Worries about artificial intelligence replacing doctors — at least any time soon — have largely given way to a vision of doctors and algorithms working together to give patients the best possible treatment. But what exact shape that collaboration will take is still being hammered out by developers and doctors.
At Health 2.0’s Dev4Health conference, held this week at the HIMSS Innovation Center in Cleveland, Ohio, Dr. Eric Louie, chief medical officer at HealthBox, moderated a panel of AI experts who discussed the topic.
“The more you learn about AI the more you’re amazed at how amazing the hundred trillion neurons in the human brain are,” Dr. Anthony Chang, who serves as chief intelligence and innovation officer and medical director of the heart failure program at Children’s Hospital of Orange County, said. “Doctor’s intelligence is so difficult to mimic, even in the best of hands.”
Chang estimates that 15 to 20 percent of the work doctors do is the kind could currently be performed by computers. And some of those areas should come as a relief to physicians.
“Deep learning and image recognition are perfect partners, I think ideally suited,” he said. “I’m a cardiologist, so I would love it if machine learning can take away the mundane part of my job, which is looking at a lot of normal echocardiograms, [allowing] me to focus on the challenging part instead.”
No matter what part of physicians’ workloads automated systems end up taking over, it’s important that clinicians can understand how they make their decisions. And the best way to do that is to involve clinicians in the development of AI systems.
“I think it’s always been man machine, not man versus machine,” Dr. Anil Jain, CMIO of IBM Watson Health, said. “And we absolutely need the experts to train these systems and provide some of the guard rails for what I think are going to be the AI models of the future, which are explainable AI models. If you’re going to get clinicians to make decisions based on these things, these models have to have some degree of explainability. And that starts with some grounded reality into the science of things. You can’t do this in a vacuum.”
Cardinal Analytix offers a service to hospitals that reviews hospital data to find and predict potential “cost blooms” — patients at risk of becoming high-cost chronic patients. To do that, though, the company uses a mix of algorithms and in-house clinicians, who ultimately make the recommendations that get sent back to clients.
“When I started my PhD a lot of people were saying ‘We’re going to replace the clinician with machine learning.' By the time I finished that tune had changed to ‘We’re going to augment the physician with machine learning,'” Maples said. “At Cardinal Analytics, we really take that to heart. We see the value in marrying clinical experts with machine learning experts to create actionable predictions.”
Brian Kolowitz, the director of product management and principal architect of the Advanced Medical Imaging Program at UPMC, said his team has embarked on projects based on the input of clinicians.
“We’re doing some really interesting things in natural language processing and also social determinants of health,” he said. “Just about every doctor I run into says, 'If you can just tell me who’s not going to show up tomorrow [and] get them into the office, that would be great.'”
Maples agreed that physician input is important, though he noted that one should never let clincians’ self-identified pain points restrict truly innovative ideas.
"The trick to this whole thing is you can listen to them say ‘Here are my pain points’ and you can build a solution to that, but it kind of goes back to that 'If you listened to people a hundred years ago they’d say ‘We want a horse that would run faster,"'" Maples said, referencing a quote often attributed to Henry Ford.
Chang says clinicians not only need to be conferring with data scientists, they need to be doing data science work themselves. The more people can marry clinical experience with data science expertise, he said, the better chance AI has of succeeding in healthcare.
“Most people out there in the data science world just build something and they sell it,” he said. “They have never really spent significant time with the clinicians. When they do have clinicians on their advisory board it’s just a quick phone call here and there and that’s it. I always try to embed myself there, just like I think data scientists should embed themselves in rounds in the clinics on a weekly basis and really try to learn the workflow and logistics. AI failed earlier because it never took into account workflow and logistics, and I hope we don’t make that mistake again.”