#262 The Hard Facts About AI in Healthcare Interpreting with GLOBO CEO Dipak Patel

0:00Healthcare is a completely different world and language services need to figure out what its role is in this new space.

0:09All right.

0:10Dipak Patel:Welcome to SlaterPod again, another episode. Today, we welcome back Dipak Patel. For those of you who watched our emergency executive order episode a few months ago, you know Dipak. So he's the CEO of Globo, fast-growing language solutions integrator focused on healthcare, language services, tech, and AI. Hi, Deepak, and thanks so much for joining us again today. I'm happy to be on. Excited for this conversation. So again, for those who don't know where you're based, what part of the world are you recording this from today? I am out of Boston, Massachusetts, 20 minutes outside of Boston. Boston, Massachusetts.

0:44Dipak Patel:All right. So quick career recap, what got you into the world of language technology, language solutions, language AI? It's been an interesting journey. I'm relatively new to the space for you. Engineer by background. Started my career in consulting at Accenture. I was a managing director there for about 15 years. Consulted across healthcare. Worked with hospital systems, insurance companies, pharmaceutical companies on everything from launching new products to acquiring other companies to revamping their supply chain capabilities. Spent a lot of time in the U.S., but I spent about eight months in Africa consulting for nonprofit and then about five months in Europe consulting for biotech.

1:22Got tired of advising companies and them not always listening. So I went into the world of private equity. I was an operating partner at a private equity company that was investing in the revenue cycle space, which is all about how hospital systems bill and collect money, et cetera, et cetera. Did that for about seven or eight years, made a number of different investments. still something was missing quite honestly it's just at some point I became a dad of two kids and I was always thinking when it's bring your dad into you know school day can you really explain what private equity is can you really explain what consulting is you know kids be you're following up the firefighter and then you're talking about private equity it just doesn't have the same ring and so I said I wanted to do something that one you know my kids would be proud of what I do two is hopefully do something that moves health care in the right direction and then three something that was easy to explain.

2:12And so I was actually looking at a number of different options. I met Gene Shriver, the founder of Globe, all this language solutions company. I had no idea what it was. The only experience I had was I'm first generation. My mom was sick for about eight years, eight plus years, and she spoke English, but Hindi was her native tongue. And what was interesting was whenever I went with her to go visit a doctor, there were times where she would not get an interpreter and the experience was terrible. Times where she'd get a bad interpreter, the experience was worse. And times she'd get a great interpreter to make a world of difference.

2:47But I had no idea this world existed until I met Gene. And that was in 2021. And eventually Gene asked me to come on and help run Global and take it to its next level. So yeah, I've been in this industry for now for about four or five years.

3:00Dipak Patel:Wow. All right. So five years ago, we had Gene on the podcast. So maybe tell us a bit about like top three, four, five things. What has changed since we had Gene on the pod? I just checked before it was September 2020, so in the middle of the pandemic. So yeah, top three things that have changed at Globo. A couple of things. One is we've doubled down on focusing on healthcare. That was one of the reasons why I wanted to join Globo. I thought there was an opportunity to really evolve the way hospital systems use language services to create a better experience for their patients and clinicians. Number two is, you know, I asked this question when I started at Glovo.

3:41I said, can't you just use Google Translate to do the interpretation? Again, I was new to the space, Florian. And he was like, no, that makes no sense. You'll get it when you actually watch what our interpreters do. Technology has evolved since then, which I know we'll talk more about, but the coming of AI and using LLMs has made an impact in the industry and will continue to make an impact in the industry. And then I think the third thing, at least specifically in healthcare, is the industry is facing a lot of challenges in terms of just financial pressures, the fact that we're going to have a shortage of clinicians coming up in the future, the fact that we have an aging demographic population and we're trying to keep them healthy.

4:21And so I just think healthcare is a completely different world and language services need to figure out what its role is in this news space. So those would be the three things that I think are different from 2020.

4:30Dipak Patel:Also, like you've moved, obviously, very much towards tech. So you have what you call now platform and product. So let me just list the three that I see on the website. So Global HQ, Connect, and Kai. So tell us a bit more about that before we start talking about the AI product, which I'm really keen to speak to you about. HQ is essentially our platform where you can access our services, But really, the point there is for administrators to access reporting, data, maybe if they want to do translation projects, things like that. So think of it like our back-end administrative platform. You got Connect, which is the actual app that sits on the iPad or the phone or the Android tablet where you can access interpretation services.

