HIMSSCast: Why hospital AI isn’t plug-and-play

Speaker A: For every use case, the impact of it needs to be very, very clear. And it needs to be clear before the use case is started.

Speaker B: Hello everyone and welcome to HIMS Cast. I'm Thiruk Gunasegaran, Managing Editor at HIMSs and today I'm honored to be joined by Vivek Rajagopal, Group Chief analytics and AI Officer at Narayana Health in India. Welcome Vivek. Thanks.

Speaker A: Thanks Thiru. Very, very happy to be a part of this discussion.

Speaker B: We're very glad to have you here. So I wanted to have this conversation with you because. Well, there are a couple of reasons. The first is HIMSS recently published an Asia Pacific AI Landscape report which has findings around where hospitals in this part of the world are at in their AI journey. So which are uh, the use cases that are seeing the most traction? What are some of the bigger pain points and what support is lacking maybe from their organizations? And also um, because Narana Institute of Cardiac Sciences, one of your hospitals, recently became the first organization in India to achieve the HIMSMM M Stage 6 validation. And for those who are unaware, the MM is our analytics Maturity Assessment model. Narayana Health, Vivek has had quite the journey with AI. There are a lot of things that you are doing well in, as we've seen from this recent validation. But from our subsequent interviews and conversations with you, I found that there are areas you're also looking at to take your AI analytics implementation to the next level. With that in mind, I wanted to frame some of our findings from this market research against your experience. So perhaps to start, could you walk us through um, that AI journey of Narena Health from the beginning, when this really started to take shape and what some of those early use cases looked like.

Speaker A: So uh, AI Journey is uh, rarely the start of a specific journey around AI. It always starts from what the business needs or what uh, the clinicians need. So that is the birthplace of any kind of digital journey of an organization. It was a similar start for us as well about a, uh, decade ago. Um, Narayana Health, from its inception in the year uh, 2000 has continuously been working towards bringing in efficiencies in healthcare and thereby bringing down the cost of care. Essentially affordable healthcare for all is our uh, vision and that is what we work towards. And everything we do is a path in our uh, way to get to this objective. So initially it was around process changes around how we bring in efficiency. That is where our first decade went. And coincidentally in the second decade technology had also advanced where it was significantly democratized the cloud was available. So Narayana was uh, probably one of the first organizations in the country to go fully on the cloud. In the year 2011. We were fully on the cloud before the major cloud players were in the market. We were already on the cloud with one centralized instance for the whole of our network. And uh, our data journey which uh, starts with uh, Atma, our EMR and our data intelligence and AI journey which is uh, the Meda AI platform that we run, all started around a decade back. Essentially these both were codification of all the process improvements and efficiencies that we had done over the years into a digital platform. So we don't believe that AI can be a one off initiative that you can suddenly start and then uh, uh, execute uh, very quickly. Uh, for your AI to be able to be successful you need to have a strong data platform. You really have a culture of data capture. You need to be able to aggregate all of the data in an analyzable form in a data lake or a data warehouse. And on top of that then your AI system can sit. That is how we believe that AI will be institutionalized. So that has been our journey over the last uh, 10 years. Starting off with data intelligence in business use cases and gradually going into clinical intelligence as our EMR adoption mature and then going into clinical AI. And today with the advent of Gen AI over the last 3, 4 years, we are into administrative and operational workflow AI as well.

Speaker B: All right, so we'll talk about meta later that you um, that you mentioned. But it sounds like what you're really describing is less about AI maturity and more about how data becomes part of how decisions get made day to day. You mentioned a bit about how you are also looking at operational efficiencies and also like clinical process improvements. So that leads me to my next question. In the study that I referenced earlier, 74% of respondents reported being still primarily focused on workflow optimization and 65% on documentation. But you've moved into clinical models like your cabg, risk scoring and ECG based detection. But I assume that wasn't an easy jump for you. What had to change internally to make that shift into clinical AI actually viable?

