Webinar: Where Medicine Meets Machine Learning

0:41 – Chip Lynch

Yeah, I’m Chip Lynch, Senior Director of Software Engineering here at Reveal. I’ve been doing data and data-related things for quite a long time. ETL, business intelligence, data science. I’ve stayed in that data space, and largely in aerospace and healthcare and life sciences have been the two things I’ve tended towards over the course of the years. So happy to be here. Over to Alexis.

1:08 – Alexis Isabelle

Yeah, it’s really nice to meet everyone. Alexis Isabelle, I lead our AI business unit and similar to Chip, I’ve spent my entire career kind of in one place, but it’s been product management and really about enabling healthcare across the entire spectrum. I spent a number of years at an EHR company building out quality management programs as it was coming up in the marketplace and really changing from kind of a carrot to a stick approach.

And then spent a few years building out cloud pharmacy solutions for pharmaceutical companies who were launching new prescription medications and wanted to support their patients in new and different ways.

So really nice to be here.

1:43 – Ramesh Sridharan

And my name is Ramesh Sridharan. I am a director of machine learning architecture at Reveal, where I lead and advise teams working on machine learning and AI projects in the life sciences and healthcare world. I’ve been working at this intersection of AI and machine learning and healthcare life sciences for about a decade across generative AI, large language models, natural language processing, but also time series, predictive modeling, computer vision, and more.

And in the past, I did my PhD at MIT in the medical vision group, working with doctors, understanding patterns in medical images. And in my spare time, I teach an advanced data science course at UC Berkeley. Excited to talk with you.

2:25 – Dr. Salim Afshar

Well, it’s amazing. Well, as you know, I’m a clinician by training, and it would be great to lay out the landscape of all these terms I’m hearing, right? We have semantic search, RAG, LLMs. So perhaps, Chip, if you can tell us a little bit about what these things are, the difference between them.

2:46 – Chip Lynch

Yeah, I think what we’re starting with today is one of the big topics is Retrieval Augmented Generation or RAG. And we’re excited about this because we’ve done, I think, four or five projects now where RAG is the cornerstone of the technology that underlies a tool that we’ve used in a healthcare setting. And RAG is built on several other technologies. So large language models, LLMs, are the new things that a lot of people have really have re-hyped artificial intelligence as a concept.

These transformer models that underlie them mean we take a ton of data and we’ve taught AI, these large language models, to kind of talk, to actually, they’ve passed the Turing test, which was, you know, back when Alan Turing years ago was working on cryptography in World War II, he talked about, hey, when will machines be able to behave like humans? And this is the technology. In the past few years, we’ve gotten to the point where you could talk to an AI and not tell that it’s not a human.

And it’s getting increasingly good at that. And that’s these large language models and these, these transformer models underlie that. And then RAG also has a retrieval component. So, it’s retrieval augmented generation. The generation is often the LLMs, but the retrieval part is based on these semantic searches and this method of taking documents, taking text and images and audio, and embedding it in a multidimensional space that we can now search.

So, there are these vector embeddings that we use to search documents. And we retrieve them and look at them and we ask a question, and we generate, retrieve augmented generation. That’s what we’re talking about.

4:23 – Dr. Salim Afshar

So, you know, I’ve had a lot of doctors who talk to me and say, look, can we just use chat GPT to look for information? And what are the differences or the benefits of having a RAG system that you just described compared to, say, just a chat GPT equivalent?

4:40 – Ramesh Sridharan

Yeah, I’m happy to take this one. I think the issue with a chat GPT is, you know, chat GPT is trained on data from across the internet, right? It knows a lot about a lot of things, but it doesn’t necessarily know about your data, right? It doesn’t know about the data that you have in your organization, whether that’s, you know, information about a product, whether that’s, you know, patient records.

And the key benefit of retrieval augmented generation or RAG is that it’s amped up search on your documents. You can search across the data in your organization without having to share that data out to the broader world. You don’t have to put that data up on the internet. You don’t have to send that data out. You can keep it in your data, internal to your organization and secure, but still search across it, and glean insights from that corpus of data.

5:30 – Dr. Salim Afshar

What kind of data are we talking about? Is it like protocols that may be in PDFs that are used in clinical care? Pharmacy, formularies, is it everything? can you just talk about what could be in this data set that’s within the enterprise and how it could be used?

5:50 – Chip Lynch

Yeah, so it’s basically everything is on the table. This multimodal approach means we can look for audio, images, and text. So yes, you can search patient history. One of the key points that Ramesh pointed out, since we don’t have to share the information with the model itself ahead of time, is privacy and HIPAA concerns can definitely be managed through this approach. So yes, if you have healthcare records for an individual patient, a doctor can search those using this tool but then can move to another patient and not worry about having polluted across them.

You can search X-rays, you can search text records, and prescriptions, pretty much everything is available.

6:35 – Dr. Salim Afshar

So, you just brought up a point about data leaving. What about training? What models? How do you choose, How do you build this? What is the model behind it? And are you training that model? Is it your model? What’s the maintenance? Can you just talk about a little bit more of this detail? Kind of the devil in the details around a RAG implementation.

