Ethical worries are marring Alphabet's AI healthcare initiative
(Bloomberg) -- DeepMind, the digital brain foundry owned by Google’s parent company, Alphabet, wants to use artificial intelligence to solve… well, everything. Last year, its software taught itself to play the strategy game Go better than any human on the planet. For its next trick, it wants to move beyond games to a very real-world problem: health care.
The London company has a fast-growing division—now 100 strong—dedicated to health. And while DeepMind’s research on Go may be years away from yielding practical applications, its health-care work is affecting people’s lives today through projects with the U.K.’s National Health Service. These include a mobile app to alert doctors and nurses to changes in a patient’s condition and efforts to research whether computers can analyze various kinds of medical imagery as well as experienced doctors.
The company believes AI has the power to save lives. But DeepMind’s maiden voyage into the field has also run smack into an iceberg of privacy and ethical concerns—and the resulting controversy has threatened to sink its ambitions of using AI to transform health care.
In July, after a year-long investigation, U.K. regulators ruled that London’s Royal Free Hospital had illegally provided DeepMind access to 1.6 million patient records going back five years. DeepMind said it needed the records to conduct safety testing of its first product, a mobile app that gives doctors and nurses instant access to medical records and can alert them to patients at risk of deterioration. The first potentially fatal condition DeepMind built an alert for was acute kidney injury (AKI). The Royal Free said it accepts the decision, but disagrees it could have tested the mobile app, called Streams, another way.
Regulators took no action against DeepMind, ruling that it acted on the Royal Free’s instructions. But the company acknowledges it made mistakes. “In our determination to achieve quick impact when this work started in 2015, we underestimated the complexity of the NHS and of the rules around patient data, as well as the potential fears about a well-known tech company working in health,” Mustafa Suleyman, DeepMind’s co-founder and head of its health projects, and Dominic King, the former surgeon who serves as the company’s top clinician, said in a statement.
DeepMind’s stumble has implications far beyond one AI research firm. Tech companies are flooding into health care. IBM claims its Watson artificial intelligence software can help doctors find the best treatments for cancer. Genome pioneer J. Craig Venter’s latest startup, Human Longevity Inc., wants to customize treatments for each patient’s DNA.
Even DeepMind Health is just one of three big health-care bets Alphabet is making. It also owns Verily, which creates medical device software, and Calico, which is trying to stretch human lifespans. Success or failure matters to more than just corporate bottom lines: the U.K.’s NHS is counting on technology to cope with an aging population and shrinking budgets that threaten to bankrupt the entire system. But if AI is going to fulfill the optimists’ hopes, the companies behind it must prove they are trustworthy.
And when it comes to Big Tech, trust is in increasingly short supply. From revelations about Russian meddling in the U.S. presidential election to new disclosures about how Apple and Google have avoided paying taxes, tech companies are no longer seen as benign — or even neutral — entities. DeepMind, founded in 2012, may still see itself as a startup, but it is a part of Alphabet, a huge conglomerate. And while Silicon Valley’s clichéd “move fast, break stuff” ethos might have worked in the past, DeepMind is discovering that sometimes it really isn’t better to ask for forgiveness than permission.
The driving force behind DeepMind’s push into medicine is Suleyman, who everyone at DeepMind calls “Moose” (an abbreviation of his first name). Suleyman’s mother was a nurse. After Google bought DeepMind in 2014 for a reported 400 million pounds, he quickly homed in on health care. “There is no other area where we invest so much money in technology and get so little back,” Suleyman said in an interview in mid-August.
Press coverage of DeepMind’s health-care efforts sometimes makes it seem as if the company is developing a software version of Hugh Laurie’s character in “House," a diagnostic genius able to deduce the solution to any medical mystery. But Suleyman said this is “total nonsense.” “We are going to be solving all kinds of other magical problems in the world before we get to that sort of general diagnostician,” he said.
Hints of what DeepMind Health does want to do can be gleaned from a project at London’s Moorfields Eye Hospital. Here, in an office cluttered with thick medical tomes, Pearse Keane is staring at an image on his laptop. Keane is a senior ophthalmologist and clinical researcher. The picture is a patient’s retina imaged with optical coherence tomography, or OCT. “It allows us to see things like bleeding and leakage into your retina and diagnose the most common causes of blindness,” Keane said.
Moorfields and DeepMind are trying to see if a computer, using AI, can read OCT scans as well as Keane. The project has achieved “impressive” results, Keane said, but he wasn’t ready to talk about them yet. He and DeepMind hope to publish their research in the coming months. The company has also announced research projects with two London universities to see if AI software can learn to read head and neck scans and mammography scans as well as or better than doctors.
