or this is how we see this is an unsafe or incorrect treatment recommendation, I think that will actually compel more folks to work within certain parameters and build algorithms that are within guidelines and practice. Otherwise, it's incredibly easy to find signal within a lot of health data noise. And I think, you know, there are companies that have suffered some hard lessons that way. So I think as long as we are, you know, kind of working with organizations that are attempting to do this, that's one way of tackling this. I think the other thing is, you know, health data is incredibly inconsistent. And, you know, there's a national subcommittee called HL7, which is a health data standards committee. They are really making a strong push for something called fhir, which is a fhir, fast healthcare, interoperable resourcing. And it's trying to create a standard where data is no longer somebody's competitive advantage, but it's something that everyone can use and it's something that is standardized for everyone. So you're not just looking at inconsistent standards. So the Centers for Medicare and Medicaid Services are really trying to push standard ontologies. I think Fire and these other organizations are trying to create kind of a consistency behind all the chaos and the noise.

Fantastic. So that actually answers the next question we have from the audience, which is from David, which concerns research into the policy implications of AI in healthcare, and in particular will fhir have any impact on AI implementation? So that's great that you were able to answer that question even before I framed it.

We'll jump right back into our interview with ARNAB after a short segment. Welcome back to Machines that Multitask with Manny Moss from Data and Society and Cloudera Fast Forward Labs. Last time we talked about why ethics matter for machine learning, but we ended on a bit of a cliffhanger. Let's talk about what kinds of ethical problems can multitask learning help solve?

I should say at the outset that multitask learning isn't a silver bullet for any one problem. No single technique is. But multitask learning has some promising applications for a bunch of ethical challenges. Multitask learning can be used to train a model to accomplish multiple tasks simultaneously, as you could guess from its name. But it can also be used to train a model across multiple data sets. What this means for privacy, which is an ethical responsibility for data scientists, is that distributed multitask learning is possible. A distributed multitask model can learn from data sets that remain distributed and disaggregated. For example, several hospitals wanting to find patterns in surgical outcomes without exposing individual patient data could do so using a federated or distributed learning setup that trains a single model across each hospital's data. Learning locally at the edges of a network without transmitting data to a central server, protects individual privacy by not dumping it into a central, vulnerable data store, and keeps personal data from later being repurposed for uses to which there was not originally consent.

So does this also have implications for fairness?

It sure does. The ability to learn how to satisfy more than one task at a time with a single model is very promising for fairness. Now, fairness can be defined in myriad different ways, but generally refers to differences in decisions between groups that otherwise shouldn't be treated differently, like different genders or different racial groups. One of the major problems of fairness in machine learning is the trade off between accuracy and fairness. This problem arises from the fact that fairness aware classifiers, when applied to imbalanced or historically biased data sets, sacrifice predictive accuracy when prioritizing balance across, say, race or gender. This fundamental problem has been called the impossibility of fairness by Sorel Friedler and her researchers. And what this means is that and I'M quoting their work here. Different mechanisms for fairness require different assumptions about the nature of the mapping from construct space to decision space. Multitask learning can be applied here to optimize fairness for each protected class. It can treat the learning of an accurate classifier for each category, male and female, for example, as different tasks that can be learned jointly.

And how about interpretability?

This is another classical ethic program problem for machine learning, transparency and interpretability. If you can't explain or in some cases even understand what a model is doing when it runs, it's very difficult to say that it's running fairly or running safely. Multitask learning in some cases can operate as a tool for interpretability. This is particularly true for sluice networks, which learn how subspaces of a deep learning neural network should optimally be shared between multiple tasks. If you set these kinds of networks up in a way that lets them be inspected, you can gain insights into how tasks are related based on differences between the general representation a model learns to solve all the tasks it faces, the shared task agnostic space, and the way a model might also learn a task specific representation for each of the multiple tasks it faces. By probing these differences between subspaces, we can verify or challenge our intuitions about the data sets we're working with and reveal blind spots that we might not have anticipated.

So are there any other areas of ethics multitask learning can help to address?

That kind of depends on how expansive a definition of ethics you want to use, but I think that as long as research funding dollars are scarce, it's an ethical imperative to do as much possible with what funding we do have. And one thing multitask learning has been used for in this dimension is to help researchers augment new data they collect with prior data that might have been collected under slightly different research protocols. Now, given that informed consent applies to this use of data still, a researcher could potentially run a drug trial, for example, across a much smaller sample of the population. Furthermore, where this kind of trial might be potentially harmful as an intervention into human lives, that intervention can be minimized by learning as much as possible from as small a sample size as possible. And training a model to learn across new and old data is one way of minimizing that harm.

