S4 #6 How to build trust in AI and digital healthcare

[00:00:00] KARTHIK: Do you trust chatGPT is the question that I keep getting asked, and my answer is simple. We have to be transparent about where we can trust it and where we cannot. If you know the answer already and can get what you're getting out of chatGPT, then it's useful. But if you don't know the answer, then you probably have to go through a series of steps to check if you can trust it or not.

[00:00:25] INTRO: Welcome to The Power of Digital Policy, a show that helps digital marketers, online communications directors, and others throughout the organization balance out risks and opportunities created by using digital channels. Here's your host, Kristina Podnar

[00:00:42] KRISTINA: Today's guest is Karthik Sourirajan, who leads ZS’s Pharma Commercial AI practice. He focuses primarily on the pharmaceutical and biotech industry, helping companies develop, evolve, and implement AI-driven product marketing strategies, tactics, and execution. His expertise areas include using AI and ML to enable sharper opportunity identification and targeting, leveraging primary MR, and other non-traditional data sources in AI/ML to enable more precise n=1 personalization, launch AI, voice of patient AI, and end-to-end patient support needs and intervention prediction using AI/ML. Welcome Karthik to the Power of Digital Policy podcast. I'm so excited to have you here today and ready to talk about all things AI, pharma, biotech, and just geekiness.

[00:01:39] KARTHIK: Thank you. Thank you, Kristina. I'm happy to be here and excited to talk about my favorite topics.

[00:01:46] KRISTINA: Well, look, before we dive in, I'm always very curious about people's backgrounds because I don't think it's really accidental that you landed where you landed. So, first of all, what inspires you to become a data scientist? Because it's such a highly niche area, you're not just a data scientist but a data scientist in the pharmaceutical and biotech industry.

[00:02:07] KARTHIK: Yeah. So, um, as a kid growing up in India, I had a variety of choices as to what I had to become, which is to say that I could either be an engineer or a doctor. So what happened? I ended up studying math and biology to figure out what I wanted to be. Within biology, there was a lot of genetics and then left the combinations of the patterns of Xs and Ys that came together to bring this to life, right? So my first real introduction to AI was when I was doing my undergrad in India, where we were applying genetic algorithms and neural networks for manufacturing problems, size production, and planning type of problems. And I had a lot of fun because it was about biology. I was, there was crossing of genes, there was mutating of genes, there generating new genes, and that was solving a math problem. There you go. That's AI to me. Then I went to Purdue for my Ph.D., and there I continued my AI journey to apply AI for the supply chain for automotive and high-tech. And there, I got introduced to AI in a different way, where one of my professors was working on a research problem of how do I cut out cancer tumors in the most efficient way. And in a way where it doesn't reoccur. That was exciting. Then I went to IBM research, where I was, again, applying AI and algorithms for high-tech and medical drug supplies, types of issues. But then I was introduced to the genealogy project where they were collecting the data that can then be used to predict a variety of different things for healthcare. That basically said, you know what? I think I really want to be in healthcare, and of course, I want to be doing something in AI, so I joined ZS, and over the last 10-plus years at ZS, I've been learning biopharma healthcare in a little bit deeper fashion. And over the last five, or six years, trying to take a little bit of an AI angle to it, and how do we apply AI and digital in biopharma? That has been my focus, and the idea is to transform how pharma companies engage physicians and patients from the point of view of creating better patient outcomes.

[00:04:34] KRISTINA: One of the things that I really appreciate that you just talked about is data, and you talked about it in terms of insights, but you kept saying data a number of times, and I think there's still this misconception when people say AI, they talk about it almost as if it's magic. There's like this notion of its magic. It's something that happens in a black box. It can't be seen or understood. Is that really the case, or are we really talking about data and understanding data usage? And there maybe there is some magic, right?