Still, DeepMind said a commercial product using AI is a ways off. Streams, the only product DeepMind has actually deployed, uses no AI. While DeepMind originally set out to use machine learning to improve an existing NHS algorithm to detect AKI, it said it never carried out that research. When DeepMind visited the Royal Free, it found the existing algorithm— which wasn’t half bad— was the least of the problem. Of far more concern were antiquated technology and Byzantine workflows that meant it took too long for doctors and nurses to act on blood test results. The real problems in medicine “are much more gritty and practical,” Suleyman said.
Those pragmatic concerns are front-and-center on the ninth floor of the Royal Free hospital in early August, when a patient’s kidneys suddenly start struggling after a liver transplant. Within seconds of a lab pathologist entering blood test results into a computer database, they are analyzed by a formula the NHS developed, and an alert sounds on nurse Sarah Stanley phone. Opening the Streams app, she sees a graph showing spiking indicators from the blood tests. Using the app, she messages a colleague to check on the patient.
“We have just triaged that patient in less than 30 seconds,” she said. In the past, the process would have taken up to four hours. A few hours delay, Stanley said, can be critical: patients with AKI can deteriorate rapidly.
DeepMind said it wants to move on from the controversy over Streams’ development. In November 2016, it replaced its original information sharing agreement with the Royal Free with a new five-year contract designed to address the initial deal’s failings. Since then, the company has published, with a few redactions, copies of its contracts with hospitals. It set up and funded a panel of outside reviewers to investigate its work and report publicly each year. The company also announced it will create a digital ledger system—similar to the blockchain technology that underpins the cryptocurrency bitcoin—that would give NHS hospitals a tamper-proof audit trail of who has accessed patient data.
But none of this has assuaged the company’s detractors. Julia Powles is a law professor at the University of Cambridge who has written critically of DeepMind’s initial agreement with the Royal Free. She said DeepMind’s new contract does not explicitly prevent it from transferring patient data to sister company Google. DeepMind said it has not and never would give any data to Google. “If they can’t put that in writing I find it hard to believe them,” Powles said.
She questions whether DeepMind was the best choice given that Streams doesn’t use AI, and thinks that other companies should have been allowed to bid on the project. And while DeepMind has been offering Streams to hospitals for free, Powles wonders if the company has an unfair advantage because it is backed by Alphabet and can afford to lose money on Streams. She said there was a danger hospitals will get locked into technology that they won’t be able to afford if DeepMind eventually decides to charge a market rate.
The Royal Free is not entitled to any money DeepMind makes from Streams. But John Bell, a doctor who chairs the U.K.’s Office of Strategic Coordination of Health Research, recently recommended that the NHS retain an economic interest in any artificial intelligence developed using its data. “Our projects so far have assumed the right way to give value back to the public is through initially providing our resources and technologies to our NHS partners for free,” DeepMind said in an emailed response, adding that it was open to discussion of other ways of valuing its services. The Royal Free said it in a statement that it was “happy with the terms of the agreement” with DeepMind.
Controversy hasn’t stopped DeepMind from signing agreements to deploy Streams to additional NHS hospitals. Suleyman said the company has received interest from U.S. doctors, too. But privacy concerns may have dented DeepMind’s hopes of integrating AI into its health-care offerings anytime soon. Taunton & Somerset NHS Foundation Trust, one of the hospitals now adopting Streams, explicitly ruled out doing anything with artificial intelligence, Tom Edwards, the hospital’s joint clinical information officer said. Nicola Perrin, the head of Understanding Patient Data, a project run by the British health charity Wellcome Trust, worries that what happened to the Royal Free might deter U.K. hospitals from adopting potentially life-saving technology.
“I think it is very important that we don’t get so hung up on the concerns and the risks that we miss some of the potential opportunities of having a company with such amazing expertise and resources wanting to be involved in health care,” she said.
Health care, Suleyman said, “is incredibly valuable, it is incredibly broken and there is a massive opportunity to transform it with AI at some point in the future.” But not today. Suleyman describes DeepMind as “getting in early” with products like Streams, so that it is well-positioned to use AI later.
DeepMind is “years away” from generating reliable revenue from health care, Suleyman said. The company earned just 40 million pounds of revenue in 2016—none of it from health work—and reported a loss of 94 million pounds, according to accounts filed with UK business registry Companies House. As DeepMind is learning, changing the way people experience health care—and turning a profit—makes beating the world’s best Go players look easy.