5:13And then you have Kai, which is our AI interpreter, which we, yes, launched earlier this year, which we're very excited about.

5:19Dipak Patel:Before we go to cut, just briefly for those who are not in the interpreting space, not in the healthcare space, who would be a user slash client slash decision maker for you? Kind of a key important user that you have. Because especially for the Europeans, they sometimes don't understand just the size of the U.S. healthcare system because it's all kind of local here. I don't know, local municipality runs a local hospital, et cetera. So how does this work in the U.S.? And who would be your kind of core user slash decision makers? So there's about 5 ,000, 6 ,000 hospitals in the country. And then there's a number of independent physician practices, and there's a number of what we call outpatient clinics settings.

5:59Usually, language services will report up to, if they're lucky, a dedicated language service manager. Sometimes the hospital system doesn't have a dedicated person. Sometimes that person is wearing multiple hats. In terms of who influences the decisions, it's multiple people. It's everyone from that specific language service manager to maybe it's whoever's running procurement to whoever's running experience. So we have a role here in the U.S. of the chief patient experience officer, which is all about trying to figure out how do you improve the experience of patients. And then obviously clinicians have a huge say at the end of the day.

6:36And you know this stat, roughly one out of four, one out of five people in the U.S. choose to speak a language other than English at home. And so those are the folks that really get help from offering interpretation services. And you think about it from a physician standpoint, you know, it's anywhere from 1 % to maybe 5 % of their patient panel where they have to use these tools in order to better interact with these patients that want to communicate in a language other than English.

7:06Dipak Patel:And currently, I guess, a substantial part of your business remains with experts, linguists being patched in into these conversations, while at the same time, you are piloting certain AI capabilities. Would that be fair? That's fair. I get this question all the time. Do you think AI is going to completely disintermediate this industry? And there's a world where you could use AI for interpretation. Um, I'm not going to say no, because I have no idea what the future is going to look like 10, 20, 30, 40, 50 years from now. I think that's impossible to say. I will say that we do not think that there's a world where you're going to be able to use AI for every type of interpretation here in the future.

7:51There's always going to be a role where you're going to need a human interpreter. It's a more complex type of conversation. Maybe the person is dealing with mental health issues. Maybe the person is wearing some sort of apparatus where you can't understand what they say. Maybe it's just a complex conversation where you can't risk using AI the way it is right now. One of the things I learned, Florian, over the last five years, and I think most people outside this industry don't have an appreciation for it, I surely didn't, is an interpreter does more than just interpret, as you know. An interpreter serves four roles.

8:24Yes, they serve as a conduit from one language to the next. They also serve as a clarifier. I saw that when my mom was speaking to a doctor, and the doctor would say something, and that could mean two or three different things. So before the interpreter interprets, they ask for clarification from the clinician. They serve as a cultural broker. Many times the doctor would ask my mom to do these things or eat this type of food, and the interpreter would interject saying, doctor. The patient's not going to do it based on their religion or background. And then they serve as an advocate, which is really important.

8:55They help the clinician understand if the patient might not be speaking up or something is off. And the reason why that is important is when you think about AI's role, AI needs to do more than just serve as a conduit from one language to the next. It needs to be able to serve in all four of these roles. With all that said, we view things from the lens of the patient journey. I think about, again, my mom's journey. It wasn't just whether or not she got a good interpreter when she saw the clinician in the ER at the hospital. It was when she was checking in at the front desk, or maybe she was admitted into the hospital and you have a nutritionist come and try and have a conversation about what to eat.

9:35Or it was when my dad got the bill afterwards and was trying to make sense of it. And if you look at the patient journey, There's gaps in terms of when we should be providing some sort of language service, whether it be interpretation or translation or what have you. That's not happening right now. And so our view is you could use technology to fill those gaps.

9:56Dipak Patel:That seems like what you built Kai for maybe now, like to fill those gaps. Maybe tell us a bit about first launch. I think you launched it within the US. Tell us a bit about language coverage, primary use cases. You just outlined a couple. You also ran a pilot and maybe something on kind of cost impact that they found from that pilot, from those results. We've been on this journey, just to be clear, for two years. Right when ChatGPT first came out, we wanted to do research in terms of how good is off-the-shelf LLMs for language interpretation. And so what we did is we basically ran it through medical interpretation scenarios.