Thanks Manny, for a wonderful and thoughtful introduction to ethical challenges in data science and in multitask learning in particular. Time to get straight back into our chat with Arnab Chatterjee.

So another question I'll just remind people who who are here listening and watching today to send through Questions in, in the chat as soon as they come to mind. Got a really interesting question from William which we hinted at earlier. But William says, what I notice is that a big part of the hype is focused on the R and D side of healthcare. For example, image analysis, drug discovery. What are the hypes and more important, the promising applications on the manufacturing side.

Yeah, so that's a good question, I think. I assume that's in relation to drug development or, sorry, within pharmacos.

Yeah, exactly, yeah.

So I think we're starting to see a lot of kind of activity in this space. It's a little bit more nuanced, but in terms of how manufacturing is trying to tackle this, I think we are now in a place where the ability to standardize and create better understanding of how drug cycling take place, where the supply chain can be optimized. I think that's where, you know, there are companies like Berg for example, that is using, you know, AI for not only research applications but for manufacturing. So it's something that I, you know, come across less but something that, you know, is still popular. You know, I think there are ways to think about unsupervised learning methods in terms of how we're trying to understand drug circulation and where we can improve supply chain efforts. There is actually a cool effort from the UK Royal Society that is looking at the role of machine learning in biomanufacturing. So you know, can we actually help sort of optimize the time factor here, like helping manufacturers reduce the time to produce drugs, lower costs, improve replication. So yeah, still very much a popular topic. It's not something that we've avoided, but discovery is where I've seen a lot of the money and interest flow to right now.

So we've got several questions coming through about the ethical nature of using AI and ML and data science in healthcare. So the first one I want to ask you is from a listener called James. So James has written, given the black box nature of AI and the proprietary nature of industry, how will external validation and reproducibility of the algorithms be assessed? So yours is basically where does open science fit in to the commercial AI space?

Yeah, that's a good question. So I think what we need to do is come up with more of an interdisciplinary, multi stakeholder way of evaluating different algorithms that are entering the marketplace. So you have, you know, these large top down institutions like the FDA that are evaluating, you know, kind of the ability to, for a physician to utilize an algorithm in practice. I think there are other organizations you know, sort of at the academic level that are more interdisciplinary. So a great one is called Odyssey, which is the Observational Health Data Sciences and Informatics group. And what they are attempting to do, and this is actually a collaboration between pharma and academics and startups. And one thing that they've done that I think is really important is they've created a common data model for healthcare. So they've looked at disparate observational healthcare databases and decided that EMRs are really important for supporting clinical care. Databases like claims data are important for reimbursement, but they both kind of serve different purposes. So we need to create a common data model that accommodates both of these different types of data. And this CDM through a partnership called OMAP is, which stands for the Observational Medical Outcomes Partnership. It basically is trying to take all this noise from random coding systems and create one standardized ontology and vocabulary. So that's one way of trying to get buy in from other people, multiple players, interdisciplinary players. And that's I think, something that helps with the ethical challenge. Odyssey is an organization that actually works on reproducing and publishing all this research. All of it's open source. They've created a number of software tools like Atlas and Achilles that are standardized databases for profiling different data and for data quality. So, you know, this is not something that we're going to solve overnight. I think regulatory bodies are going to be very judicious about what they approve and what they don't approve. And what tends to happen in healthcare is that as soon as there's some kind of adverse event or there's some kind of clinical error, you're going to see the entire industry get slapped.

Right.

And nobody wants that to happen. I think that's what's stifling for innovation. So we are, you know, hopefully trying to. It's weird kind of the world that we sit in right now where we're trying to get as much AI related work out there as possible. Also being very mindful that a successful deployment of this, once it enters the physician's hands or it starts to be part of patient care is incredibly challenging. And that's kind of what the next evolution of all of this stuff is. It's the very careful implementation of these algorithms into a clinical decision model.

I'm glad that you mentioned regulation there because we have a related question from Stephen, which is the question is from a regulatory perspective, is it or will it be enough to prove, to prove the efficacy of an algorithm or model or is a fully full Functional description required.

