When Ritu Agarwal and Gordon Gao launched the first Conference on Health IT and Analytics in 2010, they could not have predicted the explosive growth of artificial intelligence and the many ways it is changing healthcare.
“In some sense, it was prescient as to how important technology was going to become to the delivery of healthcare,” Agarwal said.
Over the years, the annual conference hosted by Carey’s Center for Digital Health and Artificial Intelligence has gotten larger, but its central mission has not changed.
“Our goal was always to bring together researchers in economics, computer science, public health, and medicine,” Agarwal said. “In addition to the multidisciplinary perspective, it's also important to have dialogue between the research, the practice, and the policy communities. You want translational work that can eventually be incorporated into what companies do, and into what policy does.”
This year, the conference, held at the Johns Hopkins University Bloomberg Center in Washington, D.C., attracted hundreds of participants and audience members over two days. Some hoped to use AI to solve specific problems, while others presented research examining the ramifications of innovations that were already in place. Research, practice, and policy communities were well-represented.
Intended and unintended consequences
In several instances, research showed the unintended negative consequences of AI interventions.
For example, Lydia Manikonda, assistant professor at Rensselaer Polytechnic Institute, won “best poster” for her work showing how chatbots that people turn to for help coping with depression can actually make their condition worse.
In another example, Jingshu Luo, assistant professor of finance at the University of Mississippi, drew interest with her research showing that hospitals that rely more heavily on patient portals have a higher risk of malpractice suits.
“Based on personal experience, as a patient, I enjoy using my patient portal,” she said. “I assumed doctors would like it as well.” But some told her it increased their workload, as patients sought explanations for the information in their files.
She also thought maybe the portals could decrease malpractice risk by improving documentation, patient engagement, and communication. In fact, “patient portal adoption is associated with higher medical malpractice exposure,” she said. The reason appears to be that it gives potential litigants more documentation to pursue a case.
Of course, many positive outcomes and innovations were also highlighted.
Research by Kelsey Moran, Rebekah Dix, and Thi Mai Anh Nguyen found that patient care improves—and costs drop—when hospitals in geographic proximity share electronic health record systems.
“If two providers can’t share information, it introduces barriers to care coordination, duplication, and even [causes] patient harm,” said presenter Moran, assistant professor at the University of Miami.
In another example, Benjamin Sprouse, an MD/MBA candidate at the Johns Hopkins University School of Medicine and Johns Hopkins Carey Business School, founded PsychGigs, a company that uses artificial intelligence to create a sophisticated matching system between psychiatrists and psychiatric nurse practitioners and potential employers.
Wenjun Zhou, a professor at the University of Tennessee, Knoxville, also described the positive outcomes of her work with a team that is developing and testing an AI-powered robot to support people caring for family members with dementia. She noted that many caregivers are under their own financial, physical, and emotional stress, and lack professional training in this kind of caregiving.
Move fast, but don’t break anything
One theme that emerged in the conference is that there are many questions yet to be answered as AI use proliferates in all aspects of healthcare. Is it better to integrate AI into existing clinical workflows, or to change the workflows to accommodate the technology? How much responsibility do AI vendors have in malpractice suits? Are there cases when not using AI means failing to meet a standard of care?
“AI is not a monolith,” said Agarwal. “We have to recognize that there are many shades of AI, and each one comes with different functionalities.”
Of course, the federal government has a role in all this, both through financial incentives--- for example, creating codes for insurance reimbursement of AI-provided care---and by setting policies and frameworks.
Jesse Isaacman-Beck, director of the Division of Artificial Intelligence Policy and Strategy at the U.S. Department of Health and Human Services, said the agency is working to promote effective and safe adoption of AI technologies.
“We want AI to provide a practical layer of value,” he said.
He touted the growth of Trusted Exchange Framework and Common Agreement adoption, saying it makes it easier for patients, providers, and researchers to access the information they need.
Yet the old “move fast and break things” mantra once used internally at social media giant Facebook prior to the formation of Meta to prioritize disruption over caution, is far too risky for healthcare.
One panel, titled Who is Responsible? Liability, Governance, and AI in Healthcare, tangled with some of the many legal questions raised by increasing reliance on AI.
“The starting point is: it’s a mess because regulatory and legal considerations have not kept up with the progress made in the technology,” said lawyer, internist, and pathologist Caroline Popper, who shared the panel with University of Pittsburgh Adjunct Professor Oliver Kim and Brenna Loufek, director of Artificial Intelligence — Regulatory & Quality at Mayo Clinic.
Loufek explained how Mayo Clinic is filling the regulatory void, particularly for software that physicians create for their own use, for purposes such as assessing cardiac risk or measuring anatomic features, which do not require FDA approval.
The Mayo framework, she said, takes into account how functional the tool is, how high stakes the situation is, and whether the output can be verified independently.
Ultimately, the answers may lie in entirely new ways of thinking about how humans and AI work together.
“Nothing is perfect, neither humans nor AI,” said Popper. “So where is the responsibility going to lie? Right now, we only have two options: human, or strict product liability. I think at the end of the day there will be a new framework of proportional liability.”