6:59 – Ramesh Sridharan

Absolutely. Yeah, I’m happy to take this one. A lot of the time when we think about this problem, you know, we like to think about choosing the right model for the job, right? It’s like choosing the right tool for the job. You don’t want to use, you know, a power tool or a Swiss Army knife if all you have to do is hammer a nail into a wall, right? Depending on the task, we want to choose the right model, meaning the right level of complexity, and the right capacity, a lot of times with RAG, one of the nice things is we don’t necessarily have to do a lot of tuning, right?

You can take a model that’s been trained, you know because training and tuning these models is expensive. It takes expensive hardware. It takes time. And one of the great things about RAG is that you can take one of these large trained models that somebody else has gone out and you know, gotten ready for you and then apply it to your data without having to do that, right? You can essentially have the retrieval augmented part take care of integrating your data with the knowledge from the model.

So we like to look at, you know, what is the use case, you know, try a number of models and a range of complexity from you know, small, smaller, cheaper, you know, faster, but maybe not as, with not as much depth of knowledge to on the other end of the spectrum, you know, larger, maybe more expensive, maybe a little slower, but may be able to answer more complex questions. And we found that, you know, looking across that spectrum, we’re able to choose, okay, this here is where we can get the best accuracy, the best kind of bang for our buck.

You know, we can address all the needs of the organization here. So, let’s go with this model. There’s no need to choose a larger model when a smaller one will do.

8:39 – Dr. Salim Afshar

So interesting. It sounds like you really have to match what your use case is to choose the right model for that. And if I did use, say, a power drill to do a small task, what is the unintended consequence there?

8:55 – Ramesh Sridharan

Yeah, that’s a great question. I think a lot of times, you could use a power tool. I mean, I’m sure you could imagine taking a power drill and just sort of using it to bang a nail into the wall, which is sort of what you get, right? The job gets done, but you’re using a lot more machinery than you need, right? You don’t really need that. And what is the cost that you pay for that? Well, there’s a financial cost, right?

Larger models tend to be more expensive. If it’s a commercial API like OpenAI or Anthropic, then, you know, every time you ask for something, you’re paying more. If you’re hosting an open-source model internally in the cloud, then, you know, you’re also paying more for the hardware to serve that. You’re probably also paying a price in speed as well. So, you know, the larger models take longer to bring answers back.

So, generally, there’s this trade-off, and, you know, we want to make sure that we’re meeting the needs of the problem that we’re solving, but we want to kind of choose the simplest model that will meet those needs and hit that problem.

9:57 – Dr. Salim Afshar

So that started to make me think about these other things I’ve been hearing about complications around AI tools in these models, some of them being hallucinations and or creep over time. Does the model size correlate with this? Does someone comment on what are some of the problems and how do we monitor for them or prevent them?

10:24 – Chip Lynch

Sure, yeah, hallucinations is one that comes up all the time, right? You absolutely, you know, one of the three pillars we, you know, hope to think about in this is the quality and safety of our artificial intelligence. And, you know, hallucinations are a high risk. So, and hallucinations, just for everybody to be on the same page, is when you ask one of these AIs a question, they’re tuned to give an answer, and they’re tuned for the answer to sound right, right?

They just spit out word after word, you know, build sentences and say the right, you know, what they think is the right thing. But they can lie and unintentionally. We call it a hallucination. We don’t call it a lie. But there’s a risk, especially in a clinical setting where you could say, hey, what should I prescribe this patient? And it says cocaine. That’s not that’s not good. Probably, you know, I am not a doctor.

But there’s a risk of that when you’re using an unbounded LLM when you’re just talking to an AI, that it will give bad information. RAG really helps rein that in. It’s not a silver bullet. It doesn’t 100% protect you from these hallucinations.

Yes, the more complex the model, I think what we see is the more subtle the hallucination can be, and the more difficult to detect even by a human. And so when you use RAG, what you’re asking it is not go a pine on the question. You’re not really asking it to answer the question. You’re asking it to search your documents for an answer. And that’s a subtle difference that really matters in the way these AIs respond.

It helps really reduce that risk.

11:59 – Dr. Salim Afshar

Alexis, I’m curious for me, say product perspective or user expensive, sorry, user perspective, say a nurse or a doctor at a hospital enterprise. If they’re listening in right now, what are some of the actual use cases where this could become helpful as an accompaniment to them in the work they’re doing, right? Because I think a lot of times we hear AI is trying to replace people. But really, how can it, you know, enable people to do their job better?

Can you talk a little bit about some use cases that you can think of or that we’ve done?

12:39 – Alexis Isabelle

Yeah, absolutely. So, you know, when Chip mentioned in terms of the data question you had earlier, like, what types of data can you actually do this with? And he was saying, essentially, the world is your oyster, right? It’s about what your particular situation is, the data you have access to. What might be proprietary and unique in your particular workflows? And so when you think about, you know, providers on the ground, oftentimes they are inundated with a lot of information at their fingertips, and that could be at a patient level, it could be at a protocol level, it could be trying to communicate and coordinate with each other.

And even thinking about where EHR started, right, and how we think about burnout really started a lot with the concept of EHRs through clicking all the time, not being able to engage with the patient. And so there are some really bold ideas in terms of where you could imagine a place where AI, RAG, and LLM are actually enabling you to engage more directly with a patient because you’re not having to do as much hands-on keyboard-type work.

But I think even in small particular use cases if you’re prepping to see a patient, or if you’re trying to identify or prioritize across all of your patients, we can see really important use cases there to say, here are the ones that we would flag to focus on first, or allow you to chat with a patient’s record. You know, you can say, I want to know what their last set of VPs was over the last three years and how does this potentially maps to what their medications were and their medication changes.

