What you need to know about the forces reshaping our industry.


August 28, 2019

DeepMind's new AI system can predict 90% of severe kidney injuries—up to 2 days in advance

Daily Briefing

    DeepMind has developed an AI tool that can predict the onset of acute kidney injury "with a lead time of up to 48 hours," according to a paper published in Nature last month.

    Innovation 101: Download the cheat sheets for today's digital world

    DeepMind's AI model

    Acute kidney injury is difficult for clinicians to predict, and once it afflicts a patient, it usually progresses very quickly, according to VA officials. So DeepMind, an artificial intelligence lab owned by Google parent company Alphabet, collaborated with Veterans Affairs (VA) to develop a new model that could detect patients at risk of the condition sooner.

    The researchers used de-identified EHR data—including blood tests, past medical history, and vital signs—from more than 700,000 adult patients across 172 inpatient and 1,062 outpatient sites to teach the model to provide "continuous risk prediction of future deterioration in patients" at risk of kidney injury.

    The researchers in the paper said their model correctly predicted 90% of the most severe cases of acute kidney injury—those that eventually required dialysis—"with a lead time of up to 48 hours and a ratio of two false alerts for every true alert."

    The model "also provides predicted future results for several relevant blood tests," Mustafa Suleyman, co-founder of the lab, and Dominic King, the lab's health lead, said in a blog post.

    Breaking out of the 'black box'

    According to Suleyman and King, earlier detection of the condition through DeepMind's AI system can "provide a window in the future for earlier preventative treatment and avoid the need for more invasive procedures like kidney dialysis."

    "Although the recognition and prompt treatment of acute kidney injury is known to be challenging, our approach may offer opportunities for identifying patients at risk within a time window that enables early treatment," they wrote in the blog post.

    Suleyman and King added that the model also seeks to overcome AI's notorious "black box" problem, as—rather than simply offering a prediction without further context—the model can display the clinical factors most relevant to its predictions. "This information may help clinicians understand the reasoning behind the AI-enabled alert and anticipate future patient deterioration," they wrote.

    However, while the model effectively predicted acute kidney injury in patients whose conditions were rapidly deteriorating, it still failed to predict about 44% of all inpatient cases of the condition. 

    "This perhaps points at the need to look into other data sources that may paint a more complete picture of the patient's clinical reality," according to L. Nelson Sanchez-Pinto, a researcher at Northwestern University who was not involved in the paper.

    According to VA officials, VA Palo Alto Health Care System in California is currently looking into integrating the new system into clinical settings (Landi, FierceHealthcare, 8/2; Metz, New York Times, 7/31).

    Have a Question?


    Ask our experts a question on any topic in health care by visiting our member portal, AskAdvisory.