The exploration of artificial intelligence (AI) in healthcare continues to be an area of great interest for industry stakeholders, and even though it’s been overhyped at times, experts believe that AI has real potential to transform hospital, medical group, and health system operations. Recently, a group of healthcare leaders convened for a roundtable discussion where they conversed on the current state of AI and what the future looks like to them.
The roundtable participants—including senior leaders from the Peoria, Ill.-based OSF HealthCare and the system’s innovation arm, OSF Innovation; Children’s Health in Dallas; the Pueblo, Col.-based Parkview Medical Center; and Pieces, a healthcare AI solutions company—came to several conclusions, including a few key takeaways specifically related to COVID-19. For instance, they agreed that while the pandemic was a catalyst for rapid innovation in AI, there is still a need to recognize areas where AI could have done more to help combat the pandemic. The stakeholders also pointed out that AI was an important accelerant for vaccine and therapeutics developers, and that it helped democratize access to COVID-19 care, via natural language processing (NLP), automated patient screening, and digital care navigation.
Following the roundtable, a few of the participants chatted with Healthcare Innovation Managing Editor Rajiv Leventhal about the group’s most important takeaways, when healthcare organizations can expect to see an ROI from their AI investments, the most promising use cases for AI in healthcare today, and more. Those who were interviewed include Ruben Amarasingham, M.D., founder and CEO of Pieces; Sandeep S. Vijan, M.D., chief medical officer, Parkview Medical Center; and Garrett Vygantas, M.D., managing director, venture group, OSF Innovation. Below are excerpts of the discussion.
What are the core takeaways from this roundtable discussion that you believe are important for readers to know?
Vijan: I don’t think AI is a one-sized-fits-all approach to solve healthcare crises; every facility has its own needs for where they need a technology solution to be deployed. My biggest takeaway is that AI platforms have a lot of versatility built into them, so we need to figure out the nitty gritty of integrating that into workflow.
Vygantas: As it relates to funding innovation, we are keenly focused on the workflow part of it. The two biggest areas for us are fitting the technology into the existing workflow, and making it easier and more efficient. The second part is where is the value is created; how is that value captured from an economic standpoint? That’s a critical question that we often ask ourselves—who is benefiting from the investment of the technology and how does that translate into bottom-line improvements? If the workflow and the value capture question aren’t apparent, then it means the pieces of the puzzle have not yet been put together for a cohesive story to wrap our heads around.
Amarasingham: As a technologist and physician working in this space, we often discuss this maxim we have when building the software, which is: if the AI is not reducing the stress or the complexity of the workflow, then you have to consider if the AI is either not working, not optimized, or not useful. And if it is useful, if it does reduce provider [burden], and improves efficiency, then there should be an economic and clinical [value] seen.
The other element is that the workflow consideration is a big one; it’s often mentioned that the ultimate AI task is to reduce human burden across all industries. We are moving from what has been described as narrow or weak artificial intelligence, where the AI is still pretty clumsy, toward a stronger AI where it can do multiple cognitive tasks at a same level of subtlety as humans. For example, early on in the 2000s you had decision support which was considered binary—if this, then that. The new AI needs to be thinking like physicians: say there is a patient who has complex medical conditions, but who also doesn’t have any family support, and may have to go to nursing home, yet there isn’t one available. As a physician, I'm worried about whether the patient is able to get his or her medications while providing treatment. Every physician is thinking about all those things in that level of detail; so, can AI match that kind of level of subtlety and sophistication? When that happens, and as it continues to happen, I think we're going to start seeing a significant reduction in stress and workflow friction for physicians.
A recent survey of healthcare leaders from Optum found that 6 in 10 executives said they expect that their organizations will see a full return on their AI investments in under three years, nearly double the 31 percent of leaders surveyed in 2018 who expected to break even that quickly. Do you think those estimates are reasonable or too bullish?