And so, I think we’re seeing a lot of use cases crop up there with just interacting with the data in a way that just wasn’t available before. And it runs that entire spectrum.

14:09 – Dr. Salim Afshar

It’s interesting you bring that up. I was just at a hospital yesterday and an individual was trying to show me something and he said, I’m doing an old-fashioned chart biopsy. And it’s literally going through each of the clicking, looking at this list of notes that are progress notes and trying to guess, not looking at every note, but skipping, trying to find what point did the decision change. And I was thinking, wow, that’s such a waste of time, because you’re really just guessing.

Is there someone made a comment about the decision not to do something? And it took a good amount of time. And it’d be just magical to be able to ask a question at a patient level, but then also at a population level, or if you’re managing, say, transplant patients. That’s fascinating. Now, what is a hospital or a health system or anyone who wants to use these tools, what’s the first step that you do from the infrastructure standpoint or from a content standpoint?

Is there some prep work that has to happen to build out a pilot and then go from pilot to say a product that’s enterprise-wide?

15:24 – Chip Lynch

Well, first you watch this webinar and you call us.

Yeah, I think, you know, it’s interesting. One of the things I think is very important is understanding the data you have available. When we come in, you know, one of our first questions is always going to be, one of our lessons learned every time we do one of these is we need to have good access to the data that you’re going to end up using. So if your problem is in the clinical space, if your problem is in the research space, where you might want to apply these, our first question is going to be, what data are we using?

And a lot of the historical work that underlies data systems has been going on for 50 years or so. Data has been huge. But people are getting various levels of data. Expertise with their own data. And, you know, so that’s going to be my first question is both what’s your problem space, like understand your need, but then what data do we have and how can we organize it and help get a grip on it?

16:28 – Alexis Isabelle

I think oftentimes it actually follows kind of a typical product, normal product development cycle, software development cycle, that’s not necessarily unique to AI. I think it’s that we see components of AI filter in at certain stages of the process, but early on, you’re really just trying to identify those problem areas and where you think it could be solved. And what’s different now than what we had 10 years ago is we can solve so many more things than we used to.

The answer used to be to the point about data was, okay, we have to structure everything. We have to actually get it to a point where we can understand it. We can put it into a table and we can pull that information out. And what’s really great is now we don’t have to say we have to get everything structured. It’s more about we have to have access and we have to understand it to a certain extent and make sure it’s clean and we know how to handle any sort of data gaps or anything like that.

But so much more of that data has become accessible. And so the ideas that can be generated really early on in that stage from either shadowing a workflow of what someone’s doing or hearing about pain points as they’re going through, all of those things are happening probably with some level of frequency within organizations already. They’re aware, right, of where they have challenges. And now it’s about just having more opportunities to solve those in unique and different ways, but you still apply kind of a rigorous assessment of what is our risk profile.

What access do we have to data? How much are we willing to invest in this? Do we have the right team to be able to do this type of work? Do we need to have a quick win versus having something that’s higher grade, more production-ready, and kind of faster. So, thinking about all those trade-offs early on is really important.

18:01 – Dr. Salim Afshar

You know, this made me think of like this concept of moderation in all things. I remember, so I’ve worked in multiple large systems, and there was a point where suddenly there was Dropbox, Box, and SharePoint, and across the institution, people just randomly developing their own because the institution itself did not have a policy and or a tool. So, all my research labs had their own tool. We had our own private little Dropbox business.

And this is making me think about how do we, what are the best practices perhaps for one extreme being apps like no, no one could do anything. You have to use this one tool and move forward to a free-for-all-around the building, using these tools in the context of research clinical care, and operational work, any thoughts or comments about that? And then also, I guess on one extreme, any tools possible versus we’re going to stick with say this cloud service provider, like say AWS.

19:06 – Dr. Salim Afshar

What are your thoughts or recommendations regarding how a policy should be set or how an institution should approach this challenge?

19:17 – Chip Lynch

I think, like you said, most of the time when there’s a new technology, someone, especially in large organizations, is probably going to jump on it and start using it, probably poorly, a little at first. And when I say poorly, you know, it’s great to have early adopters. You want to learn, you want to gather that insight. But especially, you know, we talk about like roadblocks to adoption and health systems.

Like one of the problems with technology is you’ve got a lot of these requirements you’ve got HIPAA. You’ve got security and you know, I’d say it’s a problem. It’s a problem for adoption of technology it’s not a problem for the You know, the enterprise, you want these restrictions, and they’re there for a reason. When, you know, these new AI chatbots that are just free on the internet pop up, and people, you know, we do a lot of software development, so it’s very frequent, we’ll go to one of those and say, hey, how would you solve this?

And it’ll spit out code for you. It’s a nice way to learn, it’s a nice way to try a new tool. But you run a risk of exposing your own data. If you say, hey, how would I diagnose this? Or, hey, here’s a patient’s record, and just paste it into a chatbot, you’re violating a HIPAA rule, right? So yeah, you need to meet that neutral ground of properly setting boundaries and training, you know, everyone to use this effectively and most importantly, legally and ethically, and then, you know, then open it up.

But yeah, I think, you know, we can help talk through all that.

