Eric Schmidt: Until AI’s “unknown errors” can be explained, the systems will remain advisory

Elizabeth Gardner | April 23, 2019

Key Takeaway: AI algorithms need clean, complete data to fulfill their promise of transforming health care, says former Google and Alphabet chairman Eric Schmidt

When is artificial intelligence going to change health care? Only when providers are prepared to give it the data it needs, according to Eric Schmidt, former executive chairman of Google and Alphabet, who shared his insights with venture capitalist John Doerr during a riveting discussion at the Health Evolution Summit.

AI is making remarkable strides—for example, in reading radiology images at a pixel-by-pixel level that can catch tumors that radiologists often have a hard time detecting—though the algorithms are far from ready to take over from humans. “People don’t trust computers when it comes to their own lives,” Schmidt said. “AI systems have unknown errors in them—they can’t explain why they did something or how they know something. For the foreseeable future these systems will be advisory.”

However, even in an advisory role, AI can do remarkable things, if it has clean, complete data. Current information technology architectures “incarcerate the data” in multiple non-interacting silos, Schmidt said. “You’ve got to get the plumbing right,” he observed, making a pitch for Google’s HIPAA-compliant cloud services. “You can take all your disparate data, put it in our system, and our AI system will normalize the data into a single model. You can span across all the data records that you have from your disparate systems. We can do this in an hour or two. Nobody else can do this. So anyway, run to a cloud— ideally ours—[because] you can’t achieve the vision that we’re going to talk about without having cloud-based servers.”

Even the data currently in electronic health records is only a small fraction of the data that is available and might be useful, Schmidt added. “Only 20% of the data that’s in the health care system is actually in the EHR. 80% of the data, which is what the nurse did and what the doctor did and what the radiology device did, is not captured anywhere or it’s trapped in some device that if the power’s off, you lose the memory forever.”

Schmidt had a recent encounter with the health care system that illustrated in microcosm the extra effort and cost associated with lack of connection. Feeling unwell, he went to have a CAT scan. “It goes back and forth and then you wait an hour, they hand you a DVD that you can’t read,” he said. “Then the next day they fax a report which said that I needed a calcium score in my heart, which I didn’t know what that was. I said, okay, no problem. I make an appointment for another CAT scan where they do the same thing and they give me another DVD that I can’t read and they give me reports that are faxed to me by doctors I haven’t met. This is insanity, right?”

Here’s what would have happened instead, Schmidt said, with properly connected systems and a strategic use of AI: “You go in, it does the CAT scan and then it should have said, ‘Hey let’s run the machine back and forth one more time because Eric is sitting there (and he’s going to do what the nurse said anyway), look at the heart and do the calcium score. Give that all to the doctor and give [Eric] an answer within an hour.’

“By the way, I was fine,” he added.

About the Author

Elizabeth Gardner, Staff Writer