How the cloud can advance data democratization, AI development

April 10, 2024 - A new user experience for healthcare data management is on the way through cloud innovation. When properly employed to create a strong infrastructure, cloud has the potential to achieve data democratization and improve access to AI tools. In this episode of Industry Perspectives, Aashima Gupta, global director of healthcare at Google, and Kalyan Pamarthy, group project manager for generative AI and search at Google, delve into democratizing AI development, advancing medical language and literacy, and improving healthcare user experience through AI-powered search, all with an emphasis on emphasizing privacy and security.

Kyle Murphy, PhD 

Well, welcome to HIMSS 24, you guys. Alright, so everyone pretty much knows Google and are in Google's entrenched in everyone's personal lives. You can't really get through life without Google search, email, everything, YouTube. But Google in the healthcare context, I'm curious... I guess what's unique about healthcare data? And why is it you know, and how is Google approaching it, given its its expertise elsewhere?

Aashima Gupta 

So you hit it right on, you know, Google as a consumer company--so, for the last 20-plus years, we have been building consumer products. We have nine of them: search, YouTube, maps, over a billion users each. And as we were building those products, we realized that if you look into security, the data privacy, the the way to build the entire infrastructure, the data democratization--at the heart, Google is a data company. We are an information company. And when you look into healthcare, healthcare is data rich, information poor; data rich, insights poor. So we believe we have two technologies that we can bring to help healthcare data, we need more make it more accessible and useful, which is how Google started on this core idea. Our mission statement at Google, like not just in healthcare, is: make the world's information and make it accessible and useful. So in healthcare, we apply it and it is much more faced towards the industry. Can we help customers, payers, providers, start to help make sense of the data and make it accessible and useful? And we're building healthcare specific tools. We taught cloud to speak healthcare language to FHIR, HL7. We have Palettes. We talk about data silos and healthcare is fragmented. So we have tools like Healthcare Data Engine to help harmonize, normalize, and clean the longitudinal patient record on cloud. That's how we start. Then, of course, once you have the data in place, you can talk about AI. We are all talking about AI and gen AI. But if you don't have a data strategy, you don't have an AI strategy.

Kyle Murphy, PhD 

What's it like from the engineering perspective on the on the product side?

Kalyan Pamarthy 

A lot of our customers ask us, you know, we had gotten some of the largest infrastructure from like PlanetScale, right, at Google. Essentially two parts, we've got the best some of the best infrastructure and the best AI, bringing it together in service of our healthcare customers, right? So Aashima talked about a lot of like, the challenges in managing data at scale, and the privacy and security that is needed in healthcare. We help our customers do that on our systems. Bringing that all together, I think we have some of the best engineers in the world, focusing on improving technology and AI. And now we put that to use for like healthcare use cases that our customers are trying to solve for better patient care, quality, time savings, things like that.

Kyle Murphy, PhD 

I know that not everyone, you know, infrastructure is not necessarily sexy to everybody. They picture servers and air conditioning, and all that type of jazz and a lot of loneliness. But, you know, cloud is, is different. We're not actually worried about maintaining the systems; the systems are on. And then we're also able to create, what I would say, is kind of a sandbox for people to experiment. Is cloud the way forward for an industry like healthcare that generates a lot of data, but also has restrictions on access has really high bar for privacy and security?

Aashima Gupta 

Cloud as a means to an end. We talk about digital transformation all the time. And that requires data handling, data analytics. And if in that you're building and racking and stacking the server, are you really solving the problem? So if you're focusing on innovation and not infrastructure, that's endless--so cloud becomes the only way, it's the only smart way to actually build that. Give the racking, stacking, security, privacy, because we're doing it at PlanetScale. We did it for our own products, and that's the same capability. So what is exciting? I know you mentioned infrastructure is not sexy.

Kyle Murphy, PhD 

Yes.

Aashima Gupta 

But what is actually sexy is the same infrastructure that powers our product is now going to power any organization, any size, anywhere in the world. So a developer in the Silicon Valley has the same access to the same infrastructure, same tool like AI tool, like Google products. And to me, that's where democratization happens, that's where we get excited about that we are the vehicle to enable the innovation. Of course, innovation, the main knowledge will come from industry. The gruntwork of building that stack, painfully--we have done that for 20 plus years, that technology exist. Now, what can you build on top of that, that innovation? That's where the currency of innovation lies is with the ecosystem.

Kyle Murphy, PhD 

Obviously, developers love cloud environments, because it's, it's modern. It's how we do things: fail quick, iterate many, many times over. To me, it's, you know, it's been hard for developers to get into the healthcare space, it tends to be a pretty closed-off space. How does, you know, a cloud environment maybe entice some people who haven't been in the healthcare space, but have all the tools and acumen to be able to deliver solutions?