20:37 – Dr. Salim Afshar

So I’m going to put a general counsel hat on for a hospital. So, what are the tools or guardrails that you could provide me to feel confident that this ecosystem is evolving, but not violating or putting us at risk from privacy? We’ve had some massive cybersecurity breaches in the past. Two months. And I think this is really one of the biggest priorities. So, can a RAG implementation be done in a safe, secure way?

And how do I ensure that implementation? How do I monitor that?

21:19 – Ramesh Sridharan

Yeah, I’m happy to take it, maybe we can kind of split the, I feel like that’s a two-part question. I’m happy to take the kind of guardrail side on the, you know, when we build one of these models, you know, everyone’s always, everyone knows that there’s this risk of these models hallucinating, right? It’s sort of one of their hallmarks. And so, when we go out to, you know, build a solution using these technologies, we spend a lot of time on evaluation, right?

Evaluation is, you know, as important or maybe more important than actually building out the capability because we want to make sure that the answers coming back are reliable, that they’re kind of grounded in the documents that we’re looking at, you know, and that they’re not just, you know, random knowledge or worth something made up. So, we spend a lot of time on evaluation, both sort of automated using another layer of language models to look through and see, okay, are we catching the right thing?

And also human-driven evaluation, right? We want to have subject matter experts look at the results, make sure, yes, that looks good. No, that doesn’t look good. Okay, then we can go in, and catch the result. We can have automated kind of filters and things along the, the software engineering path to make sure we catch any phrases or ideas or, you know, anything that might not be directly what we’re looking for with the answer.

So, there’s a lot of ways that we can address this, you know, find any potential issues that are coming up, go back and address them using a number of different tools. And then I think the second piece of that is the data privacy piece, right, which maybe I’ll hand off to Alexis and Chip.

22:57 – Chip Lynch

Alexis, do you want to chime in on that?

22:59 – Alexis Isabelle

Well, I guess, yeah, if you want, definitely you can jump in on data privacy. I was also just going to mention that Ramesh said, you know, a human in the loop. And I think that’s really critical, is one of the conversations that you should be having is not if there’s a human in the loop, but when is there a human in the loop? And what do we need to show them so that they can actually do an evaluation that legal privacy regulatory teams feel comfortable with?

And so, for us, even thinking about working with a client, you know, they all have these different functions within their organization and you want to engage them early and you want to engage them often. So, they understand as you’re developing these products, kind of what’s going on and you’re getting a sense for where their risk tolerance is so that it can actually then feed into the decisions made in the evaluation, the process that Ramesh was talking about.

If we don’t know what the threshold is we’re trying to achieve, it’s going to be really hard to actually say at the end of the day, we have a product that someone’s going to feel comfortable launching. But I think the expectation should be that with most things, especially in a healthcare setting, there is going to be a human in the loop, and you just have to figure out what that looks like, And actually, this is very much of what you were talking about Sal, but it really does bring up this whole other side we haven’t talked about really at all, which is, it is still a functional part of a product, and so a workflow, a design, an experience is really critical for the user because you can’t just say, I’m going to go do a prediction model, or I’m going to go create a chatbot, and it’s kind of, throw it out there, right?

Because it’s maybe relatively easy to build these days. You still need to incorporate it into a product, into a daily process so that you’re building that trust really early on. And it’s trust with the product and trust with the AI. And that’s not just for the user, but also for those internal teams that are really critical.

24:37 – Dr. Salim Afshar

I love that you bring that up because I think trust is fundamentally one of the key pillars in doing this. And trust from everyone involved, right? There’s a lot of fear and worry in many dimensions. And a lot of it can be, I think, addressed if we had more discourse and education and engagement at all levels of the organization. A question I had regarding the workflow, how do you envision a workflow that, say, impacts a primary care physician who’s managing a panel of 3,000 patients Oftentimes, things get lost in translation as I go to the data engineering, and data science group, and it comes back.

How do we close that gap where the needs of, say, the clinicians, the nurses, as well as the business needs of, say, the management teams are being prioritized and we’re fine-tuning the human-to-human communication around what they want to build? Do you have any thoughts regarding things you’ve seen that have been effective from a communication standpoint?

26:06 – Alexis Isabelle

Well, I wish I could easily solve the provider with 3,000 patient panel and how to use AI to make that so much easier, but breaking it down into some smaller components, I guess. One thing that we’ve done, I think, has been really successful when we think about building these products is making sure that we’re designing a team with the right individuals who will have different perspectives. You want to have really strong engineering talent that is able to dig into that model and understand the complexity.

And to be perfectly frank, I’m never going to understand that complexity. You know, oftentimes I’ll get into a room with Chip and Ramesh and they’re talking like way over my head. But what I can do from a product perspective, or just from, you know, a human perspective that is not that smart, I can sit there and say, wait, I’m confused about this, or I don’t understand how we arrived at that outcome, or wait, what am I supposed to do with this information next?

And it’s those probing questions, I think a little bit that what you’re getting at, which is really, we can do so many super cool things, but how are we actually helping someone at the end of the day? And what are we truly solving for? And I think that does require a collaborative approach of understanding what they’re doing today, understanding how it could be improved. And that’s really where you have that kind of engineering component of what’s possible and what do we need to make that actually happen.

And then you have the design piece. So even when we build teams, we think about kind of a trio of an engineering tech lead, a designer, and a product manager, and those three are together collectively responsible for making sure that provider has a good experience and making sure they understand it well enough that we’re solving the right problem the first time. And they were able to then expand that in the future into other areas that continue to make improvements in their life.

