Computer scientists have long promised that artificial intelligence could make healthcare work easier for doctors and ultimately better for patients. Around 10 years ago, IBM said its Watson technology would help doctors diagnose and treat patients. But no company has yet come close to delivering on that promise; in fact, according to two leading AI experts, artificial intelligence probably won’t transform healthcare for another decade.
“In the tech world, progress tends to happen slowly and then very quickly,” says Andrew Ng, founder of Google Brain, former chief scientist at Baidu, and current adjunct professor at Stanford. “I think we’re still in the progress-is-happening-slowly process in healthcare.”
On Thursday, Ng and Fei-Fei Li, codirector of the Stanford Institute for Human-Centered Artificial Intelligence, discussed how they see AI integrating into the healthcare industry at a panel hosted by the institute. Ng said that he doesn’t anticipate major breakthroughs in the next few years. Rather, he hopes to see AI in healthcare “blossom” over the next decade.
The Obama administration paved the way for this eventual future by incentivizing healthcare systems to digitize patient health records through the HITECH Act and the Affordable Care Act. But the promise of AI-powered health has not materialized, even though the healthcare industry could certainly use the help. More than 40% of doctors are suffering from burnout, according to a 2021 Medscape survey—and the vast majority of them say they were burned out even before the pandemic. The main agents thrusting them toward exhaustion are bureaucratic tasks, like inputting notes, that often need to be completed in their off time. There are big concerns that pervasive burnout among clinicians could lead to poor outcomes for patients.
Some companies, like Nuance AI, which was recently acquired by Microsoft, are developing tools that could potentially write a doctor’s notes for them. But so far, the technology is little more than a transcription service. There are still major hurdles to making technology good enough that it can truly assist doctors in a way that reduces their workload.
“We need AI and technology to prove that they make a fundamental human difference to the care, recovery, or the well-being of . . . the patient or to the healthcare worker’s work,” Li said.
Healthcare is a high-stakes business. When there are errors in medicine—a missed diagnosis, a wrong medication administered—people can die. Ng says there are multiple barriers to getting artificial intelligence to work inside of healthcare. He says the algorithms need to be better and they need to be tackling the right problems.
For instance, when Ng’s researchers test and train their algorithms in Stanford’s hospital using its data and machines, they can prove that the technology is as good as Stanford’s radiologists at spotting various conditions in X-rays. But, he says, there is a gap between this proof of concept and actual in situ use. If a researcher were to take that same algorithm down the street to another hospital where the X-ray technician uses a slightly different imaging protocol, the AI system would perform poorly, because the conditions are not exactly as the ones in which the algorithm was tested.
“In contrast, any human radiologist can walk down the street to the other hospital and do just fine,” Ng said.
Another problem concerning AI in healthcare is that not all of the necessary data is digitized. Li noted that an important aspect of healthcare that isn’t documented is human behavior, “whether you’re talking about patients living at home with chronic disease and hoping they never have signs of deterioration or you’re talking about doctors carrying out their important procedures, hoping that the protocols are met and there’s no clinical error.”
What’s called ambient artificial intelligence, or technology that sits seamlessly in the background of an environment passively collecting data, could play an important role, for instance, in monitoring whether doctors are conducting procedures appropriately. Another possibility: alerting a doctor to a patient in intensive care who’s taken a downturn. But right now, the technology is not there yet.
“We need, for example, in radiology, a proven product or proven story that truly has not only helped diagnose patients but also made radiologists work better and collectively moved the needle,” Li said. “We’re inching toward that watershed moment.”