Speaker A: So firstly, I believe that uh, workflow AI and documentation level AI are still very, very good use cases that can have a real clinical impact as well. In addition to operational impact, uh, the moment clinicians are able to document accurately and effectively or extract information out of a document accurately and effectively, data is available at the point of care, at the point at which it is Necessary. All of it also have a bearing on clinical impact. But like you rightly said, developing core clinical AI models or clinical ML models is a different ballgame altogether. Uh, primarily developing core clinical AI models requires a lot more compliance to high quality structured data. You need the data to begin with. If your data is not properly organized, you can never build a clinical model around it. So the discipline of data capture, the discipline of capturing it in a particular format without any gaps of it being high quality is the first and fundamental, most critical barrier that comes into developing clients clinical AI models. Second thing, the gestation period for clinical AI models is very very long. Um, you will have to co develop it with clinicians, it cannot be developed independently in a lab and then brought to the hospital directly to be plugged into operations. It has to be co developed, which means you need to have a strong clinical team that has the bandwidth, the intent and the interest to be able to support the initiative where you're continuously interacting, providing feedback, iterating over a model, understanding where there are gaps, understanding the usefulness, understanding how it is actually applicable in real uh workflows. All of that needs to be there for you to be able to successfully develop. And even after having developed a clinical model m, the uh work doesn't end over there. How it actually integrates into the core workflow. No clinical model can function as a separate system. It has to be embedded into the clinical workflow. And finally you will have to go through all of the clinical validation processes. Uh, especially a model as critical as predicting a risk score and thereby driving action from that risk score needs to be thoroughly validated because any decision could be irreversible. So you will need to take time, six uh, months to 12 months of uh, uh on field validation where doctors are essentially using the model as an additional observer and not directly acting on it. So it is an audit phase where you are studying the performance of the model. And finally for the model to get into production, every country will have its own uh software as a medical device or uh, some sort of clinical decision support validation body. It will have to go through that as well. So for a clinical model to go from conception to final use, it realistically takes anywhere between two to three years. That is why it is a lot more challenging for the organization to not give up and follow it through to completion.

Speaker B: So clinical workflow integration is something that you brought up and this was something that came up in our study, in our responses, um, a lot of uh, our respondents specifically in the area of generative AI, a lot of them are using generative AI, 81% in fact, but um, 38% said um, these gen AI tools need more support, specifically in the areas like workflow integration that you mentioned, governance, uh, and training. Now with MetaScribe, which is your AI clinical documentation assistant, feel free to expand um, on this, on this scribe, you've embedded a gen AI driven capability directly into your clinical workflows. Can you talk about maybe that experience? Like what didn't work at first and what did you have to change or to iterate to make it actually deliver value?

Speaker A: Um, right. Very interesting. That was a, that was a fantastic project for us. And uh, it is still going on in terms of embedding it into every kind of documentation in the organization within the next uh, few months. We want to ensure that all documentation happens in an ambient manner. Not just the OP consultation, not just the OT notes, but any EMR documentation that happens in the hospital network. And um, it is a very challenging change for the organization and it is a very challenging project to develop as well. Um, the main roadblocks or problems uh, that we faced initially is to be able to understand what the actual accuracy of this system is.

Speaker B: Mhm.

Speaker A: How do you determine it? Um, and how do you determine the where the accuracy is breaking? Is it breaking because there was a poor transcript or is it breaking because post transcript processing, uh, the intelligence layer, there is some problem? How do you understand what was spoken, um, what was captured, what was the report that was presented and what was the final report that went to the patient? So it was not about the just developing the core system, but it was essentially about developing the governance around the system that was critical to solving this project. Once we were able to get a clear hold on accuracy and where accuracy is poor, is it in recognizing particular clinical terms, is it in understanding certain acronyms or abbreviations that people are using? Uh, is it in filling the report in a certain way or filling certain fields in a particular way? All of those were the initial challenges that we had to go through where we kept iterating with different ASRs, different kinds of uh, prompt engineering, different kinds of LLMs that we had to put together and different user interfaces that we were experimenting with. Those were the initial phases. As the uh project started to mature, the next level of challenges came in surfacing data that is required to be able to populate the report that is already there in the system. M It is not that people are going to start off a fresh report every time where they are starting from scratch. Very often they may want to refer to a finding in a previous report or just capture uh, the changes over a previous journey, uh, of the patient. So how to surface all of that as and when something is necessary and how to integrate it into the core workflow of the MEHTA scribe. Those were the subsequent, uh, challenges that were there. Finally, on how it embeds into the actual workflow itself. The scribe has to be there, but the scribe has to be invisible. The clinician or the clinical staff still need to feel that they are operating within the EMR while the scribe is somewhere in the uh, uh, uh, under the hood and it is operating and orchestrating everything. So those were the challenges. And uh, you need to have a, uh, very passionate and patient clinical team to be able to get your scribe off ground successfully. Otherwise it starts off with a lot of failures. People shouldn't give up, they should keep at it and that will reap very, very rich rewards once it stabilizes.