[00:05:10] KARTHIK: Let's talk about what AI is, and this is going to be a simplistic explanation, but that's kind of how I think about it, right? The traditional way of doing analysis is that you take data, you apply logic, and you get some output. AI is; basically, I have data, and I have the output for some situations. Now I have to figure out the logic or the underlying pattern. Which I can use to explain things, which I can use to predict stuff, or I can use to create new, right? So that's ai. So is there magic? Yes. But the magic is also, in a way, it should be explainable. And the reason I'm calling, talking about data a lot, is the AI fad has been there for a few different cycles. So AI has had multiple cycles where there are a lot of things we talk about AI, and then it disappears. Now, of course, with data and technology, it's coming back, but if I really look at what is the best AI, the best AI has two factors. It is data-centric, where it's less about, Hey, I'll just apply, and I'll go with them, but more about how you bring the data pieces together in the right way to solve a specific business problem. I'll give you an example. Pharma companies have data from patient claims and other sources. There is also market research that is conducted by pharma companies with the customers. Those are two data sets. Typically, you might say, well, I need AI. I need to use large data, so let me just use claims. But claims is not perfect. Market research is small and noisy. How do I bring the two of them together to create an insight for all the customers in the universe? That's an exciting problem, where you're engineering the data and putting it together the right way to solve a specific business problem, which could be which position is going to write the product or what they think about your product. The second part of it is in addition to data being data-set, I mean AI being data-centric, you also have to have AI that is invisible. The best AI is invisible. Part of the thing is there is the statement that I've heard, which is it should be like Tylenol. Nobody knows what's in Tylenol, but people trust it anyway. So we have to get to a point where the trust part of AI comes in. But AI is invisible. What is visible is what you get out of it. So yes, there is some magic, but the magic is constructed, right? Like, like anything, the magician knows what is going in. And how you do it is what creates better outcomes from AI.

[00:07:56] KRISTINA: I love that analogy. I think I'm gonna start talking about Tylenol more.

[00:08:01] KARTHIK: Yeah. I stole it from somebody. So happy to share.

[00:08:04] KRISTINA: There you go. You've heard it here first, but what are the key challenges you're seeing around sort of the data aspects and biopharma today that AI can help us solve?

[00:08:16] KARTHIK: The data available for all the customers for biopharma is broad but limited in its own way. For example, patient claims is a very important source of data for biopharma. Patients go to a doctor, the doctor files a claim, and that claim is recorded. And so you can track patient journeys along their continuum, right? But that data has gaps because it is collected at different places, at different level levels of granularity. Sometimes, I might not know what stage of cancer a patient has just by looking at the data because it's just not captured. So can I use AI to actually bridge those gaps? Sometimes the data is available in a different dataset. So I was talking about market research earlier. You could talk about EMR data as well, where the data is available. Can I learn from that? And then use that to bridge the gap in a different dataset. That's also AI transfer learning, right? So there are different ways in which AI can help in the biopharma world, and it is actually happening quite a bit because of pharma, AI, and pharma. They realize that it's not just about the ML models; it's all about the business problem and how you use the data to solve the business problem.

[00:09:40] KRISTINA: So if we're talking about data gaps because we don't necessarily have 360 views of patients and their claims, that would be helpful. Is there a place for synthetic data there, or is that even something we're thinking about?

[00:09:58] KARTHIK: Yes, so absolutely. So, I'll give you a classic example of synthetic data, right? This is sort of where in some ways, it started in the old world; there was this concept of systems dynamics and agent-based simulation where what you were trying to do was, and in manufacturing, if you had to detect failures and if you wanted to understand if certain things are working, you run a simulation, and the simulation is synthetic data because I'm trying to create widgets and I pass the widgets along a specific set of activities, and then I'm mimicking what happens in reality. How is that applicable to pharma and healthcare? If there's, there is an application where, let's say, I'm doing contracting and rebate with payers. There are; it's game theory, right? There are different reactions that can happen. There are different strategies that you can take. Can I simulate that through the use of synthetic data at digital twins and so on? So that's something that's basically applied quite a bit, right? It can also be applied at a patient level where I'm mimicking the patient journey to understand where are the gaps in the journey. But I may not have all the data. So then can I use what data I have to create, call it statistical distributions, a sample from that, and then run more synthetic patients through their journeys, understand where the gaps are, and then apply that to reality because the real data is not collected the right way?

[00:11:31] KRISTINA: When we talk about synthetic data, what does that look like from a privacy and regulatory perspective? And is it the case that we need actual data to create synthetic data?

[00:11:41] KARTHIK: Yeah. So synthetic data is typically created with a sample of actual data, right? Let me take a step back and talk about trust and data privacy in a different way, right? So, there is this aspect of consumers starting to get more involved in healthcare treatment decisions. So one part of that is as a consumer, there are many situations where I might want to share my data because I get something out of it. For example, let's say I have diabetes and I'm using a diabetic pump, and the medical manufacturer wants to track my data to ensure that if there's a pump failure or something happens, they're able to detect it properly. I might be more willing to share because it's important for me, for my outcome. Now, let's take a different example. Let's say that you are searching on your favorite search platform, and let's say I'm able to predict that, based on your search terms, you have pancreatic cancer. Should I tell you, or should I not tell you? If it's suicide help support, we find it imperative to tell the person that, yes, call this helpline if you need support. What about other tricky situations like cancer, other things about which you tell? What about depression? So these are the cases where you have to be very careful about what you do versus what you do not do when it comes to data privacy, right? It's a very tricky topic in some ways because there is no one rule that fits all. But fundamentally, it comes down to two things in my mind. One is informing people what data you're collecting and how you're using it. The second piece of it is, do you have the right systems in place so that the data security is maintained?

