IBM Watson manager, academics describe challenges, potential of healthcare AI

By Dave Muoio
Share

Last week, hundreds of digital health entrepreneurs, investors, and executives met in Boston for the annual Digital Healthcare Innovation Summit. Among the more frequently discussed trends and technologies was artificial intelligence — its promise, its stumbles, and how it should be implemented to best serve healthcare.

“I started marketing Watson for Oncology in January of 2016. I’m almost approaching two years,” Deborah DiSanzo, general manager at IBM Watson Health, said during a session. “By the end of this year, I will have over 20,000 patients [and] over 120 hospitals using it, and really seeing helping oncologists all over the world. Nothing that I have done in my life in healthcare technology has gone as fast as that, and that is not hype.”

On AI and health data

DiSanzo said that AI systems such as Watson or Google’s DeepMind have an opportunity to redefine how large sets of data are interpreted. This capacity can lead to a number of clear advantages, including streamlined clinical trial enrollment, discovery of new interventions, and, with the help of blockchain, consolidation of personal EHRs.

Unfortunately, data management and collection still has a way to go before AI will be able to make a noticeable impact, Dr. Atul Butte, director of the Institute for Computational Health Sciences at the University of California, San Francisco, said.

“As a country, we spent hundreds of billions of dollars on these systems, and [relatively] zero on using any of this, analyzing it, interpreting it — and I would also argue that this is the most expensive data in America because we’re asking doctors to enter it,” Butte said during the summit. “We are trying to teach computers how to do a lot of our thinking with half the game, and we’re going to have to integrate to get to better ways to teach.”

A major contributor to these shortcomings is the ubiquity of siloed health databases, Butte explained, but the lack of readily available data from clinical trials or diverse patient populations is also limiting what AI can accomplish. For DiSanzo, these needs pale in comparison to what could be accomplished once technology is able to capture and process how an individual’s behaviors are affecting their long-term health.

“Seventy percent of our determinants in health are not in our clinical or genomic data,” she said. “We analyzed 650 million data points from one activity. We could find, really, no good actionable insights [because] it’s very difficult to get good social determinants in health data.”

Watson's MD Anderson setback

DiSanzo also fielded questions from session moderator Lisa Suennen, senior managing director of healthcare investing at GE Ventures, about IBM’s well-publicized falling out with MD Anderson earlier this year. While some critics viewed the company’s failure to properly integrate Watson into cancer care as a setback for AI, DiSanzo reminded the audience that the partnership was always viewed within IBM as experimental. Further, she argued that the extent of the failure and its implications of the field were largely overblown.

“We are out in front. I have 1,000 data scientists working on AI in healthcare. They’re smart people, they’re making a difference in the world, but when you’re out in front you are the target of controversy,” she said. “I wasn’t there, but I’m very proud of what the Watson Health Team did with MD Anderson. … We don’t believe that we failed at MD Anderson.”

Dr. Isaac Kohane, chair of the Department of Biomedical Informatics at Harvard Medical School, lamented the both what he saw as the failure of the MD Anderson project and the public response, which he said detracts from meaningful progress by scaring away potential investors and innovators.

“One of the top reported facts was that MD Anderson had created a platform for leukemia-based diagnoses, over 150 potential protocols, researchers use Watson to make platform, blah blah blah,” he said during a session. “It was never used, it didn’t exist, and those of us who were in the know didn’t have to wait for this thing to happen. It didn’t exist, and this is very bad because … it actually distracts us from investment in what’s really happening.”

Augment, don't replace

Both DiSanzo and Kohane reaffirmed that AI should not be viewed as a replacement for trained practitioners, with the former describing her distaste for buzzphrases such as “Dr. AI.” 

“They like to say that AI is going to kill radiologists, and I’d really like to kill that [phrase] because that’s not what’s going to happen," DiSanzo said. "AI is there to augment intelligence and help our physicians, radiologists, everyone else to do their job, and I would like all these 'Dr. Things' out of technology.”

Kohane took a more clinical approach, telling first-hand stories of how AI has been used to quickly draw diagnoses from complicated and disparate health records, and to identify potential treatments through bulk gene analysis. Both of these approaches are examples of how healthcare AI can augment, instead of replace, physicians.

“These are superpowers, and that’s what we should be looking at for AI,” he said. “We shouldn’t have AI for what doctors can do, and we should have AI for what doctors can do but won’t, but let’s really focus on what doctors can’t do.”