Kalyan Pamarthy  

Yeah, one of the things I think the cloud has done really well is free up time for these developers to focus on higher-order problems, right? So like Aashima mentioned, years ago, we would spend time tuning databases and scaling up infrastructure. I think now, the same folks can now think about like the hardware problems, right. So building AI, building use cases that our customers need, and your technology is becoming just faster and easier to learn. And so a lot of the lot of the folks who understand cloud, can now you know, pursue their passion, if that's healthcare, right? I mean, in building applications for saving lives, for example. And so I think cloud has enabled a lot of developers to start thinking about building in healthcare.

Kyle Murphy, PhD 

Now, we talk about large language models--healthcare, medical language is a whole nother language, it might as well be alien in certain contexts. How challenging is it to develop a model that can speak healthcare, not only from the clinical side, but I imagine you have to extend that to the patient? You know, given health literacy and things like that, to ensure that they're able to follow the directions that they're that they're given?

Aashima Gupta 

So that's a great question and if you look into large language models and learn from the [inaudible] they have learned a lot of information. But in healthcare, that information is where all the biases exist. There's all, like, valid information and valid information. So we about a year ago, there was research done by our Google research engineers on can we take the same model that learned on the internet, but fine-tune it on healthcare data, publicly available datasets, CDC, PubMed, medical journal articles, and some of our own Q&A datasets? And it was a question. We didn't know how the model would do or not do. And what we found was, with that fine-tuning, the model was actually able to pass a US medical licensing exam with a 60 percent passing rate. And now within three months, it became--actually, it's passing the same exam was given to the medical doctors at 90 percent. That means it's, it's learned. And those exams have multiple choice questions, long-form questions, it has--requires some medical reasoning. And that started as an exploration. And we're seeing the promise where a medically tuned model where you require scientific knowledge, where you require clinical knowledge graph can be helpful in a healthcare scenario. So when you talk about health literacy, if I want to now know about something, can I help reduce some of that jargon. So that's one part, MedLM is, we actually announced our last HIMSS: we are being both bold and responsible.

These are early days and we are now building on the capability. We gave it to the hands of a few customers, we're seeing what use cases. And there are two. Search and summarization: MedLM was actually very good at to be able to [summarize]. And we are seeing some of work done by our customers in applying MedLM in those use cases and see what the results. The second, when you talk about--medical has lot of terminology, HL7, FHIR, abbreviation, line code, and all that. And we announced also the Vertex Search ability to search beyond the terminology, a clinical knowledge graph that's able to understand the nuances. So there's a lot of abbreviations. We use this term a lot. There's an antibiotic versus a name of an antibiotic. So if you want to know "Okay, here's my patient, I have three years' worth of history; I want to know were they taking amoxicillin has an antibiotic." Now, if you just type the exact word with the exact spelling, if a note has a misspelling, or if there's a different generic versus band name, you would miss it. But if you have a clinical knowledge graph that has the understanding that these are the antibiotics, people can sometimes misspell it, there are different names, and then you can able to collect that full picture. And that's what Vertex Search does is a healthcare tool search. It works on FHIR.

And that's what we also announced [inaudible] to help and why that is key from the industry perspective, right? We talk about burnout. Burnout happens for two reasons: reading documentation, writing documentation. When you read it, can you apply that medically attuned MedLM, medically attuned search, so that you are now able to read the document with the context and to be able to write. Again, that's where we believe the MedLM with the scientific knowledge with the medical grounding is able to do that.

Kyle Murphy, PhD 

Then last question for you: What's it like engineering what is the new healthcare user experience for providers--coming back to the burnout part--but also we're all patients, consumers, so I'm curious what it is... It seems to be a lot of responsibility and a significant challenge. But, I mean, how exciting is it?

Kalyan Pamarthy

It's incredibly exciting, right? So just to give you an example, like when you go to google.com, and search for information, you're not thinking about whether you're entering the right spelling, or even if you're asking the right question, right? Obviously, the web has--has evolved over many, many years. And Google's invested 20-plus years of research in understanding, just, language, right? And so when you bring that into healthcare, the new user experience in healthcare could be around that, right? So, reducing the effort that our clinicians and eventually end users are putting in to find the information to find the answers find validated, authenticated sources of information. So that's, I think, the new user experience. And so many of the customers we talked to are trying to create that amazing, beautiful experience, but still with like high quality information backing, whether it's a search or summarization. And I think for us, it's, it's an exciting challenge, right? I mean, how do you take something that we've invested in for the general internet, and bring it to a cause like healthcare and just, you know, make sure patients' lives are easier, doctors can provide better care, and they can do it faster, and they can spend time actually with patients, instead of writing documentation.

Kyle Murphy, PhD 

It's wild what a little search engine has become. So thank you so much for your time guys, I appreciate it.

Aashima Gupta 

Thank you, thank you.

Kalyan Pamarthy 

Thank you, it's a pleasure.