10:37And then we tested how good were the results using medically certified interpreters as well as third parties. And we published those research findings. I think it was a year ago where it showed it wasn't very good. And this was for Spanish, by the way, not the 300 plus languages and dialects across the U.S. It wasn't very good. However, what it did show us was it created a roadmap in terms of some of the things you would need to do work on in order for it to work. And so we've been on this journey for a while. And then last year, we finally finished developing the prototype for Kai. We wanted something where we could actually test it in a real-world setting.

11:16And then earlier this year, January 1st, we started to do pilots with a number of different hospital systems across the country. I think now we're on number 10. And in these pilots, what we do is we're deploying Kai in the app. And then we have a medically certified interpreter in the room while the clinician or person is using Kai. And the medically certified interpreter can do one of three things. One is they don't interject because Kai got everything perfect and there was no need to interject. Two is they have to interject because Kai got something so wrong that it changed the course of the conversation.

11:54And that's the last thing we want to do. or three is Kai got something wrong, but not to the point where it impacted the conversation. We found the level of acceptability. And the reason why that was important was, it's like, I get this question a lot too, you know, how is AI ever going to be good as a human interpreter? And the question I usually ask back is, well, there's no standard way of assessing what an acceptable human interpreter is, at least in the US for healthcare. We know that there's great interpreters. We know interpreters that are not so good. What we don't have is a standardized way of assessing quality across every interpretation.

12:33And so I talked about my previous experience. I used to run a revenue cycle company on medical coding. And within the medical coding industry, there was a standardized way of assessing the quality of medical coding, meaning you would review X number of charts. This was the formula you would calculate to determine if the chart was correct or not. And then here was the score that meant it was acceptable or not. There's none of that in the US for healthcare for interpretation. And because we don't have that way of assessing what is acceptable for interpretation, it's hard to make the comment, is Kai or any sort of AI as good as a human?

13:10So the way we're approaching it is by running systematic pilots with multiple hospital systems across the country, where we're actually using a medically certified interpreter to assess how good CHI is. And then on top of that, we are surveying both the patient and the staff after using CHI. And we're doing it in the form of the MPS, net promoter score. And so I think we've interpreted somewhere between 4 ,000 to 5 ,000, you know, we call them messages across almost a thousand sessions. And our net promoter score is 60, meaning they, and it's crazy. That's a really high net promoter score. So people are happy, people are acceptable.

13:49We're doing it again in sessions where it's administrative in nature, it's not clinical, it's use cases where typically you're not using any sort of interpretation service.

13:59Dipak Patel:In a sense, it's still in a pilot phase, right? So when would you expect it to become much more widely available and maybe with a bunch of things you need to click and to agree that maybe that certain preconditions are met when you're using the app. But yeah, when do you expect kind of wider deployment of this? We already have three of the hospital systems that have expressed interest in wanting to use it. Come September, you'll officially be able to use Kai, meaning we'll be able to invoice for you. So it's right now fully integrated into our app, meaning if you use Kai and you need to kick it off to a live human interpreter, you can kick it off to a live human interpreter.

14:41September, though, will be the formal launch. The other thing we're doing flooring is it's also about what's the outcome of using Kai versus not using Kai. And so we've also started to assess savings, not from a cost standpoint, but actually from a time standpoint. So I'll give an example. For one hospital system, they would have nutritionists that would round in the hospital system and try and have a conversation with the patient in terms of nutrition, food, needs, et cetera, et cetera. And they struggled having that conversation because they were trying to do it without any sort of interpreter.

15:15It didn't make sense to bring an in-person interpreter. It didn't make sense to try and find the iPad on wheels. So they would just try and suffer through trying to have that conversation. They viewed this as a really novel solution where they could have a more efficient conversation. And what we did was we essentially asked, how long was that conversation before using any sort of AI interpretation tool, Ty? And then how long was it after you used Ty? And in that scenario alone, we found about 20 % time savings.

