Lisa Farren (Host): Hello everyone and welcome to Crushing Healthcare, where we explore diverse perspectives regarding the state of healthcare today and gutsy visions for a more affordable, accessible, equitable, and sustainable healthcare model.

My name is Lisa Farren. I'm your host. And this is going to be the first in a series of podcasts focusing on artificial intelligence and specifically its use in healthcare. AI touches us all. Its tentacles are spreading wider and wider each day into all areas of life and within all industries, including healthcare.

So, to start unpacking today's topic, we have not two, but three guests today. Each bring a unique and specialized viewpoint and expertise to the discussion. It's my pleasure to introduce George Beauregard, DO; Renee Broadbent and Steve Melinosky. Our guests, all hale from Southern New England Healthcare, which is also known as SoNE Health, a physician-owned and led clinically integrated network. SoNE is a leader in value-based care, providing resources, support, and advocacy to help independent providers remain sustainable and to thrive in today's healthcare ecosystem.

So, George Beauregard DO is SoNE's Chief Population Health officer. Dr. Beauregard leads SoNE's extensive pop health programs, performance improvement, clinical integration and health equity. His clinical experience in internal medicine spans over 20 years in the Boston market. Renee Broadbent is the Chief Information Officer and Information Security Officer at SoNE. Renee is a senior level healthcare executive with extensive background in information technology, digital strategy, and data security. And Steve Melinosky serves as the Chief Compliance and Privacy Officer at SoNE. He is a seasoned professional in Compliance. Steve is certified in Healthcare Compliance and certified in Healthcare Privacy Compliance. So, we have a lot of expertise here, clearly, and this is going to be a really good conversation.

So, I say we kick it off. Renee is our IT expert here. I'm hoping you can set the stage for us as we get started. I feel as if the topic of AI is everywhere, as I said earlier. More and more frequently, AI is heard about in the news and just in our day-to-day conversations. So on a basic level, for those who may not know or may not understand AI completely, what is AI ?

Renee Broadbent: Well, thanks, Lisa. I think it's a really good idea to kind of set the foundation because we use the term very liberally. And so, I think it's really important for people to understand that there's different variations and they mean different things. So to get us started, let's talk about artificial general intelligence, otherwise referred to as AGI. It's strong AI and it represents AI capable of performing intellectual tasks across domains, very similar to human intelligence. AGI is a hypothesized type of highly autonomous AI that would match or surpass human capabilities across most or all economically viable cognitive work. So, that's kind of intimidating when you think about something that could actually think like a human.

So for some examples, self-driving cars. Imagine self-driving a car, piloted by AGI. It can not only pick you up from the airport and navigate unfamiliar roads for you, but it also adapt its conversation in real time, answer questions about local culture and geography, and personalize the experience based on the passenger's interest. Other AI systems like LaMDA and GPT-3, these systems excel at generating human quality text while accomplishing specific tasks like translating languages and creating different kinds of creating content. While they might seem like AGI, it's important to understand that they're not quite the same.

Another level is artificial superintelligence. So, we'll call that ASI, right? And this envisions a level of intelligence surpassing human capabilities in all domains, including creativity or reasoning and emotional intelligence. And here's some examples, and these are kind of going back in the way back machine, but HAL 9000 from 2001, a Space Odyssey. HAL is an AI that surpasses human intelligence and takes control leading to disastrous consequences. Or Skynet from the movie Terminator, Skynet is an AI that becomes self-aware and decides to eliminate humanity. These are sort of scary types of things that we start to think about what AI is capable of.

So, we'll talk a little bit about LLM. This is a term that people are more familiar with, and that's called large language model. And it's a type of machine learning model designed for natural processing tasks such as language generation. LLMs are trained with self-supervised learning on vast amounts of context and can be fine-tuned for specific tasks or guided by prompting and engineering. These models do require predictive power regarding your syntax, your semantics and ontologies inherent in human language. But they also inherent inaccuracy and biases present in the data that they are trained on. They are foundational models, and because they're trained on huge amounts of data to understand and generate natural language and other content. They can perform various tasks such as chatbots, content generation, translation, co-generation. Some examples of them include like OpenAI, which is GPT and GPT-4. Meta's Llama models and probably what we're most familiar with is Google.

