Getty Images

Views on artificial intelligence and machine-learning in medicine alternate between two extremes: They either tout AI as the cure for every ill, or vilify it as a vector for spreading systemic bias.

Researchers find themselves managing expectations even as they strive to improve outcomes.

“One of the things that our patients have told us really, really strongly is that they really resonate with the idea of AI as a tool, because they can envision a person that is accountable for their care who is making a decision that also incorporates their values, their perspective and their context,” said Dr. Ziad Obermeyer of the UC Berkeley School of Public Health.

Obermeyer numbered among the experts at a STAT News roundtable this week who agreed that AI systems are, and should remain, tools for doctors and nurses to use — each with their strengths and weaknesses, just like any test or scan.

They underlined the need to identify an AI’s purpose and then to test it with that specific outcome in mind.

'A trust-but-verify situation in the algorithm world'

UC Berkeley

Dr. Ziad Obermeyer is an associate professor of health policy and management at UC Berkeley who works at the intersection of machine learning and health.

“We don't tend to take Merck or Pfizer at their word, when they tell us, ‘Oh, in our data, it looks great, you should approve this drug,'” said Obermeyer. “It also would be weird if we were taking people's word on the fact that the algorithm's performing great in their data, and we didn't have a trust-but-verify situation in the algorithm world as well.”

But finding people with that particular blend of skills remains one of the undiagnosed problems facing medical AI.

“You need people who both really understand the data, and people who understand the clinical realities and how to turn that data into a diagnosis,” said Obermeyer.

He said he worries that neither medical nor technical education programs are producing graduates who can straddle both worlds effectively.

“It takes a ton of very specialized knowledge to go through this incredibly messy data that's the exhaust of the health-care system and turn that into something that is credibly a real diagnosis that AI can learn.”