Yeah, I don't know if we have proper guidance around that. I think there are a lot of organizations that are trying to demystify how the AI or how the FDA should be thinking about this. So, you know, a few things that are sort of standard in terms of like how much you have to prove something. You know, one is that, you know, what is sort of your benchmark data. And you know, are you using organization? Are you using ground truth data? Meaning is it kind of a trusted claims data source? Is it a standardized ontology for EHR data? I think companies grapple with choosing a whole variety of different data sources and collection methods and then they realize that their algorithm isn't that good or they're hoping for its approval. So, you know, there are, you know, good definitions on what like ground truth data is, and that's one way of kind of creating a really strong model. I think, you know, there are other ways of thinking about what's the intended use of the algorithm. How do we think about this interfacing with a physician? You know, does it have any sort of downstream patient implications? Is there going to be algorithm bias? Meaning are you going to deny care to a certain patient population? You know, that's something that the FDA considers in terms of how it considers it to be ethical or not ethical. And then I think there's a whole regulatory approach on refitting models so that they're continuously learning. You know, Scott Gottlieb, who runs the fda, has talked a lot about, you know, how will patients evolve and how do we think about when companies have to do a refit of the model and how it should be validated. What is the right refitting cadence? Is it every hour? Is it six months, Is it annually? So I feel like a few organizations have taken a stab at trying to create, you know, these list of guiding light questions that can help us come up with a good model versus a subpar model and one that's more likely to be implemented and accepted by the clinical community versus ones that have a great finding but may not some holes.

In them in terms of the regulation as well? We've got a great question from Harsh. Harsh asks that due to healthcare being a regulated area, do you feel that AI and healthcare will be a focused area for only big companies like Google and Facebook to be able to make advances in? Or do you feel small companies have a space? And you mentioned earlier that there have been several hundred startups in this space in particular, so maybe you can speak.

A bit more about that.

So I think It's a really interesting question. I think I'll give you one example that I think raised a lot of eyebrows. So to your point about big tech companies like Google, Facebook actually announced earlier this year that they are looking at using AI to monitor suicide and understand which one of their users might be more likely to commit a suicidal event. And that is a highly ethically challenging place for a tech company to play in. And I think the large tech companies, while they have great applications and great computer science, are treading very carefully because they realize that this isn't something that you can just wade into. I think they actually have a tremendous amount of organizational risk as they enter this market. One, because patient care is totally different than selling ads. And I think the ability for them to use their science and use the computational power comes at a great risk. They have to evaluate whether it's worth it or not. But they all want to make the world a better place and that's sort of their aspiration with other companies. I mean, there are so many powerful tech companies, like I mentioned voice technologies in this space. Companies like virtual assistants and voice technology are going to be major players in this space. And I don't just mean Amazon, but I mean companies like Orbita and others are doing incredible work, Robin. AI, they are basically, basically trying to help reduce physician documentation burden. And you know, these are well funded, well capitalized, strongly supported startups that are doing great things. There's, you know, patient data and risk analytics. You know, there's companies like Narrative DX that are doing really compelling work. They're partnering directly with healthcare systems to do this in a safe and compliant way. So I don't think, you know, large tech companies are the only ones that can play in this space. I think it's, you know, if you are very calculated and methodical about how you enter, if you engage in the right partnerships, a startup, I mean there are quite a few startups that have made a lot of headway in this space in the drug discovery world. These are startups that have raised hundreds of millions of dollars literally that are now kind of functioning quite well and successfully. Pharma companies are making multi year bets on companies like Numerate and Nipple and AI and investing quite a bit of money. This is not just a large tech company space anymore.

There are actually an overwhelming number of fantastic questions coming, coming through, but we're going to need to wrap up in the, in the next 10 minutes or so. So I've got two more questions. The first is from Christopher, which I, which I'M very interested in. Christopher asks, what are the main limitations on the rate of adoption of AI in healthcare?

Yeah, it's a good question. So, you know, we talked a little bit about, you know, what are sort of the policy hurdles and we've talked about how do we think about, you know, approaches. I think what, what is the biggest rate limiting step in healthcare is going to be the ubiquity of good quality data. And this is sort of the biggest challenge that I think has been plaguing healthcare for decades now is that, you know, as soon as a novel data set is released into the healthcare wild, everybody gets really excited about it. So, you know, as soon as the government made standards for EHRs, electronic health records, EHRS became sort of competitive domain for any organization that had the record, right? And the challenge was getting access to that data. Same thing with genomics. Now we're starting to see biobanks and the ability to have genomically sequenced data, genetically sequenced data. That is sort of the next domain. That's where people are sort of trying to get to. But none of this matters unless the data is linkable, unless there is a standard, unless there is labeled data. You know, we talk a lot about imaging today, but radiologists suffer from the fact that imaging data is stored within these pacs warehouses. PAX is the archiving system and then they're not labeled. So we don't know what we're necessarily looking at. And what that all goes to show is that the biggest hurdle for AI adoption in healthcare is good quality data. Which is why I mentioned standards like FIRE and others trying to create some kind of harmony and consistency in the data in a very chaotic world. I think the other thing is that hospitals and other organizations that have the data are very much willing to work with players. But there's quite a bit of overlap in terms of what companies are promising. We're starting to see a lot of companies tread into different spaces and say that they're doing development or they're looking at molecule identification or target validation and they're sort of trying to be jack of all trades. And I think that is sort of obfuscating what the company actually does. So my suggestion is just having a very clear, refined focus on what you think you are doing and what you're good at versus trying to then wade into a lot of other murky waters. With that said, the market is so hot right now that you will see a tremendous amount of partnership and opportunity. For startups. The biggest step is access to the data, finding the right partner being able to demonstrate a use case and then the application of that algorithm within clinical practice.