The next one is avatars. These are kind of cool because they are digital representations powered by the artificial intelligence and machine learning. They can mimic users' appearances, their voices, their movements, and they engage in natural conversations and exhibiting emotional expressions. AI advertisers are used in like content creation, advertising, corporate communication, and that alike. But we can also use them for other things. We can use them to search through large amounts of data too, which is something we're thinking about. Some of the more popular ones include DI-D, Vid, Synthesia and Vidia. So, those are just some of the ones that are top in the field.

The third one is age Agentic AI . So, this is a super interesting type of AI that's really important in healthcare, and we'll talk about that in some of the later episodes. Agentic AI systems ingest vast amounts of data from multiple sources, third party applications to independently analyze challenges, develop strategies, execute tasks. Examples include AI-powered agents that can plan trips or make travel arrangements, act as virtual caregivers for the elderly, or optimize inventory and response to real-time demand. These systems promise to transform human machine collaboration by enhancing productivity, innovation, and insights. So from a healthcare perspective, when we talk about Agentic AI, we talk about analyzing and looking and predicting large amounts of clinical data at a much more rapid pace than a human could do.

The last is machine learning. So, machine learning is a branch of AI that focuses on enabling computers and machines to imitate the way humans learn, perform tasks autonomously, and improve their performance and accuracy through experience and exposure to more data. Machine learning algorithms can be used for various applications including image recognition, sorting and categorizing images based on their content. So, think about radiology images, salesforce casting, predicting future sales, and big data analysis. Taking large data sets to uncovered patterns and insight. And this one is super important, again, in healthcare because we do deal with vast amounts of data and being able to quickly go through there and find out information that we can use to maybe treat a patient or treat populations of patients is super important.

Host: Wow. Well, that is a broad topic. I did not know that there were that many different types. I'm guessing, like before I heard all this, most people might be thinking when they hear AI, they think of the most common AIs like ChatGPT or Claude or Microsoft Copilot. But clearly, it's bigger, it's wider, it's grander. So, break it down for us a little. Are there different types or classifications of these AIs?

Renee Broadbent: Yeah. There's like three categories based on the capabilities. And one of them is artificial narrow intelligence. So, that's ANI. it's also known as weak AI, right? And it's designed to perform specific tasks with predefined boundaries, right? Examples are virtual assistance. Everybody's familiar with Siri and Alexa. Artificial narrow intelligence refers to goal-oriented AI designed to perform a single task based on, you know, a very specific data set. Some examples that are included in this, digital voice assistance, which we talked about, Siri and Alexa, probably the most common ones that everybody's heard about. You know, everybody has Siri in their car if you're looking for directions, and Alexa in your house to do certain functions like turn the lights off, right?

There's also recommendation engines. So, Netflix and Amazon think about that. They use ANI to suggest movies and products based on user preference. When you go in and you load Netflix, if you have it, it will suggest movies. It will suggest a TV series based on your previous viewing. Search engines, Google uses ANI to process search queries and provide relevant results. Chatbots, so many organizations use ANI-powered chatbots to handle customer inquiry. You probably don't log into a system these days that doesn't offer a chatbot to answer some of your most basic questions versus first talking to a customer service representative. You go through the chatbot and that sort of helps you through what you need. Autonomous vehicles, self-driving cars use it to make driving decisions. That one always kind of scares me a little bit. Facial recognition, it's used to do facial recognition systems to identify specific individuals. And those are kind of some of the ones we more commonly think about.

Host: Well, that is so interesting. From what you've described, we've all experienced those. We all have Siri and/or Alexa, and certainly making our lives easier. Well, thanks for all that. So, this is a healthcare podcast. So, let's shift gears a bit and talk about AI in healthcare. So Dr. Beauregard, given all the types of AI, can you share some examples with us on how AI is actually used in a clinical or healthcare setting?

George Beauregard: Sure. Lisa, there are several use cases that currently exist in healthcare today. One is diagnostic assistance in specialty areas like Radiology and Pathology, where AI can analyze images to improve the diagnostic accuracy and shorten the time to diagnosis. And I'll provide a couple of examples.

There was a landmark study recently published in a world-leading medical journal, Nature, that demonstrated AI's ability to improve breast cancer screening accuracy, you know, the AI compared to radiologists, human beings. There was another study published in the journal, Radiology, that showed a significant reduction in turnaround times for critical findings, like something like a tension Pneumothorax, where the finding delivery time got reduced from 36 seconds to 12 seconds. So, those things typically result in better patient outcomes.

