As described in Healthcare Innovation’s Ten Transformative Trends May/June cover story package, “The phenomenon of artificial intelligence (AI), which encompasses machine learning, has been the subject of tremendously hyped speculation and commentary for several years now in healthcare. A star area of endeavor at recent annual conferences sponsored by HIMSS (the Chicago-based Healthcare Information & Management Systems Society) and RSNA (the Oak Brook-Ill.-based Radiological Society of North America), AI ended up being weighted down with enormous expectations, and inevitably, as a result, the concrete results have lagged behind those Mount Everest-high expectations.”
Still, even in the midst of the current COVID-19 pandemic, clinician and non-clinician leaders in patient care organizations nationwide are moving forward to adopt AI and machine learning technologies and apply them to patient care delivery and patient care organization operations.
Two of the industry leaders interviewed for that Trends article this spring were Ryan Pretnik, director of strategy and research at the Orem, Utah-based KLAS Research, and Tim Zoph, client executive and strategist at the Naperville, Illinois-based Impact Advisors consulting firm.
In the interview this spring, Pretnik and Zoph referenced a report entitled “Healthcare AI 2019: Actualizing the Potential of Artificial Intelligence—a KLAS-CHIME Report,” which was published by KLAS and the Ann Arbor, Mich.-based CHIME (College of Healthcare Information Management Executives) last November, and authored by Pretnik and Lois Krotz, KLAS Research’s director of research strategy. Among the key findings of that report were the explication of widely held misconceptions, including the idea that “Building the models is the most time-consuming AI task.” Indeed, the report noted, “Don’t underestimate the time and effort it will take to prepare the data needed to test and build the models. Healthcare data is hard to clean and comes from many sources, and your organization may not have the expertise to feed the right variables or features into your models. Vendors and tools can help, but you need to do your own evaluation of the time and effort required to be successful with your models.” The report recommends that patient care organization leaders study current clinical workflows and find ways to embed AI tools appropriately; that they “promote interdisciplinary collaboration,” and that they “take a social engineering approach to get staff engaged in implementing changes.”
Below are excerpts from the interview that Healthcare Innovation Editor-in-Chief Mark Hagland conducted with Pretnik and Zoph this spring, on the subject of AI and machine learning.
What does the AI landscape look like to you, from a 40,000-feet-up view, right now?
Ryan Pretnik: When we look at the realm of healthcare right now, there’s a lot of excitement around AI right now, in terms of what it could be and how it could improve care quality while also driving financial outcomes; but not a lot of organizations have dipped their toes in yet. Why is that? Number one, because the knowledge is lacking in healthcare organizations. They don’t know the use cases yet.
They also don’t know what vendors to work with; whether they should do it internally. And especially for those without a data science team or knowledge of how their process should work. And sometimes organizations don’t feel the quality of their data is high enough. We have data in all different formats: claims, HER, structured, unstructured. And providers are realizing when they start to look into AI, there are different data sets that you have to collect that are very specific for the model. It’s what you put into it that tells you what you’ll get out of it. And it involves collecting the right data. You need a specific type of data; sometimes, providers realize they don’t even have the data they need.
Those are the two issues. And number three, how do they change behavior within an organization, to actually use this? We can dig into best practices. We looked at roughly 90 to 100 different use cases within 57 unique organizations, in clinical, financial, and operational areas. The number-one area was clinical. 75 percent of the use cases we validated were clinical, around pop health, CDS, clinical research, patient engagement, and patient education. Everyone wanted to do clinical.
This is what you’ve found in the past year?
Yes. And the second-biggest focus area was operational. Employee experience and bed/patient/staffing management; and third area was financial, with the fewest use cases, such as value-based care reimbursement, revenue cycle financial health, waste/cost/fraud avoidance.
Tim, what are your thoughts, at that 40,000-feet-up level?
Tim Zoph: KLAS’s study confirms that we are very early on in this journey. And the uses cases focused on the sweet spot of clinical, for which there’s a direct impact on financial. Sepsis, readmissions, were areas. Narrow-banded and fairly pragmatic applications of this. I would say, too, that it’s going into analytics teams that are already working on this but need a better tool set and data for its effectiveness, but culturally, AI is still a departmental function and not an enterprise strategy. We have some powerful tools and some potential, but we’re starting out a bit modestly in terms of its application.
Pretnik: AI has actually been around for 50 years in some form. But what really makes this hum is the EHRs and the amount of data we’re now collecting. Years ago, we weren’t collecting this data, or collecting data at any level of scale. Now, providers and payers have all this information to sift through. So when we start bringing in some of the different vendors that have been around for a long time, if we look at those capabilities, what happened when we first started implementing AI three or four years ago? It was a disaster. There were a lot of failures. And it created a scare in the market, and most of the vendors I’m aware of are starting out from a bit of a scared standpoint. So they’re just trying to dip their toes into the water. They don’t know how many people will be needed to support this, and the amount of data involved. I see the hype; and I’m not surprised it’s taken this long. It’s still very departmental and use case-focused.
Zoph: We’ve seen this in the industry before, where the tool is outstripping the adoption. And we really have to stand back and ask why that is. I really think it’s going to require an awakening of institutions to the value of data.
Is one of the underlying challenges a broad lack of data scientists in healthcare?