27:43 – Dr. Salim Afshar

Oh, that’s a great point. I just had someone reach out to me recently, and they are from the cryptocurrency world and Bitcoin world. And I asked about that, like, oh, RAG implementation, like, oh, we could build it for you in two weeks. And I stood there thinking, like, hmm, but you have no understanding of healthcare. Like, you may technically be able to do it, but I think there’s a, like, the technical part could be pretty fast, I assume.

I don’t know. Again, when we’re mentioning Chip talk, it’s over my head as well, but it’s the nuances of that, how do you engage, how do you capture other requirements? So, I think, and that leads to trust as well because we don’t have enough time to kind of iterate in ways that are so deviant. From what we really want. We’ve had, I think, a history of, from a clinical standpoint, really bad experience with technology over the past decade, where it really didn’t unlock the ease, and it’s been a lot of work.

28:46 – Dr. Salim Afshar

So, I’m very sensitive to that concept, and people quoting real quick two weeks, we can build it out.

It takes time to understand and connect with people.

28:56 – Alexis Isabelle

And you mentioned infrastructure too. I think even now, you know, it’s really accessible to do a proof of concept super quick, like on your local computer, right? I’m sure that if we threw a question at Ramesh & Chip, they could go away for, you know, a few days, a week, two weeks tops, right? And come back with, here’s what I was able to produce. Isn’t this cool? But you still actually need to do a ton of infrastructure to make sure that is a production-ready product that you might’ve gotten past that initial hurdle of, can we establish that it works?

And can we establish that there is a potential to build lots of trust in it, but you still need trust in the product as a whole? And that really means putting production infrastructure around it. And that’s almost never two weeks, right? So, I think you get that first like proof that it can be done in a few weeks. And that’s amazing that we can do that. But it’s the work that is around that, that is sometimes the part that I think gets forgotten when we think about how magical AI feels and can seem when we think about what it can do to help solve some of our problems these days.

29:59 – Dr. Salim Afshar

Yeah, this is going to be another point I’m going to bring up around the kind of philosophical debates I hear regarding open source versus private, but even more around, again, I don’t know the nuances between Google Cloud and say AWS and the different cloud service providers, but from an enterprise standpoint, my understanding is most hospitals, at least a lot of life sciences and hospitals are using AWS.

What are the challenges if someone in that institution wants to use another cloud service provider? Is it easily doable or should people tend to try to use the same platform and the tools on one platform?

Is there any thought regarding the pros and cons, the difficulty with maintenance.

31:23 – Chip Lynch

Yeah, I mean, there are a million choices that go into choosing a cloud provider or choosing to do something on-premises. You talked about open source. I’ve been an open-source fan for decades, using very early versions of Linux and what not on my desktop. I find the power of that free open-source world amazing. But what you pay for with cloud providers is often you’re offloading a ton of internal And a lot of that is very hidden from the end user.

There are things like, you know, we talked about how to avoid your data getting out. You know, if I’m chatting with a bot, what’s the technical. So if you’re in an AWS ecosystem, it’s not just that you have tested your AI and you can use AWS as a guardrails tool now, which is actually designed to help make sure your LLMs don’t hallucinate or don’t give bad information or don’t give out private identifying information or healthcare information, so don’t leak your medical records.

But you can also build infrastructure that no one sees, like a VPC, a virtual private cloud, where only the people inside your organization can connect to it, and the networking will not allow it out. It doesn’t connect to the internet in that way. You use tunnels, you use a lot of technical things that take time to set up, time to test, time to, you do penetration testing, you know, you make sure you hire security experts and say, hey, every 12 months, try and hack my system and see if you can get it to lie to you or get it to you, you know, offer you a prescription that you shouldn’t or even just get information that you shouldn’t have access to.

All those things are time-consuming, but they’re part of any large organization’s, you know, responsibility to manage their data. Now, in terms of which Cloud provider, I think they all have their pros and cons. These are very large organizations and you’re getting smaller niche players in too. Open AI, is very cool, very hot right now with its AI offering and you’re seeing them start to build up more of these abilities around their core GPT offering so that you can opt out of, you know, having them do your, you know, take your data and train it.

And you can maybe move into a private space. You see a lot of people offering their tools on AWS or on the Google Cloud so that you can, again, put it inside your network wall, even if it’s in the cloud, have a little more control over it. So, there are a lot of things we can do to protect it.

33:23 – Dr. Salim Afshar

No, it’s interesting that you brought up the security about hiring people to test your models to see if they’re violating your rules and guardrails. And made me think that although a solution may have, say, its own, quote, guardrails, an institution may want to create its own just to double-check and verify the performance. Because one of the things I wonder is how often I should be checking it.

Any thoughts? Is it a daily check of the model for it shifting? Is it a monthly? Is it a yearly? What’s the speed in regards to the change that’s occurring in these models?

34:09 – Ramesh Sridharan

Yeah, I think there’s kind of two… I think that’s a really interesting question, because, you know, to me, on the technical side, I almost hear two separate questions, which is, there’s one of, you know, how often do we go out and look for, okay, how do we pull out the latest and greatest? You know, should we be upgrading from, you know, version 3.1 to version, like, 4.2? But there’s also, how often do we want to check and make sure that what we’re doing is delivering the right thing?

