Sanford Health partners to overcome data ‘bottleneck’ hindering medical AI tools

• Read original article here

SIOUX FALLS, S.D. — The power of computer-driven artificial intelligence has offered tantalizing promise as a tool to help doctors diagnose and treat illness, but so far has failed to live up to lofty expectations.

Elliot Green and his co-founders of Dandelion Health Inc. decided to investigate why artificial intelligence in the clinic had so far yielded disappointing results.

“We saw it has the potential to do so much good and we couldn’t understand why it never got off the ground,” said Green, the company’s chief executive officer.

“We discovered it was a data problem,” he said, “it wasn’t a technology problem.”

It turned out that the massive volume of information required for computers to teach themselves how to perform tasks wasn’t enough, a lack of data that was stifling innovation in clinical applications of artificial intelligence.

So Green and three partners formed Dandelion Health to provide a data platform that aggregates massive data to enable researchers to develop new artificial intelligence applications for the clinic.

They quickly sought health systems as partners to provide the data, and Sanford Health was one of the first to sign on, along with San Diego-based Sharp HealthCare. Sanford is the country’ largest rural health system, with a 250,000 square-mile service area.

“We believe this collaboration with Dandelion and Sharp will strengthen the ways we are able to deliver care and improve lives,” said Kent Lehr, Sanford’s chief business development officer.

“We see a great opportunity to leverage data in order to build innovative clinical tools and products and solutions,” he said.

Sanford’s participation ensures that rural populations won’t be ignored in the data sets that are used to develop artificial intelligence applications for clinical use.

To date, some of the leading examples of artificial intelligence have been in radiology. Computers are good at examining the pixelated information in medical images, Green said, and sometimes can spot things the human eye can’t detect.

Given enough data, the potential to help physicians diagnose and treat patients through artificial intelligence technology is great, Green said. “Realistically, we’re barely scratching the surface,” he said.

The idea driving the search for clinical applications of artificial intelligence is simple: “Where can a machine do what a human can’t?” Green said.

Artificial tools in the clinic are to assist, not replace, physicians, he said. "That's what it should always remain, an augmentation."

Dr. Jeremy Cauwels, Sanford’s chief physician, agrees that artificial intelligence holds promise that will benefit patients.

“For me the real benefit of this is to identify paths that may not be evident to a doctor,” who lacks the time and is not as well equipped to sift through massive amounts of information, he said.

Computers, on the other hand, excel at analyzing reams of information. “Right now, many of these projects are in what I call the infancy stage,” Cauwels said. “We don’t have a great database to train the computers on.”

When fed truckloads of data, computer intelligence grows. A famous example from 2012 is the development of an artificial brain by Google. After analyzing 10 million YouTubes, it was able to recognize a cat — a skill that it performed with 75% accuracy after examining more images for three days.

Sanford has physicians and researchers that have ideas for artificial intelligence applications, and they will have access to the combined data provided by Dandelion.

Sanford also will get priority in the applications that are derived from the pooled data.

“We’ll get kind of first opportunity to deploy those tools,” Lehr said.

The partnership, he said, provides the opportunity “to try to bring the promise of AI” — shorthand for artificial intelligence — “to reality.”

All identifying patient information is stripped from the data provided to Dandelion, where it is stored securely in the cloud, where it can be analyzed by developers, but can’t be downloaded or copied without express permission from the health system, Green said.

“We’re not interested in selling data from our patients,” Cauwels said. “We get to keep the data. You have to come into our ‘sandbox’ and do it our way.”

Protecting patient privacy in the use of the data is of paramount concern, Lehr said. “I can say without question the No. 1 priority we’ve had in the conversations we’ve had is patient privacy and protecting that,” he said. “We do get to maintain control of that data at all times.”

Individual patient information, which is unavailable to those using the Dandelion database, is of no use to the developers, Green said.

Dandelion is in “active discussions” with other health systems, with the goal of signing up five partners, he said.

Artificial intelligence developers will have access to the dataset soon, as early as one or two months, Green said.

“We have people waiting to use it,” he said.

In recent years, IBM has announced Watson Health, an artificial intelligence initiative to efficiently identify symptoms of heart disease and cancer.

Sanford University has announced it is working on an artificial intelligence-assisted support system to detect behavioral changes in elderly patients who are living alone or in intensive care unit patients.