Zoph: I think you’re right. To really have it be powerful, you have to scale a number of things. You have to be able to scale your data definition work, your data quality, and the repository of data itself on which you can act. It requires a great deal of care and feeding. And you have to scale your talent. Maybe there’s more of a consortium model that comes out of this, where institutions might band together, and the vendors/solution providers become more involved? We have to figure out a different model to allow us to harness the power of AI. Otherwise, some institutions will move ahead, but really, the whole industry will need to benefit. So can we get there from here? And if we approach AI from the traditional buy/sell approach, it won’t work. This isn’t a classic buy/sell situation; it requires an underlying set of resources and talent, so that to do this effectively may require a new model in the industry. The buy-in and implementation of this. I’d like to be more optimistic, but I think we’ve got some fundamental challenges here that we have to be creative about solving.
Pretnik: Yes, there’s a lot of “I don’t know what to do with this,” out there. And the data scientists will help with two areas. There are three areas where people haven’t succeeded with in AI: knowledge, tools, and change management. Data scientists can help you cleanse the data. What they can’t do is help implement change management.
Clearly, change management is a different animal from AI.
Pretnik: Yes, and AI is lacking a recipe for success. A lot of providers want to put that onto the vendors, and the vendors don’t want to do that, either. We have some data that finds that when you have data scientists or a data science team involved, the interest in working with AI rises. Same thing in an academic medical center, that has mathematicians and Ph.D. guys. But in those organizations without data scientists, they’re nervous and don’t know where to start. And putting this on a CIO or IT manager will be rough; they’ve already got so much on their plates.
What will happen in the next couple of years?
Zoph: I think that the power of artificial intelligence will be more readily understood by HEALTHCARE organizations when they see other institutions taking the lead, or begin to see the power in the embedded solutions that they buy. That will be hopeful. And there are areas of transformation in healthcare showing great promise, and I see that in at least three areas. One is in the area of decision support. And I recall this intelligently guided care. Look at all these assessment and screening tools out there now for COVID-19. This idea that you can have self-service care guidance, which underneath it, is AI-driven. Those are the kinds of tools that are going to stick. So it will be a combination of what patients can do for themselves and what else can be off-loaded from physicians. The second area is what I would call process automation, or RPA.
Especially with the labor challenges, we’ll be looking to use process automation in areas like revenue cycle automation, we’re starting to see that already. And even in help desk work, we’re seeing that. And it’s not that the labor goes away, but rather, it’s applied to higher-end tasks. And the third area that’s important is NLP, with voice-enabled documentation. We’re in the hole now around physician documentation and burnout. BTW, language processing is really good now; we just haven’t figured out how to push that into physician workflow, but I’m optimistic about where that will go two or three years from now. I think all three of those have real potential to start to make visible where data-driven workflow transformation changes can offer very clear value propositions for HEALTHCARE.
Pretnik: I’m very optimistic on where this could go. If an organization has not adopted some sort of AI for documentation, imaging, RPA, machine learning, NLP, they’re looking at it. Whenever I ask people, people are saying, no, but we will be doing so within six months to a year. So there’s a lot of optimism in organizations, and I’m seeing that the c-suite level. Tim laid it out really nicely, on focus areas. Let’s start with machine learning and NLP; those are the data science platforms that people want to use. But large, progressive organizations want to be able to use those. Smaller, less-progressive organizations are very dependent on their vendors to help them. So we’re starting to see the vendors put pre-built content into their systems or software, and are trying to figure out how best to implement those in the consulting realm, with their customers. And NLP is starting to get huge. Huge amounts of physicians, per burnout. If we can just go back and turn these recordings into something useful, that will be huge. And robotic process automation, RPA, on the finance side, absolutely. It’s heavily embedded into finance already. And it also affects the operational/supply chain focus as well.
And the last area is imaging, which is uber-important. A lot of these platforms are putting AI into imaging platforms. Those are the four techniques, and we’re starting to see providers expand.
Zoph: Imaging is one of the areas where you’re starting to already see some of the most tangible results in AI, and that’s because you already have the data sets from supervised learning, around these models. We’ve seen a lot of early promise; but you still have to get out and clinically validate it. If one institution does it, there’s likely bias in the data, so you have to get beyond this early-stage, clinical trial-level learning set, and scale it across organizations as an evidence-based protocol that can be widely adopted.
We’ve reported on the consortium established by the American College of Radiology for radiologists from different universities and academic medical centers to collaborate to create algorithms that can be used discipline-wide.
Zoph: Yes, that makes complete sense. Imaging is a good area. I think it’s going to be really difficult for an individual institution to really scale its own AI. So there’s a consortium model, a cross-industry learning model, that will emerge. I think that’s what you’re going to see; no one institution will have the data, the talent, or the environment in which to apply the transformation to help everyone. We’ve got to figure out a way to collaborate together, fast.
Is there anything you’d like to add?
Zoph: I think that we’re at a real inflection point for the industry and its ability to operate more broadly. We can no longer simply look back and manage things retrospectively. We need to figure out how to deal with issues prospectively. If ever there were a time in healthcare in which we needed intelligent and data-driven predictive models, now is the time. But for an industry that’s always relied on retrospective modeling, I don’t underestimate the cultural transformation that will need to take place. So I’m optimistic on the one hand, but it won’t be the data or technology that will hold us back, but the culture change.
Pretnik: I agree. I’m optimistic, too, but it will require a cultural shift, and it’s new. So, don’t underestimate the change management needed to implement it and to use it to drive outcomes. In our surveys, we noticed that most organizations take anywhere from one to three years to be able to successfully change outcomes using AI. So don’t underestimate what it will take to make this successful in your organization. And provider organizations aren’t used to that prospective, future-oriented outlook. It’s going to take time, and providers need to be patient and cautiously optimistic as well.