And one of the things that I learned fairly early on in my career doing machine learning and AI, and this is long before the current crop of gen AI models, is monitoring and checking the outputs of every model, every prediction, is important. And so that level of monitoring should happen on a daily basis, right? Sometimes, whether it’s like dashboards that are tracking, okay, here’s how many responses came in, here’s how many, you know, the automated evaluation, you know, determined we’re good, here’s how many should be reviewed by somebody or double-checked.

That can happen almost at a daily or even weekly level. But then when it comes to going out and evaluating and, you know, choosing a new model, I think that really depends on the context, right? We like to think of, I think it’s important to bring in the latest and greatest technology, but it’s also important to make sure that you’re focusing on delivering value, right? I mean, I think, you know, you mentioned this level of mistrust in technology.

And a lot of that comes from the focus on, like, technology as the be-all, end-all. But really, we’re trying to solve this very human problem, right? We’re trying to solve the problem that care providers, and scientists have. And we need to make sure that, you know, when we go out and choose a new model, it’s because, well, they want to solve their problem better, or, you know, they have issues with what’s going on, or this new model can improve their lives rather than just, oh, like bringing in the latest technology for technology’s sake.

36:09 – Dr. Salim Afshar

So, that’s really interesting. So, I’m wondering, In the work that all of you have done in this space, can you give some examples of use cases that have worked well what were some of the roadblocks that some companies or institutions may have faced, and how they were addressed so that people can learn? I’m sure there’s many people who are interested and have certain roadblocks in their own enterprise and there may be some learnings that they could be from actual implementation.

And then also the use case that was used.

36:45 – Alexis Isabelle

So, I can tee up one use case and Chip and Ramesh can go into more detail about it. But, you know, we’ve been giving a lot of love to the healthcare space. And I think actually one of our most interesting use cases in the last year has actually been in life sciences with a pharmaceutical client. And they had, not unsurprisingly to their industry, they had this wealth of unstructured information that came from a number of different sources.

They just weren’t able to tap into it. Um, and they also are in, you know, obviously a highly regulated environment. So there were certain data sources that maybe we can’t use today, but we can use in future things where maybe there’s PHI or PII that you don’t want to include into a product. But really, they started out with this ambitious goal of saying, we know we have this information. We just can’t serve it up to any of our users in a functional way.

How do we actually pull it into a product where we can then have a user interact and glean some insights from that? And I think we also hear insights a lot when we think about AI, right? So you’ve got to dig down into what does that mean to you? You know, what is an insight for your perspective? Because you want to be able to say, here’s some information provided, and I’m able to take a next better action as a result of it than I was able to before I had this piece of information.

And so with this particular client, we thought through a number of different applications, and they really came in with some really great ideas that they wanted to implement quickly. And it actually, thinking about a previous conversation we had, it ended up being a fair amount of infrastructure to start that. And where we are now is really refining that process and figuring out what do we continue to enable from that data?

So, whether it’s an interactive chat experience that feels like two years ago, it was really out of scope for them. There’s no way they could imagine interacting with their Twitter data and just asking a question in the product. Great, we can totally do that now. If it’s a dashboard that is a more dynamic dashboard, adaptable dashboard with information, that’s something that can be done. You mentioned early on semantic search.

So, there are so many different things that you can incorporate into this, but it starts with this one really ambitious idea of, we know we have this, we want to do something, help us figure out how to approach that. And I can imagine you could look at every pharmaceutical organization out there, they also have similar unstructured data that they just aren’t using to the best of its ability. And that’s something they could be getting more value out of and thinking about how to actually apply AI to those situations to do that.

39:08 – Chip Lynch

I’d love to jump in a little more. So Alexis mentioned Twitter, and I know we had some questions about this from the perspective of a large pharmaceutical organization, but you can look at this massive amount of text that Twitter sends you, and you’ve got to go through and say, hey, look, I think a natural question for a drug company is like, look, are people talking about our drug? So, your drug has a name, right?

And, you know, so you can search Twitter and see what people are talking about it. But you might also search for, you know, get all the Twitter data from, you know, maybe you have a conference that you go to which produces or like, you know, the pharmacies or the places you sell it. And drugs have two or three names, you know, which is which is an interesting technical challenge. It’s normally an FDA approval under some code name.

It might have a, you know, chemical name that is well known to people, but then also a marketing name. So you’ve got to look at all three of those and combine them. That’s a traditional data problem right there. But then you look at something like X, Twitter, that has all this weird unstructured information. One of the problems we got trying to go through that in a real case was spam. You know, random people just using a drug name as a hashtag and a message that has nothing to do with it.

They’re trying to sell cryptocurrency or they’re trying to, you know, whatever, you know, in various languages too, right? You’ve got multilingual data that may or may not be reasonable and something you want to include. So, you’ve got to now have artificial intelligence that can go through it. And then we talk about like maybe different agents, you know, Alexis was talking about a dashboard. So, we want to see, hey, just in general, over time, what are people saying?

Do a sentiment analysis. Are they saying good things or bad things? Maybe you want to look at adverse effects, which are huge, of course, in pharmaceuticals. Are people going to Twitter to say, hey, this drug made me sick for 24 hours? Like, you really need to know that. That is vital to pharmaceuticals, to be able to pull that information out. And you can’t trust that people will tell their doctors that, or people will call the number on the prescription and tell the company that.

