The human body is frail and people end up in intensive care units for all kinds of reasons. Whatever brings them there, more than half of adults admitted to an ICU end up sharing the same potentially life-threatening condition: kidney damage known as acute kidney injury.
The Veterans Administration thinks artificial intelligence could reduce the toll. In a project that drew on roughly 700,000 medical records from US veterans, the agency worked with Google parent Alphabet’s DeepMind unit to create software that attempts to predict which patients are likely to develop AKI. The VA hopes to test whether those predictions can help doctors prevent people from developing the condition. AKI manifests as a sudden failure of the kidneys to properly remove waste from the body, and often occurs as a complication of surgery, infection, or other stresses of hospitalization.
The project is an example of the worldwide push to save lives using the AI techniques that power internet companies’ virtual assistants and facial recognition. The spread of digital health records offers a torrent of data about patients, including subtle patterns that algorithms can interpret in ways doctors cannot. In the US and other rich countries, AI is seen as a way to improve care and cut costs. In places like India and China with chronic shortages of medical specialists, the technology could improve access to care.
DeepMind’s collaboration with the VA fits into a broader push into health care by Alphabet. The company hopes to use AI to diversify beyond advertising, which supplies nearly 90 percent of its revenue. Other Alphabet projects are training algorithms to detect eye disease and cancer. Google recently hired veteran health system executive David Feinberg to take charge of its health projects.
The VA collaboration also illustrates a challenge to Alphabet’s health care ambitions. The company has a world-beating roster of AI researchers. But in health care it lacks the kind of data troves that power Google’s dominance in search and online ads. Only by teaming up with organizations willing to share piles of medical data can Alphabet get the feedstock needed to train machine learning algorithms. The VA’s millions of electronic health records represent one of the largest collections in the US. A DeepMind spokesperson cited the VA’s leadership in kidney disease and health analytics, and the fact it has “one of the most comprehensive electronic datasets covering patient care.”
The VA’s engagement with DeepMind began a few years ago, when the agency’s director of predictive analytics, Christopher Nielsen, received an unexpected phone call. “It’s not uncommon to get calls from people saying I can solve all your problems with AI,” Nielsen says. He has learned to be wary of out-of-the-blue AI pitches.
But this call came from Mustafa Suleyman, who cofounded DeepMind before it was acquired by Google in 2014. The company has a track record of breaking new ground in machine learning, including bots that beat Atari games and masters of the board game Go. Early in 2018, the VA announced that it had signed a formal research agreement with DeepMind.
Right away, Nielsen and his VA colleagues had to tackle a common hurdle for AI health care projects. The machine learning algorithms driving the AI boom need large amounts of example data to learn from; typically, the more data, the better the results. But when the data consists of people’s most private information, it must be treated with special care.
VA researchers and engineers developed a process that uses cryptographic hashes to obscure lab results and other data in a health record, Nielsen says. It was used to give DeepMind access to a sanitized collection of hundreds of thousands of health records from a 10-year period. AI experts at the company used some of Alphabet’s US computing infrastructure to train neural networks—the guts of much of today’s machine learning—to predict when a patient is likely to develop AKI.
Full results will be detailed in a forthcoming scientific paper, but results have been encouraging, Nielsen says. “It’s been fairly successful in predicting AKI at an early enough stage to prevent it,” he says, declining to discuss any of the factors that were identified. Data provided by the VA during the project remains the property of the agency, and will be destroyed after use.
The project’s next phase will probably be to feed in live data from the millions of patients in the VA’s system and track the accuracy of DeepMind’s AKI predictions over time. If that goes well, Nielsen wants to test the system with doctors in a VA clinic to see if it helps improve care. He anticipates that will be at least a year away.
DeepMind is working with the VA under what is known as a Collaborative Research and Development Agreement. The two organizations work together without money changing hands, and can both make use of ideas developed in the project. Laurence Meyer, head of specialty care services at the Veterans Health Administration, says the VA could end up offering tools developed in the program to others. “We are interested for our own purposes, and in developing things that would potentially be useful outside the VA,” he says.
Scott Sutherland, an associate clinical professor in nephrology at Stanford, says getting AKI prediction technology into the clinic could be revolutionary. The condition is very common in critically ill patients, but once tests detect it, doctors can only prevent further damage, not directly treat the injury itself.
Previous attempts to use technology to predict AKI have not yet been fruitful. “I have not seen any truly successful big data or machine learning algorithms to date,” says Sutherland. Most work in the field has used more established statistical techniques, he says, not the neural network technology that is DeepMind’s specialty.
Getting the AI software to produce accurate predictions will only be part of the effort needed to transform care in hospitals—a common feature of AI health care projects. Because doctors haven’t previously been able to forecast AKI, it will take additional clinical research to figure out the best ways to stave it off, Sutherland says. “There’s not a ton of data to say this is clearly what you should do,” he says.
More than half of adults admitted to an ICU are afflicted by acute kidney injury.
DeepMind has spent two years testing an app with hospital staff in the UK that could be a vehicle to examine that question in the clinic—and eventually to productize its research with the VA. The app, called Streams, helps hospital staff monitor patients’ test results to spot AKI, without the help of AI technology.
One hospital involved was censured by the UK’s data regulator for giving DeepMind over-broad access to patient data. The company escaped official blame, and announced in November that the Streams project will be transferred to Google so that it can be made into a product under Feinberg, the company’s new health boss. The DeepMind spokesperson said the company hopes to see AI-powered alerts in Streams, but that it would require extensive work, as well as regulatory approvals.
The way DeepMind is handing off Streams suggests that it will remain primarily a research unit of Alphabet, in line with its founders’ interest in making AI as capable as humans, rather than become a sustainable business more like Google. Financial statements filed in the UK indicate that the division lost £302 million ($390 million) in 2017, triple its losses in the previous year.
Streams is not part of DeepMind’s research collaboration with the VA. Nielsen says the VA project is not transferring to Google, but may expand. The agency’s rich data trove and the protocol it developed to scrub data before transferring it to DeepMind offer the potential to try and predict other health problems in hospital patients early, he says. Possible targets include septicemia, heart attacks, or falls.