Speaker B: So you mentioned a patient and passionate clinical team. Did you have any pushback from your clinicians ever?

Speaker A: Uh, they do get frustrated. They do get frustrated because, uh, for us we are sitting in a software lab and uh, developing all of this. But for them, then there are hundreds of patients waiting at their door and then this causes a malfunction or this causes an additional uh, extra workload for them. That is not a desirable state. But being embedded in a health system, we are blessed with a passionate group of clinicians who go beyond their call of duty to be able to support all of this and make uh, healthcare more efficient for us.

Speaker B: All right, fantastic. So Vivek, I wanted to change tracks a little bit and get into the budget and finance of it. And this is a very important question. Specifically in India, um, in our report, about 47% of respondents told us that, um, or rather reported their AI budgets, um, over the next year was below US$25,000. And in a system like India, which is largely self pay and your cost pressures are very visible, many of these organizations are working under those constraints and they struggle to move beyond their pilots. For Narayana Health, clearly you've moved beyond the pilot phase for many of your AI initiatives. What do you think made that difference for you to actually start to be able to scale across the organization?

Speaker A: See, I think, uh, the biggest misconception is that scaling, uh, AI is primarily a funding problem. Uh, it is not, it is a structural problem that needs to be solved. Um, first thing, as I said, AI is a long journey. So if you haven't built up your foundation right, AI is never going to Scale. No matter how much funding you throw at it, it is important to take a foundational approach to AI. Ensure that your data landscape is properly structured and placed on top of which your AI can operate and scale. The second thing is it is important to take a platform approach to AI. Um, meaning, uh, we don't look at a, uh, single use case and see how we can execute this use case. Whenever we are looking at a use case, we are also thinking how we can execute use cases of this kind at scale. So we take a platform approach to AI. For example, if it is patient interaction with the health system, it may start off with one simple bot that we built. But we are thinking about what all bots will eventually exist in the patient journey and how a single framework can support all of that. The next is ROI discipline that needs to be there, um, for every use case. The impact of it needs to be very, very clear. And it needs to be clear before the use case is started. So before the use case is started, um, we co develop the impact framework along with the business function on what this will change in operations. Is it going to improve efficiency? Is this going to give a direct financial gain for the organization? Is it going to improve clinical outcomes? Is it going to improve patient experience? What is the baseline today and how are we going to measure the improvement? That is very, very important at the start of the project because that will help us prioritize meaningful projects over experience. And once meaningful projects, uh, are prioritized for us, it is very important to track it to completion and to track the impact around it. That is very critical. Once you start demonstrating success with your AI projects, then there is no reason that the organization is going to be hesitant or going to put a stop to the AI initiatives. And the last one is at a pilot or a POC level. Many of these AI projects look very, very attractive or cool. But uh, the real work comes in again on how well we are able to embed it into a workflow. It can't exist as a standalone system that way. It will never scale. It has to embed M itself into the core workflow and the associated change management has to happen in the organization. That is what is critical in scaling AI.

Speaker B: It sounds like it's a very strategic, almost scientific approach that Narena Health has to not just AI, but digital transformation in general. And you talked a bit about tracking, tracking those, um, like tracking the end goal, like looking across at the entire journey. So on that note, let's talk a bit about ROI. Right. Um, our data shows 80% of respondents are uh, already seeing improved operational efficiency from AI. But far fewer of them are actually able to clearly tie AI initiatives to their uh, clinical or financial outcomes. And in a previous interview with us, Vivek, you spoke very candidly about the need for better m, better outcome attribution in Narena. How difficult is it in practice to actually prove that this AI intervention is driving those results?