Vijan: I think those estimates are very reasonable. I believe what healthcare executives have started to realize is that on an accounting spreadsheet, to drive revenue, either you're going to drive more volume or you're going to increase the quality of care. In order to achieve one or both of those goals for physicians who are practicing, you are going to have to give them the tools to practice more efficiently. Whether volume or quality, they just need to be able to get the information they need to make therapeutic decisions. Moreover, they need to get the information that is on the surface invisible to them. So if you [as the physician] are seeing 10 patients in a day, and that was all you did that day, you would have the luxury of doing an hour history and physical exam, and an interview [with the patient] to get into the nitty gritty of social determinants of health and their home environment. But that’s not true of medical practice today; most providers are responsible for seeing several times that number of patients each day. So AI has become an investment in efficiency to find the information that traditional EMRs just don't provide in an easy, readable format.
Vygantas: I don't see any reason why you couldn't expect earlier value capture. I understand that the pace of change in healthcare in general is slower than in other industries, but when there's top-down leadership, engagement, and a culture around efficient deployment, I can't see why you wouldn't expect to capture value earlier. We've seen examples in this current pandemic where technologies were deployed at a much more rapid pace. So, from the standpoint of whether it can be done or can't be done, that question is answered. It’s really an issue of culture and the leadership driving those efficiency changes.
What are the most productive use cases for AI in healthcare today? Where have there been recent promising breakthroughs?
Vijan: For Parkview, I'll just speak to some of our real-life experiences from engaging with Pieces. We are a nonprofit community health center that serves a suburban and rural area of Colorado. What we have not been in tune with for a long time has been the entire spectrum of continuity of care for patients. We’re an acute care hospital, so we fix your broken hip and we send you out the door. But our ability to mine data and evaluate the social determinants of health that are really the big predictors of a successful clinical outcome has been very limited. Pieces has helped us identify those vulnerabilities and then help us build action plans around them, so that they get addressed during the hospitalization, and the patient is set up for success post-hospitalization. That is a program that would have taken 20 full-time employees staring at the computer, pulling data every day for every admission to try and capture some of that information. AI has allowed us to do that in a seamless manner, presenting a dashboard that we can build human workflow around. So I think that is where AI in healthcare today has the greatest value—bringing some of the invisible things that determine healthcare outcomes to the surface so we can stare at it and figure out what those challenges are for our patients.
Vygantas: I'm going to add a different example in thinking about technology being deployed for precision medicine needs. There are companies out there like Gauss Surgical that have developed technology using computer vision to, in essence, quantify a real-time blood loss in the operating room, augmenting the clinician’s actual vision in estimating blood loss. So, again, you’re bringing a more precise, potentially visible data point, but getting even more precise about it by using technology. Another example is on the radiology side, and there are a number of companies deploying technologies. One of them is a company named VIDA that is looking at CT scans and very rapidly aligning anatomical parts from one scan to the next, making the radiologist’s time much more efficient in looking at longitudinal cases. It’s also highlighting biomarkers that are visible, but quantifying them rapidly, allowing radiologists to process more scans more efficiently and thoroughly.
How do you see the next few years playing out and how will you measure progress in your respective organization?
Vijan: Organizations such as ours set goals around readmissions and length of stay, as well as a variety of metrics on quality. And the achievement of those goals—driven by AI integration into our care delivery—is how I would measure success in the next two to three years as we continue to explore these possibilities. Identifying vulnerabilities and opportunities through data mining to match the right patient with the right piece of care that they need is how Parkview would measure success with AI.
Amarasingham: I think the progress we have made on interoperability is a substring for greater AI value creation. When you're able to have more and more different systems create different ways that can interact with each other, that's going to catalyze the value so that it's not just one technology provider that's bringing the value to providers and patients. I'm excited about the progress we've made on interoperability over the last 10 years.
The second thing is that we will just continue to see more really impressive reflections of AI in terms of the subtlety at which you can make judgments—multivariable judgments that we typically just associate with human thinking. I think will start seeing that more in medicine; a more complicated, subtle understanding of the patient, of what might happen, what is happening, and what needs to be done, hopefully all of which will increase the joy of practicing medicine for providers, and lead to better outcomes for patients. So I'm really looking for that Turing test of if you can tell this was a human or a machine in any given element of medical workflow.