[00:13:42] KRISTINA: Who owns the responsibility around that? What if you have a device that's implanted in you and it's using AI? Can I block updates to the device? If I don't want them, for whatever reason, I just don't want updates anymore. I just don't want any kind of treatment for whatever ails me. Is that choice that I have any more? How far into those types of ownership decisions are we at this point?

[00:14:12] KARTHIK: I think the ownership has to be with the consumer. I actually strongly believe that, right? So I need to have a choice. And the choice aspect goes across different things, right? In the same way, I want the choice in what drug I take. I also want the choice in how my data is used, and those choices are very important for people to have trust. So the ownership should certainly lie with the consumers, and how you make it easy for the consumers is also important when it comes to ownership. Because one of the things that I've heard is that healthcare providers talk about data, and they have a lot of data. They just don't know how to use it because whom do I inform? How do I use it? And that part is also onerous, right? So that's another problem to solve. So the ownership, of course, is clear, but how do I use it, or what do I do with all the data that also needs to be thought about?

[00:15:13] KRISTINA: And so, how do we start to build trust in AI? What does that look like from a consumer perspective?

[00:15:19] KARTHIK: I'll take chatGPT as an example; which is the rage today, right? Do you trust chatGPT is the question that I keep getting asked, and my answer is simple. We have to be transparent about where you can trust it and where you can work. So one of the rules of thumb I've heard is if you know the answer already and you're able to get what you're getting out of chatGPT, then it's useful. But if you don't know the answer, then you probably have to go through a series of steps to check if you can trust it or not. And even OpenAI CEO talks about this quite a bit, right? It's evolving; it's learning. So there's a lot of messaging that is being done to say that, Hey, this is still not ready yet for prime time. But at the same time, it is already starting to have an impact. In many cases, when people are using it, for example, it can actually write code for you. And more often than not, that code is smarter and better than what you made manually. It becomes a co-pilot for you. So to build trust in AI usual approach has to be, again, transparency in how it is built, whether it's the data or how the data is used or what output you get, but also talk about it in terms of how do you maybe think about it as a co-pilot first and then sort of doing other things later, right? Even if I think about how doctors are using AI, doctors are using AI versus a co-pilot before; basically, other things happen. For example, the doctors get an alert saying that based on this imaging, this patient has a specific disease, they still have the opportunity to get it and say, okay, do I believe in this or not before they pass the diagnosis. That said, over time, many of these diagnosis have become better. I was reading the news about Mayo Clinic rolling out something where they're able to predict GI disorders faster there. They're able to predict which pre-diabetic patients are at higher risk of developing diabetes. All these are important things where they've tested it, the doctors have actually approved it, and now there is larger acceptance. So you have to go through those stages of trust. Trust involves explainability. Trust also involves other aspects that we have to bring together.

[00:17:46] KRISTINA: So I'm personally trying to reconcile this really crazy kind of thing where I can't send my doctor a message, like an email, I have to actually still use the phone to schedule an appointment, and I have to come into the office. With the fact that my doctor might be using AI, I'm thinking to myself like, how do I like literally help me reconcile this? Like, where are we?

[00:18:13] KARTHIK: You're bringing up a very important point, which is the proliferation of digital in healthcare; what we need is not digital health. What we need is digitally connected health. So there has to be this aspect of there are all these different ways I have to call the doctor. In some cases, I can do a digital engagement. In some cases, I can do different things. How do we bring all these pieces together into a connected healthcare ecosystem? And how do we bring the different stakeholders together? Biopharma, the way they look at digital might be different from how providers look at digital and may be different from how payers look at digital. But that collaboration and the closed loop aspect has to come into play for the patient to actually get the benefit out of this. Where I'm going with this is, in some cases, digital is helping quite a bit, whether I look at my trackers or alerts and so on. In other cases, what you are seeing is which is that I have to call the doctor because we have not connected the different pieces of the ecosystem well enough yet through digital.

[00:19:27] KRISTINA: And do you see different levels of connectivity depending on where you're geographically based? Is there more connectivity, for example, in the EU or China versus the US or in Israel? Or in Saudi Arabia versus Canada?

