DeepMind’s Latest A.I. Predicts Kidney Injuries 48 Hours in Advance
DeepMind, the artificial intelligence company owned by Google parent Alphabet, has developed software capable of predicting which hospital patients will develop a common and potentially life-threatening condition up to two days in advance.
The development, which the company made public in a research paper in the journal Nature on Wednesday, brings DeepMind’s health unit closer to being able to market a product that would push A.I.-enabled alerts to doctors and nurses.
Dominic King, the clinical lead for DeepMind Health, said in an interview that the company was “involved in several parallel discussions” with regulators —including the U.S. Food and Drug Administration—about the approvals needed to deploy this kind of technology commercially.
DeepMind’s research was conducted using more than 700,000 anonymous patient records from the U.S. Department of Veterans Affairs hospital system. The company focused on a condition called acute kidney injury (AKI) that contributes to more than 300,000 deaths in the U.S. each year.
By using a kind of machine learning loosely based on how the human brain works, DeepMind said it was able to accurately forecast more than 55% of all AKI occurrences more than 48 hours in advance, and more than 90% of the most severe AKI cases, in which patients required on-going kidney dialysis.
DeepMind is best known for its work creating the A.I. software that could beat the world’s top human players at the strategy game Go, considered a major breakthrough in computer science. But the company had an entire division called DeepMind Health dedicated to using technology in healthcare.
In November, Alphabet announced that it was merging DeepMind’s Health division into a new Google business unit dedicated to healthcare, called Google Health. It has hired former Geisinger Health CEO David Feinberg to head that business. The merger is not yet complete, DeepMind Health said.
DeepMind Health already sells a system called Streams that pushes mobile phone notifications to doctors and nurses if a patient is showing signs of deterioration. So far that system, which has been used by a number of U.K. hospitals to send alerts related to AKI, doesn’t use any artificial intelligence. Instead, it relies on a static algorithm developed by doctors in Britain’s National Health Service that provides a score for how likely a patient is to develop acute kidney injury.
In a blog post published Wednesday, King and Mustafa Suleyman, DeepMind’s co-founder, said the company’s research into an A.I. algorithm to detect AKI along with its Streams alerting app were “the building blocks” of its vision for preventative healthcare. DeepMind, the two said, was committed to “using our machine learning expertise to explore how we might prevent patient deterioration, and building transformative mobile tools that gets that information to the right person at the right time.”
In additional research published Wednesday in other journals but timed to coincide with the Nature paper, researchers at University College London showed that London’s Royal Free Hospital had significantly improved its treatment of AKI through use of Streams. The hospital reduced the average time it took to review urgent cases from four hours to less than 15 minutes. It reduced the number of AKIs that doctors and nurses missed from 12% to just 3%. And, as a result of being able to intervene earlier, the hospital saved 17% on the cost of the average AKI—and that doesn’t even count the cost of on-going dialysis for patients with the most severe kidney injuries.
DeepMind’s work on developing Streams with the Royal Free Hospital has been controversial. In July 2017, the U.K.’s data privacy regulator concluded that the Royal Free had illegally transferred 1.6 million patient records to DeepMind to help it test the safety of the mobile alerting app.
More recently, the merger of DeepMind Health with Google Health outraged some data privacy advocates, as the move seemed to violate earlier assurances by DeepMind that it would never transfer patient data to the U.S. technology giant. DeepMind has said no data has been transferred to Google to date and that no data will be moved unless the hospitals give their approval.
Privacy advocates were also troubled that a review panel made up of outside experts that DeepMind had appointed to scrutinize its healthcare work and report to the public annually, has been disbanded as part of the merger.
The A.I. software DeepMind developed using the VA data scrutinizes some 600,000 different features in patient medical records, Nenad Tomasev, one of the DeepMind researchers involved in the study said. Since doctors can’t understand so many variables, the algorithm is designed to give a sense of which two or three factors were most statistically significant in its prediction that a patient was at high risk for AKI. The algorithm also tells doctors and nurses how confident it is, at any given time, of its prediction. In addition, the algorithm tries to forecast several different blood test results for the patient in the future, as this can give doctors some further insight into why the A.I. thinks the patient is at risk, Tomasev said.
While the results sound impressive, the A.I. system wasn’t perfect. At 48 hours from AKI onset, it wrongly predicted someone was at risk for AKI twice for every accurate prediction it made. The researchers said this false positive rate could be tuned downward by doctors, but at the expense of having to wait longer to pick up AKIs and potentially missing some altogether. The VA dataset also skewed heavily toward men—only about 6% of the patient records were from women—and, as a result, DeepMind’s A.I. performed worse on detecting AKI in women.
DeepMind is also hardly the only group of researchers to apply A.I. techniques to AKI. Relying on somewhat different technology, and 300,000 patient records from Stanford University Medical Center and Beth Israel Deaconess Medical Center in Los Angeles, California researchers were able to detect AKI up to 72 hours in advance with almost 73% accuracy.
DeepMind, in its paper, said it had tried this other machine learning technique—which is called a boosted-gradient decision tree—and found it was far less accurate when using the VA data. It was able to detect just 36% of AKIs 48 hours in advance compared to DeepMind’s 55.8% result.
The disparity between how this technique performed when DeepMind tried it compared to what the California researchers found may reflect differences between the data used in the two studies, which is why either technique would have to pass a multi-site clinical trial using real patients before doctors can start using it regularly.