A reality check on AI in healthcare
There is a lot of talk in healthcare about using artificial intelligence to improve patient care and boost operational efficiencies. But is the technology being overhyped? Or are AI applications making a real difference? James Golden, senior managing director for PwC’s Healthcare Advisory group, is an AI expert. He discussed those topics, AI adoptions rates and more with Health Data Management at the recent HIMSS conference. An edited version of the interview follows.
HDM: People are buzzing about artificial intelligence and machine learning. But it seems the terms still need to be defined. How do you define AI and machine learning when talking about healthcare?
Golden: It’s important to have the terms defined correctly. AI is a branch of computer science that basically is about making computers do things more like human beings. So that’s a very simple explanation. Machine learning, which is a sub-branch of AI, is really about statistical learning, where we don’t exclusively derive a program. We actually expose algorithms to data and it’s self-learning.
HDM: What are the healthcare applications of AI and machine learning?
Golden: There’s a select group of problems that I’m particularly interested in. In terms of the clinical setting, it’s clinical decision support. We’ve been doing that for a while. But anything that is rules- and protocol-based really lends itself to automation. Even if that automation isn’t true artificial intelligence, we can do a lot with it.
I’m also very interested in whether we can use AI to make the physician experience better, to relieve the clerical burden. There’s been an incredible rise in the amount of data that’s available – which is a blessing and a little bit of a curse. With electronic medical records and imaging, or with the data that’s captured in the insurance process, there are petabytes and petabytes of really interesting data. And it’s this ability to actually integrate data – imaging data, clinical data, genomic data, payer data – that allows us to query the data, generate insight and derive new ideas for both patients and physicians. This is why it’s a great time for AI and why you’re seeing all the buzz.
HDM: Can you give us an example of how it could relieve the clinical burden?
Golden: A great example: There’s an oncologist I work with. She says she gets nine new protocols a week. And then there’s the data from EMRs. With patients, she now gets 250-page EMR reports. She asks: “How do I process that? I’m seeing patients and all this information is coming in.”
HDM: Could AI or machine learning techniques be used to summarize that patient data?
Physicians are having to rely on more and more data about patients. With the right tools, that data can be summarized precisely and accurately, and predictions could be made. That’s where I think AI is going to relieve the clerical burden. And, when I say clerical burden, I really mean the data burden.
HDM: And then what about on the operational side?
Golden: The value chain is quite Byzantine. The technology, however, gives us the ability to actually take complex provider workflows – regulatory compliance, for instance – and automate and streamline the process. I think in terms of operations, realizing revenue, understanding billing, understanding payer preauthorization–those are all things that lend themselves very well to machine learning approaches.
But the real use case that will play out in the next couple of years is radiology, where we’ve really made fantastic inroads. I think we’re going to give a radiologist who now does 50 readings of a scan a day the ability to do 100, or 200, a day.
HDM: Where talking about things that will happen in the future. What’s the AI adoption curve going to be like?
Golden: One of the things I’m particularly interested in is what constitutes standard of care and how long does it take for something to become standard of care at a facility, for it to disseminate across the entire United States? And I was surprised to learn it can be years – four, five, six years – especially in oncology where things are constantly evolving.
I also ask my provider clients if they see different adoption rates between older doctors and younger doctors. We do find younger physicians are much more interested in technology–just like your kids and my kids know how to use a cell phone far better than we do. But I find everyone is eager for data. And I think when you get to places that have a larger data burden, they’re more eager to adopt.
HDM: Are there any specific factors holding AI back?
Golden: The regulatory burden. One of the good things about AI, machine learning and predictive modeling is that we can do really interesting things with machines. But they deliver statistical models. And it’s hard to get the computer to explain how it arrived at a conclusion. So there are liability issues.
HDM: And implementing AI in a healthcare environment is no easy task. There are pitfalls.
Golden: Many of the initial implementations were ambitious. Moonshots are great. But you don’t have to do moonshots. When I talk to people about bringing in AI, machine learning, I tell them to start with a problem that’s really important and has high value. Use that to solve that problem. You don’t have to spend a lot of money. You don’t have to spend a lot of time. Show that it works. Gain goodwill. Get buy-in.
Can we actually use clinical data to do a better job predicting outcomes? Can we talk about readmission rates? Let’s solve a problem. Get people excited. Solve the next problem. And if you do that thoughtfully, you’ll string those together into an AI capability.
Having said that, hospitals need to have access to data scientists. They can either hire them or use outside experts. But, in either case, they need to work hand in hand with the physicians. Bring in the physicians and clinicians early. Get them to define the problem they’re trying to solve.
HDM: A lot of people still say that artificial intelligence is overhyped. But sounds like you don’t think so.
Golden: I think the answer is no. It’s not overhyped. The general idea that machines are as intelligence as humans – that’s not what we’re talking about. We’re not there. In specialized AI, things are really happening. Again, radiology is low-hanging fruit. I think this is a great time for optimism.