[00:19:42] KARTHIK: Yeah. So ZS did a study recently, and one of the things that came out is China has the highest adoption of digital. Let's think about why? The population size and the fact that there are long wait times to actually get to adopt. It was imperative for them to adopt digital to ensure that they were getting the right care at the right time. So you see that in China where whether it's telemedicine or user apps or other things, China is ahead in terms of adoption of digital. The U.S. is second on the list with other EU countries coming into play as well. So why does that happen to me for it to become digital and for digitally connected health to happen in addition to the technology, AI, and other aspects coming into play? We also need to be; the problem has to become a little bit more acute in some ways when people are like, yeah, I can actually do this better. A classic example is covid. We didn't have access to doctor offices during Covid, so everybody was okay with digital and virtual doctor visits. Now, virtual doctor visits are starting to become a thing, even though it's going down after Covid. Many companies are starting to explore how do I use telemedicine as an example to connect with patients, right? Telemedicine has a few different flavors, right? It's not just your PCP that you're talking to. There is this aspect of the digital front door where you are going back to our search term example. If somebody's searching for something or going to a webpage, can I give them a number to call? And then that basically translates into let me do the triaging first at one level before I take you to the right specialist. These kinds of things are also happening where you're connecting different pieces of digital to get to the right outcome.

[00:21:32] KRISTINA: So do you see technology, specifically AI and other digital capabilities pushing the boundaries to the point where I, as an individual citizen, can access healthcare from anywhere in the world? Or are we still bound by geographies? Will I get to the point where I can get faster, better treatment from the EU, from China, or will I continue to be bound by what's available in the US?

[00:22:01] KARTHIK: So in the near future, we are going to be bound by the US and potentially local geography as well because of two reasons. One is, of course, regulations, and there's also the aspect that different countries are in different stages when it comes to healthcare. So yes, you want to go there, but maybe they don't have the supply to serve the demand and also have something to think about. I mean, the classic example is that there are many people who come to the US for cancer treatment, right? So that is still possible, but not virtually. So that's the piece that we have to figure out a bit more, and regulations and other things are in place that sort of are going to make it difficult in the near future unless they change. The second aspect of it, in full honesty, always bothers me is the aspect of the digital divide, right? I, as a person, might have access to all the digital tools, and so maybe I can do it, but what about the underrepresented, underserved population? How do we create health equity if we go towards this model? That is a very acute concern, and we need to take a local approach to solve that, so it has to be a local approach so that it's not like the digital aspect is creating an imbalance in society in terms of how we provide healthcare as well.

[00:23:31] KRISTINA: Karthik, we always talk about opportunities and risks; that's what the power of digital policy is all about. What is the biggest risk you see right now for biopharma when it comes to AI and big data, and what is the biggest and biggest opportunity in your mind?

[00:23:47] KARTHIK: The risk in my mind is always about privacy and how do we manage consent and privacy. That, to me, is something that we always have to think about. The biggest opportunity in my mind, I'll take a different tack on this one. When you think about AI and digital in healthcare, this feels like, it feels like haves and have-nots type of approach that is coming in. I'll explain. When we spoke about the digital divide, where do you have access to digital or not and how do we ensure health equity, there's also the aspect of companies that are bigger and can provide more resources into it, maybe are getting much smarter, much faster, much more agile, much more efficient and gaining a competitive edged, and the smaller companies are still not able to capitalize on it because they don't have the investment. So how do you bridge that? Now let's expand that aperture a little bit, right? Where do you, where can you actually create an opportunity? In my hometown in India, there is an app that has come out called E-Paarvai, which basically means eyesight. So the idea is that somebody can come into your home and scan your eyes to see whether you need cataract surgery, and that enables them to figure out who needs it, who doesn't need it, and their source of sending them to the right specialist. So when you think about developing underdeveloped countries, are we focusing enough on AI and digital in those countries where some of these things can have a very, very big impact? China took the bull by the horn, and they're doing it now, developing underdeveloped countries. How do we do it there? How do we create better patient outcomes globally? I think there is the have and have not to divide that is also happening there. It's an opportunity and a risk.

[00:25:49] KRISTINA: Well, there you go. I feel personally challenged, and I hope the listeners do as well. But wonderful insights. Certainly, a lot of activity is going on in the biopharma space. Really appreciate you coming by today to give us your take on it, and certainly, lots and lots of insights here to think about. So, Karthik, I appreciate your time and look forward to chatting with you more in the future.

[00:26:10] OUTRO: Thank you for joining the Power of Digital Policy; to sign up for our newsletter, get access to policy checklists, detailed information on policies, and other helpful resources, head over to the power of digital policy.com. If you get a moment, please leave a review on iTunes to help your digital colleagues find out about the podcast.

‍