An Interview with Rafael Rosengarten, CEO of Genialis and Founding Member of the Alliance for Artificial Intelligence in Healthcare - Part 1 of 2
“There's no area of healthcare that will not be touched by AI.” “Physicians or companies who do not have an AI strategy in place will get left behind by those who do.” We've been hearing a lot of trite expressions lately. Will Artificial Intelligence be taking on a much larger role in healthcare and diagnostics like it has in other aspects of our lives? Dr. Rafael Rosengarten, CEO of Genialis, a data science and drug discovery company, and founding member of the Alliance for Artificial Intelligence in Healthcare joins us.
Joseph Anderson, MD: Rafael, unless you've been under a rock for the last couple years you would think AI is going to take over our lives. Certainly, in health care, we're hearing a lot about it. Just to get everyone on the same page, could you tell us, very simply, just what is AI and what would be some examples of it in everyday life that we're used to seeing but might not even realize are AI?
Rafael Rosengarten, PhD: Sure, and thank you for starting at the top. I will try to put it in simple terms but I'm not going to give you a short answer. I'm going to give you kind of the long answer. I want to do this because I'd like to tell people where my definition is coming from. I am part of an organization called The Alliance for Artificial Intelligence in Healthcare. We formed last year in 2019 and launched at JP Morgan. At the end of 2019, I had the privilege of collaborating on a fairly hefty technical white paper with some real industry thought leaders. And so, the definitions of AI, and its subsets that I'm going to provide, come directly from that white paper. The reason I want to do that is that the whole point of the white paper was to create a lexicon - something that we all in the industry of AI in healthcare could agree to use - the same terms in the same way, so we're all on the same page. If you’d like to read this, its at https://theaaih.org/publications
In the meantime, I can just answer your question. AI, in simple terms, is the study of artificial intelligent agents and systems that exhibit the ability to accomplish complex goals. I know that’s sort of circular definition; I used both the words “artificial” and “intelligence” in that definition. What we're really talking about is some sort of constructed agent that can accomplish complex goals in everyday life. Usually when we talk about “AI,” we're talking about some form of “machine learning,” which is a subset of “AI.”
The notion here is that this is the study of algorithms and statistical models that computer systems can use to perform specific tasks without explicit instructions. “Machine Learning” is a subset of “AI” that I think that we usually actually mean when we talk about “AI.” “AI” is actually a higher-level taxonomic term.
JA: A computer program, that can carry out somewhat complex tasks, or even an algorithm, which has branched decision making - are these “AI?” How are these different from the true definition of “AI?”
RR: In a machine learning system, the idea is that the computational programs and the algorithms will actually learn patterns. They'll learn this from some initial data set, but then they can apply these to future data sets. And if they're built in a certain way, they can continue to learn from future data sets, so that they can get “smarter” for everything that comes after that. This is in contrast to a standard computer algorithm that just does the same thing over and over again. It's kind of static or fixed in its capacity to ingest information and output product - even if it can do that at massive scale.
An AI system, on the contrary, is able to actually evolve in a way. It gets better at a task or at least changes the way that it does the task based on information it's already seen. An example of this that I see in everyday life is my Amazon shopping recommendations. Sometimes I joke that it's not very smart. I just bought a giant package of diapers and the next day they recommend I buy more diapers - they should probably give me a few weeks. Then they start sending you lots of other stuff for baby items - so clearly, they're learning my preferences. Netflix has done a really good job of making the recommendation engine a standard part of our lives.
Most content products on the web today are relying on some sort of AI to learn what you like and to try to sell you more of it.
JA: The idea of “Artificial Intelligence” has been around since the 1950’s - is that correct? And if so, why haven't we seen more progress?
RR: That's right. The term was coined in 1956 by John McCarthy. I don't know that it's fair to say we haven't seen much progress across the board. It's kind of come in fits and starts. Like virtually any other topic that has hype cycles - there have been summers and winters. Right now, people are really excited about AI in practically every industry. Everyone from Forbes to Harvard Business Review has articles about how if your company isn't generating an AI strategy, you will be left behind. But there are definitely winters too. There are times where you get laughed out of the room if you talk about AI. You have to come up with some euphemism for it so that you can be taken seriously.
I think that we've seen, in just the last decade and a half, a massive increase in the prevalence of AI thanks, in large part, to cloud computing - people using the internet to aggregate large data sets. I think, if anything, the limitations before were some combination of 1) computing only recently becoming decentralized and as powerful as it is and 2) the data. This is really, I think, the issue in the slowness of healthcare to adopt AI in a meaningful way. We can point to limitations in our ability to generate and aggregate the right kinds of data sets to actually be able to apply AI in a meaningful way.
JA: You mentioned the term “AI winter.” Many of us have heard of that term. Could you explain that a little more? Is it a real thing? Is it based on sentiment? Does it have to do with money flowing into the space?
RR: That's a good question. I don't think I've been in the space long enough to really understand the history intimately. In other words, I wasn't working in AI during the last winter, but I've spoken with people who were. My favorite anecdote about this was from Daphne Koller, a world-renowned AI scientist. She would call herself a “machine learning scientist” in order to avoid using the term “artificial intelligence.”