So, these third-party just blobs of data, Twitter, Instagram, everything, you know, important avenues, and very difficult to manage this well, because you have to make sure the signal-to-noise ratio is good, right? You want to get useful information out of it, but the amount of junk is voluminous, and you’ve got to be able to parse through that. And the AIs don’t do that by themselves, right? So, there’s some traditional data mining, and you can maybe say, hey, build a little custom AI, an agent that says, hey, is this even relevant to this question?

Is this a real treat? Does this feel like it came from a real person talking about a real thing? Is it somebody quoting somebody? Is it, my friend said this, is it a direct or an indirect? You know, what’s the sentiment? What topic is it about? Is it adverse effects? Is it taste? Is it something that matters or doesn’t? Is it cost? Are people complaining about the cost? All of these little things are vital.

42:09 – Dr. Salim Afshar

You know, so we’ve now, so Alexis, you talked about, how you come in and you discover other issues, right? The infrastructure was an issue in our new use cases. So people who haven’t done anything yet, but are very interested. Is there a best, easy-use case that allows for the institution or enterprise to start building momentum and learning and to discover the other elements that need to occur to really unlock the power?

I would love to hear ideas, both from the healthcare and life science side, where it’s something that’s tangible, that could do. How long it will take in general, but it’s something that is going to build both institutional and enterprise learning, helping them find their policies and so forth, and a demonstration project where they understand, they start to see what they can do.

43:09 – Alexis Isabelle

Yeah, I’m really glad you mentioned it as a demonstration project. If you’re trying to start and you don’t have anything yet, it is a pathway that is not yet paved. And so, you actually don’t want to go after the most ambitious idea, I think, unless you’re going to put an entire organization like OpenAI behind that, right? They went after a really ambitious idea, but they spent a lot of time doing that, a massive engineering team.

So, unless you’re willing to invest in that, you actually want to start out with something that you can slowly start to chip away at that path and make sure you’re making those connections with those privacy teams, legal teams, those groups that we were talking about and understand use cases. And one that we’ve seen that I think has been Particularly useful as a 1st step, just because it’s relatively quick to implement.

You could do something say, in. You know, in 4 weeks, you can have your 1st version of it’s pretty robust and you kind of refine it over the next 2 to 4 weeks, and so you’re, looking at a total of two months, but is looking at documents. Where do you have a massive set of documents and you just want to be able to interact and ask questions about those documents? That’s one of the easiest things to think about.

It allows you to interact with something you haven’t been able to do in the past. It opens up this new use case. The documents themselves already exist, right? You’re not trying to create something. It’s just a matter of getting them into a state where you can start to query them and interact with them. And the risk there is pretty low. You’re not making a recommendation yet. You’re not making a prediction.

You’re just providing access and interaction with it in a different way than you’ve had before. But it lets you start to establish those conversations with the people you would need to have when you want to go after something more ambitious. Predicting a disease state or predicting the disease progression, is really critical. Or if you’re thinking about the pharmaceutical drug space, how do I actually predict which one of my clinical trials is going to be more successful than another?

Or what’s the next best transcription pathway to go around? So There are all these unique specific use cases you can have, but we have seen it be really valuable to start with something that feels like low-hanging fruit from a number of different perspectives, both risk, engineering investment, and kind of user experience.

45:13 – Dr. Salim Afshar

Well, I love that use case because in the era today of the high turnover that we have in nursing and staffing, the effort for training is high. And to be able to give them a tool that after their training, maybe even shorten the amount of training, but they’re able to retrieve easily and not have to figure out how to, you know, I call it dumpster diving in the SharePoint to try to find the right PDF.

That’s a huge impact, but yet it sounds like not a big lift from a technical implementation standpoint, or at least a good demonstration project around that. And I think there are a ton of areas that are beyond the typical kind of thought process, where you think of revenue cycle management encoding, that’s one, say, bucket, But then there are other things that are very valuable, predicting when someone will be discharged from a hospital or who’s at risk for an adverse reaction in the clinical trial, right?

So, I think that it’s really exciting that these tools are emerging and that we’re able to implement them.

So, one thing I want to say is clearly, what I’ve taken away for the last, say, 10 that context matters, right? Having a discussion, a deeper dive within a person’s enterprise’s context allows for people to move forward. And I think we welcome anyone who wants to reach out to us to go deeper in exploring how a RAG implementation, what a demonstration project could look like, and do a discovery together.

Any other thoughts that Ramesh or Chip or Alexis, you want to share?

or things I’m missing that you think would be important for, say, executives and management who are saying, gee, I have this team. How can I work to see where we can get an implementation that ties to an ROI?

47:18 – Alexis Isabelle

Yeah, just 1, um, folks might want to add in more, but 1 comment I would say it’s also your last point, right? About starting that conversation, you know, reaching out is even that can feel daunting. Sometimes as well. I’ve got so much work going on, right? My team is already kind of overworked and understaffed. We don’t need to have to explain what we’re doing to someone else. And I think there’s. The potential for this misconception is that even starting down the path of trying to do an AI transformation or integrate it into your experience is going to require a significant investment of time ahead just to get someone used to what you’re doing.

And we talked about the value of having context, even from the engineering perspective or the team that you’re working with. And so, I just wanted to demystify that a little bit. It can feel really daunting to say, I’m going to invite someone into what we’re doing and try and have them help me figure it out. But I think actually the lift there can be. Less intense than it might feel because really what we’re looking for are specific pain points.