So I've got one more question from one of our listeners and then I'm going to going to ask you one final question. This is from Gamal and it's a relatively general question that I'd like you to interpret in whatever way you see fit. Do you think artificial intelligence will democratize medicine?

Interesting. I think that we will get to a place where I'm going to be liberal about the use of the word democratize. I think what that means is enable access to care perhaps, or maybe that's the definition that we'll choose to use. I think patients are increasingly interfacing with the health system in different ways. And the fact that the majority, the vast majority of patients go online to look up health information over almost 90% now, the fact that there is still there's a lot of ways for patients to share their data now with physicians, with technology companies. We all know the work that Apple's doing in sort of its work in Health Kit and Research Kit to try to get at more access to data. I think there is going to be some greater role for AI and maybe for technology to help with access to care. And hopefully I'm addressing your question, but feel free to rephrase. At the same time the US is suffering from tremendous endemic health policy challenges that I don't think AI will fix. I think AI will enable and help certain things. It'll maybe power diagnoses, maybe it'll improve better health outcomes over time. There's still a good portion of the population that will never be AI enabled for lack of a better word, or have access to healthcare resources. And I think that's the greatest hurdle in our system.

That definitely does answer the question. So my final question for you is for all our listeners out there, do you have a final call to action?

Yeah, I think we've talked a lot about challenges. We've also talked a lot about promise and, and where the industry is going. I think this concept of fixing a lot of low hanging fruit problems, we've picked on a lot of sexier things like drug development. But our healthcare system suffers from tremendous waste and these are enormous problems and AI can fix a lot of these things like insurance and billing claims. And my old mentors used to say that a lot of the most lucrative work in health care is also the least sexy and sort of the back office application type stuff. If we're able to do things like predict better waste or fraud or if we're able to improve billing and documentation processes, these are incredibly important problems to go after. And I think there's something there and you should use AI and your powers to go fix them. I think the other thing is that these problems shouldn't be done in isolation or fixed in isolation. You are going to see a lot of different and perhaps unique partnerships in healthcare take place. You know, hospitals and tech companies and patient groups working with startups and I think that whole model has flipped on its head. So I encourage everyone to be very inventive about how they want to work with different parties. There are a lot of non traditional folks, you know, that are getting into the healthcare space. So think about, you know, where the cross, the intersections lie and kind of where the cross functionality lies. And you know, that's where you usually find more inventive solutions as opposed to kind of working through the same channels.

Thanks Arnab.

Yeah, thank you Hugo. And thanks to Data Camp for the time. I really appreciate the opportunity.

Absolutely.

All right, thank you.

Thanks for joining Our conversation with arnab. We saw that fundamental advances are being made in healthcare due to data science, machine learning and artificial intelligence in the varied areas of diagnostic, imaging, diagnosis, prediction, the scut work of the insurance industry and pharmaceutical R and D to name several. We also cut through the hype to see that robots and AI aren't going to replace any major parts of healthcare in the near future, but that they will make practitioners jobs less burdensome. We also saw so much more, so make sure to check out the transcript online as well. Also, get yourself excited for next week's episode, a conversation with Peter Bull about the importance of human centered design in data science. Peter is a data scientist for social Goods and co founder of Driven Data, a company that brings cutting edge practices in data science and crowdsourcing to some of the world's biggest social challenges and the organizations taking them on, including machine learning competitions for social good. We'll speak about the practice of considering how humans interact with data and data products and how important it is to consider them while designing your data projects. In doing so, we'll discuss the role of empathy in data science, the increasingly important conversation around data ethics, and much, much more. I'm your host Hugo Bown Anderson. You can follow me on Twitter hugobown and datacamp@datacamp. You can find all our episodes and show notes@datacamp.com community podcast.