In her career, there was a time where she had to start putting “cognitive computing” on all of her grants instead, because “AI” couldn’t get funded. What's the chicken and what’s the egg? I'm not really sure whether it's a lack of funding from the public and private sectors that makes people start doubting it or whether industries just kind of sour on it. Anything where there's massive hype around some sort of change, where in reality, the change is going to be more incremental than the hype leads us to believe, can lead to some sort of disillusionment.
JA: In the last one to two years, it seems, we've definitely hit a high point in terms of hype. Can you explain how we got here?
RR: Good question. Where did the hype come from? I think - and this is in general, not just in health care - we can look at the success of companies like Google, Facebook, Netflix and Amazon. A huge amount of what drives their success is their ability to deploy AI seamlessly in the background - where you don't even know that's what's happening. Somehow it just knows what you're looking for and knows what you're thinking. These tech giants have come to dominate our everyday. They paved the way and showed us what was possible by integrating AI if you spent the time to get the data all in one place.
Both, the folks who hold the purse strings to funding new innovation, and also the people who want to innovate, see this huge opportunity to take what we've learned from these pioneers in web-based AI and to apply that everywhere. If you can make something work faster-better-cheaper, you might be on to something.
I also think, just in that way, venture-backed startups are prone to this hype cycle or any kind of hype cycle because the entire nature of raising that first seed round is going out and pitching a big vision for something. The very nature of it is to get out there and scream at the top of the mountain about this giant vision you have, which inevitably is going to be bigger than you can achieve over a short time span.
JA: We’ve all heard the clichés: “There's no area of healthcare that's not going to be impacted by AI… Those that are not using AI are going to be left behind by those who do…” Specifically, though, what you see in terms of applications in healthcare for AI?
RR: The space that I spend the most time in is, frankly, around drug discovery, which is only a small subsection of healthcare. In fact, I would argue it's almost adjacent to healthcare. We can talk about things like drug discovery. We can talk about things like drug development. Those are like two concentric spheres that a lot of smaller companies and new companies are playing in, because there's capital there. There's also then, of course, clinical care, which I think is really what you're aiming at.
Diagnostics is an obvious place where AI is already having a big impact and will continue to have an outsized impact. In diagnostics, the whole point is to take some measurement and have that be a surrogate for a much more complex biological state or disease or malady.
Maybe it's not just a diagnostic. Maybe you want to have a prognostic. Maybe you want to know how this patient is going to do. Or maybe you want it to be predictive. If we give them intervention “A” or “B” are they likely to respond favorably or adversely. Those kinds of questions are especially well-suited to a system that learns patterns from previous data and then can continue to learn as we iterate with new data. The whole diagnostic aspect of clinical care is already starting to be disrupted by AI and, I suspect, will continue to be.
JA: We are very much interested diagnostics. Could you explain to us how AI is going to take us to the next level? Let's say, in this current era, in the development of a diagnostic, you start with maybe a few clinical trials doing what you might be called “discovery.” You identify a set of a few hundred genes, that in a univariate analysis would correlate with outcome. That gene set can then get whittled down to maybe twenty or thirty genes and then get woven into a score or an algorithm, which then can do a much better job than any of the individual genes. Strictly speaking, it's just a score or an algorithm isn’t it? Is there an AI component to it? How can AI help us here?
RR: Well, just to get to that score, to have a high degree of predictive accuracy and to do it in an efficient way, we will employ machine learning at multiple points along the way. What you just described is actually something my company does a fair bit of, and you described it very well. The key here is what we want to measure for this patient or what we want to be able to understand better.
For more complex diseases or more complex conditions, where maybe the treatment options are also rather complex, where there’s just a ton of available information that is large and too varied for human eyeballs to do a terribly good job with on their own - to go from being able to measure either variation in the sequence or in the expression of say thirty thousand genes to something that's clinically actionable and could be put on a panel that's closer to thirty genes, it takes a huge amount of work. That work is made much more accurate and also much faster through the use of learning systems. That's an area where AI is already in use. Some of the pioneering companies in molecular diagnostics are already using machine learning as a tool. I saw a great talk by a scientist from Flatiron Health the other day. She wanted to make it very clear that machine learning is a tool - not a product - but it's absolutely essential to them in being able to parse millions of patient records in order to find patterns.
Foundation Medicine is another one that uses genetic variants and has quite a lot of data science behind it. I also think that with new high throughput data modalities, where you have things like single cell and spatial genomics, these are areas that because they're so fresh, are converging with the new algorithmic approaches. It will almost be native to these new data generation methods to have an AI system that helps interpret it. But even old school approach, like traditional pathology, is probably an area where machine learning is making a large impact already, because we're quite good at doing image analysis.
The idea isn't to make doctors irrelevant - quite the contrary. It is to return physicians to the job of treating patients and to take away a lot of the rote work. And I think almost everyone in the industry would say that's really the goal.
Continued in Part 2 …