Another area where they're currently in use is drug discovery and development, where AI is being used to design new drug molecules that in turn accelerate drug discovery. Drug discovery typically takes a long period of time, and AI is facilitating that in a much faster manner. Personalized medicine, everyone's familiar with wearable monitors that can not only monitor, but also serve as early warning systems for various chronic diseases and conditions. Hospital operations and administration are areas where AI is currently being put to use to predict and optimize workflows to adjust resource utilization or plan for resource utilization and administrative tasks to optimize staffing ratios, throughput, and improve operational efficiency.

And last but not least, virtual health assistance and chatbots where you've got AI-powered symptom checkers and care navigators that run through algorithms or different types of logic and come up with potential diagnoses and care navigation pathways. And then certainly transcribing live conversations into clinical notes, so-called ambient listening, which takes away the task of, you know, having to type on a keyboard or actually speaking into a microphone on your computer, where the computer is now equipped to just take the conversation live and transcribe it into clinical notes. Those are some a few examples of how AI is currently being utilized in healthcare today.

Host: It's pretty amazing all the ways AI is being used in healthcare. Somewhat of a sci-fi ring to it, I would say. Things like you mentioned personalized medicine, designing drug molecules and even increasing diagnostic accuracy, which is all very exciting. But I'm going to say it's a little unnerving as well.

I think maybe we need to take a look at AI through a different lens, thinking all the innovation and increased use in uncharted territory like healthcare, where we're talking about people's lives and wellbeing, makes me think there should be some maybe guardrails in place. So Steve, now I'd like your perspective around AI. What are some of the compliance considerations or concerns on the use cases that George described?

Steve Melinosky: Yeah, thanks Lisa. I think each use case is going to have various points of vulnerability, because AI and machine learning, they start with the data you put in there. So, there's an old saying in IT, garbage in, garbage out. If you give a machine bad information or poor instructions, or even like an unclear prompt, it's not going to give you the information you wanted.

And I mean, just last week we had a pretty extreme example, the large language model Grok by xAI, it started posting troubling content on X, formerly Twitter, and it was looking at bad data, so it gave a bad output. And unfortunately, this time it was in the form of, some racist tweets actually from an AI, which was troubling.

But that leads us to consideration of the human element, which is still a very much required element in the use of AI . When we talk about guardrails, the human is the one that can interpret information. Whether that's incorrectly or inaccurately, they can create a narrative based on their own biases, regardless of the information they received. And with any technology, AI is vulnerable to viruses, ransomware, worms, hacking. In fact, you know, when we look at guardrails and having that human element in there, there's a new, troubling trend in AI, it's called an AI hallucination. And that's where artificial intelligence, given accurate data, it kind of makes up its own data, its own information, and uses its own reasoning, however wrong that is. And it says, " This is what I figured out," and it's called the hallucination. It's fascinating.

So from a compliance perspective, we have to take all of this into account because different types of AI are going to become more powerful, and they're going to be powerful enough. Like Renee said, they're going to make decisions that humans can't or won't make. I mean, HAL 9000 from a Space Odyssey was an extreme example. But if you give an AI, in this case, access to life support systems, and then give it conflicting orders, it may have to come up with its own "creative solution," right? And that movie did not work out well for the humans.

But that's the problem with using AI in healthcare is that we are dealing with humans and we're dealing with patients and their lives are at stake. So, if you ChatGPT to make a presentation on finances and it gets it wrong, you risk financial loss. But in healthcare, there's a certain element of human life at stake, and I think that's going to be the biggest risk that compliance considerations we'll have to put into the use cases as described.

Host: There's a powerful and chilling example for sure. Thanks for keeping us grounded. There is clearly a lot of considerations. I'm sure we've just barely scratched the surface, particularly in the world of healthcare when we are talking about people's lives. Since today's conversation is just part one of a broader conversation over several episodes, I think this might be a good place to wrap up today, but more to come.

So, thank you everyone. Renee, George, Steve, you've all done a great job explaining it and have graciously agreed to return and talk about AI some more. So, thank you. I look forward to picking up the conversation again and taking a deeper dive. And I want to thank everyone for joining us today, listening in. Remember, until next time, we all have a role to play in healthcare transformation. So, join us in Crushing Healthcare.