Speaker A: Attribution is one uh, of the hardest problems that is there in AI because uh, no matter how well you design an experiment practically, there are several factors that uh, influence an outcome. You cannot simply say that I did this initiative and therefore revenue increased. Because there are several factors that influence revenue. Uh, it can be things that are strategic in nature. It could be capacity, addition, it could be something that uh, reflects the market scenario, the competition landscape, or several things. So attribution is extremely complex, which is why, uh, we will need to understand how we can frame the problem at a very, very granular level. Instead of saying that this will improve revenue, which is a financial metric, will we be able to say this will improve revenue through this particular channel for people coming from this area within this specialty, Then it becomes very, very pointed and it doesn't get muddled up in uh, other set of uh, factors. But having said that, it is still practically very, very challenging. We've built our own internal framework around impact measurement on what all perspectives you need to look at to be able to quantify impact. Starting from man hours saved to um, how you are able to quantify the efficiency that it brings to decision making in terms of earlier interventions or in terms of more frequent interventions that you're able to do, or impact on data quality, impact on informed decisions that you are able to take. And finally direct KPIs and indirect KPIs where we try to project it to a financial metric. Uh, but uh, many times projecting some of these to a financial metric purely becomes an academic exercise. I think the organization and the leadership buy in also has to be there in what they are trying to improve. If for example, we are trying to improve something like a patient experience or efficiency, or we are adding capacity to the organization by improving throughput, those are things that need to be understood at uh, a leadership level as well that we are on the same page, although the financial projection of it will just be an academic exercise, but this is what we are actually bringing to the table. So it is a ah, complex process. You need to have some science behind it, but you also need to have your leadership be aligned to it strategically for it to actually come together. Right.

Speaker B: So it's, even though you mentioned it's a scientific process, it's, it also sounds like a very. What I'm hearing throughout all your responses, it's a very practical approach to AI. And you also mentioned about how you shouldn't just be looking at whether the technology is cool. Um, there's a lot of momentum around AI right now, but there's also a lot of noise. And based on your experience, what do you think, uh, healthcare organizations or hospitals overestimating and what are the underestimating in fact, when it comes to scaling AI

Speaker A: in healthcare, as we spoke earlier, I think what is being overestimated when we speak to organizations is the speed of AI transformation. Uh, especially in the clinical settings. As we spoke about, the gestation period is very, very high. So that we will get one AI model today and in one month everything will change. That is not going to happen. It takes a lot of discipline and change management to be able to uh, stabilized it successfully. Um, the second uh, area where I think uh, there is overestimation is the role of the model alone. Um, where we feel after we've cracked the accuracy of a model that uh, the core work is over. But actually that is where the rest of the work starts, which may be 50 or 60% of the total project. So building the model alone is only half the work done. Rest of it is how you're able to integrate it into the workflow and get people to use it. Um, what is being underestimated I feel is again reflecting back to the first point that AI can be picked up and plugged on. The ease of implementing uh, AI I feel is being underestimated. Uh, the importance of, of data and governance as a foundation is what is being underestimated there. That is very critical for us to be able to run successful AI projects. And the difficulty of workflow integration as well, how it naturally and seamlessly embeds itself into the core workflow of a person as a standalone product, it may be great, but if it is operating in a silo, then it will never get picked up. Um, third thing is around change management. That is a discipline in itself. Organizations and projects where the change management team is very, very strong, we've observed that those take off more than the others.

Speaker B: Throughout this whole conversation I think you've been very consistent with what, what needs to work. It really isn't about building AI or building AI maturity. It's getting to work, uh, inside real clinical workflows, um, which takes more time and discipline than many people expect. I think discipline is something that Narayana Health has been hammering on. Vivek, it was really nice speaking to you today. Thanks so much for joining us. And um, it was really, really, really nice to hear from the perspective of someone who has had, I think, what that 10 to 20 year, uh, journey. It's like two decades long journey. And also an organization that has achieved one of the highest standards that we have, which is the stage six for the hims. Mm. So thank you again Vivek.

Speaker A: Thanks. Thanks Tiru. It was fantastic interacting with you. Great, great questions and great conversation.

Speaker B: And for those of you interested in reading the report, what I mentioned earlier, you can find it on our HIMS website or through a quick web search. It's the AI landscape in APEC Healthcare 2026 report. And if you like what you heard today, feel free to check out our other episodes by subscribing to our podcast on Apple Music, Amazon Music and Spotify. Have a great day.