And we’re going to be working with ideas that are probably already percolating, and it’s just about refining those ideas and helping put some structure around it, and I think that’s something where you would go talk to kind of normal, you know, software engineering teams. They do that all the time, right? Thinking about ideas and then bringing them down into kind of breaking down into the components that you can actually go and build and refine.

And that’s not something that I think is a normal, natural part of the healthcare space because you’re so focused on the patient care and so focused on the day-to-day actions that you’re not going to break out of your day-to-day to say, I’m going to go identify all the places today where I took 10 extra clicks or where I couldn’t find this particular document that I was looking for. But even having just someone shadow that experience or someone hearing about it a few times, you can pick up on those insights pretty quickly and then really layer in this approach to understanding what’s the right one that we could actually tackle first.

Where is that low-hanging fruit? Where do we have data? Those types of pieces. So, it’s really about an early introduction and then you’re relying on kind of the expertise of the individuals around you to build that. So just want to make sure it didn’t feel super daunting when you think about inviting someone into that experience with you.

49:28 – Dr. Salim Afshar

I like that a lot. What I like about that is today there are a lot of institutions that have, say, implemented on your journey with a technology platform, a new EHR implementation, a new tool. The reality is EHR platform does a lot of what you need to do. There’s like that 25%, 30% stuff that’s outside.

It sounds like this could complement that process, right? It’s not, that we’re not replacing the process, but those friction points that people feel around accessing knowledge content, right, within a system, and so forth. And, you know, seeing from that perspective could be less daunting as well. It was great. Ramesh and Chip, any final thoughts as we wrap up, get towards the end of the hour?

50:22 – Chip Lynch

Yeah, I think one of the things I’ll say is, you know, we’ve been very focused on kind of a very easy-to-understand patient, you know, clinical interaction here. But Alexis did drop in a quick little note. I want to expand upon it. She mentioned transcription as kind of a chemical process we can work on. We’ve taken some of these AIs, we worked on a neural node project, ordinary differential equations inside a neural network to help predict whether like working on PKPD interactions right are these drugs going to interact well are they going to cause adverse effects if you take two drugs at the same time and this is deep in the research right um I just wanted to point out as that example that you know, there’s AI applications happening at every level of the organization.

It’s not just to help with clinical care, although we have a ton of use cases there. It’s not just to help pharmaceutical organizations that have already released a drug or are looking into the data that’s coming back from that. Early in the chemicals or in FDA approvals, you know, there are lots of opportunities for wherever paperwork is difficult, wherever a product is difficult, but you can narrow it down and quantify it into something that you could just ask an AI to try.

They just released AlphaFold3, I think, recently, which is a protein folding major piece of artificial intelligence that’s doing something which has been difficult in pharmacology for decades. And it’s, you know, doing it well. So, yeah, I think, yeah, I just wanted to add that, yeah, at every level of an organization, we have this opportunity.

52:01 – Dr. Salim Afshar

That’s great. Thank you. Because we all get biased to the area that we are focused on, but the magnitude of the scope of how these tools can be applied is mind-blowing, right? To think at a molecular level, full level, to like operational level. Thank you for pointing that out. And that’s also exciting, right? To think about all those opportunities and things that I wouldn’t be able to imagine because it’s not my, I don’t have that lived experience on those challenges.

But that the playbook, the skillset, is a key component, plus understanding that context. So that’s wonderful. And Ramesh, anything you wanted to take away or something to share at the end?

52:51 – Ramesh Sridharan

Yeah, I think just, you know, to maybe build on what Alexis and Chip were saying, I think there’s a lot of really exciting opportunities for AI in the, you know, healthcare, medicine, life sciences world. And all of those, I think one of the most exciting things about this space is that there’s such an interesting interaction between, you know, what are the problems to be solved and what are the technologies that we have available, right?

And I think that as the technology has kind of improved over the last five years, or really over the last 10, 15 years, there’s so many more problems that we can solve. But we can only solve those problems if we work closely with the people who are actually going to be benefited by solving them. There’s, you know, all these interesting use cases like Chip mentioned around, you know, this sort of like PKPD modeling of like drug dynamics and, you know, recovery over time and, you know, concentrations of the drugs.

We’ve worked on problems involving, you know, like prediction of activity or health levels or, you know, sickness from wearable sensors. We’ve worked on problems, you know, looking at large-scale imaging, you know, like histopathology imaging, you know, both like at a small scale and also looking at like whole slide, like large-scale data. And there’s so much potential across, you know, I think now everybody’s excited about Gen AI and large language models, and understandably so because we’ve seen such large leaps in that tech.

But I think there are also so many other applications. I think one of the things that I think is really exciting is getting to work on those problems, but also being able to really understand how this is affecting the people who are actually doing the work and how that technology can improve their lives.

54:40 – Dr. Salim Afshar

Wow. Well, I’ll just end it there because I’m excited and I’m ready to accompany and connect with everyone. And I welcome people to reach out to us We would be happy to begin this journey and looking at how we can leverage these tools within the context. So, thank you, Chip and Alexis and Ramesh for joining today. I want to thank everyone who’s joined us live on this webinar for being part of this community and the questions that have been posted.

Thanks, All.

Bye, everyone.