15:45Dipak Patel:Did you also hear that maybe in those scenarios in the past, they would have just brought their own device in a way and maybe pulled out their cell phone and try to use Google Translator? or something like that? They didn't say that, but we know, I've seen it, that physicians and providers are still doing that across the country. And I guess the way I think about this, because I have a lot of friends that are doctors, if you don't have access to interpretation for whatever reason, you're going to try and do whatever you can to understand the case, even if it's not perfect. And I think that's one of the interesting things about this industry, Florian, where we really think about it in terms of the experience of the patient.

16:25What is best for a patient? And at the end of the day, a patient wants some sort of effort made to try and communicate in their language. And what's unfortunate is we don't have the right set of tools to support that in every use case across the patient journey. You know, I think about this industry. The only innovation that has really happened, I would say, in the last 10 years was this idea, instead of always using in-person interpreters, we're now going to use an iPad on wheels that's dedicated for interpretation, right? And so what we're trying to do at Global is trying to figure out what other tools, what other innovations can we bring to this space so that we can level up the experience of the patients.

17:09We're always thinking about things. How do you make the patient experience as well as the clinician experience better?

17:13Dipak Patel:And you probably have to really make it fairly frictionless because if there's too much friction, even though the tech's awesome, maybe people are going to be like, well, you know, I don't want to do these three steps, just two. But if it's just there and it's widely deployed and it's superior to whatever other app you have in your pocket, it's going to be used. That's exactly right. You think about the world of just video interpretation right now, where you're accessing an interpret through an iPad or some sort of, you know, tablet. Lots of times the way it works is when the clinician goes up to the iPad, they have to put in a bunch of information.

17:48They have to put their username, password, put the department, put the language, put maybe some sort of medical record number or some sort of thing so that they have data on that interaction. It's a lot of friction, right? If you're a doctor, why do I have to type in all this stuff in order to just access an interpreter? And this is where, again, we've made a big push in terms of integrating into EMRs so that you're able to access those interpreters with just one button, right? You already know the language, you have all that information.

18:16Dipak Patel:Removed to friction. And you guys are also using AI for monitoring life, the quality of life calls of human interpreters, right? That was another thing I found surprising. Whenever we are bidding on an opportunity to work for a hospital system, pretty much every hospital system asks the same thing as it relates to quality. They ask, how do you, vendor, monitor your own quality? And pretty much every vendor says the same thing. Well, we hire the best interpreters. We have a rigorous process for vetting the interpreters. And then what we will do is we'll listen in on a sub sample of calls to determine whether or not that interaction went well or not.

18:56We're on a mission. And right now, we as far as I understand, we're the only company out there where you can understand the quality of every single call. And we think about it in terms of three buckets. The first one is relates to backgrounds and the way the interpreter looks on video calls. So we're using technology to make sure that there's a professional background, the interpreter is dressed appropriately, their head is in the right position. And the reason why this is important is if you're the patient, you don't know that interpreter works for Glovo. You think that interpreter works for the hospital system.

19:27And, you know, how the interpreter works and how you view the interpreter is really important for the brand of the hospital system. So that's number one. The second thing we're about to roll out this summer is understanding if there's background noise or if the call drops on every single call. And the third thing, which we're very excited about, which is going to be released by the end of this year, is accuracy. So the ability to understand how accurate was every single call. What we want to be able to do is look at our customers and say, when you work with Globo, you truly understand the quality of every single call so that you can make an informed decision on how good our services are.

20:02Dipak Patel:So when you're building this, there's a lot of components that are going into something like this. A lot of these components would be provided by the big kind of what they call now hyperscalers, the Googles, the Microsofts, maybe like OpenAI. When you're technically building something like this, how do you take this into account? Which components are you getting from something like Azure? Where do you kind of spend time custom coding? Yeah, how do you put all this together? Especially this is moving so quickly. Every day there's a new announcement of something, new mousetrap. We're not in the market to build our own LLM, and we want to be flexible.

20:37So the way we do interpretation, it's a three-step process, voice-to-text, text-to-text, text-to-speech. And for each of those three, we've built our solution in a way where it's flexible. It can use not just an LLM, and some of those you can get a great result using some traditional machine translation tools, but it's been flexible across all three. The real, I'd say, way we've improved on that is three ways. One is a lot of prompt engineering. And you know this, whenever you interact with any sort of LLM, how you ask the question gives you a different result. And so through these pilots, we've learned how to ask these questions in a way to get the best type of results.