Artificial intelligence may be able to help doctors figure out which already-sick patients are at serious risk of sudden kidney damage, as much as two days before it happens, according to new research. The technology could potentially save lives and lessen the need for uncomfortable, invasive treatments such as dialysis.
A paper describing the work — which was conducted by Alphabet’s AI research company, DeepMind, and collaborators including the US Department of Veterans Affairs — was published this week in the journal Nature.
Kidney issues are well known to make already sick patients even sicker. The researchers built and tested software that can predict acute kidney injury with the help of a dataset made up of electronic health records from over 700,000 adult patients. The software was able to look at anonymized health records of hospital patients and predict nearly 56% of all serious kidney problems they would encounter, and about 90% of kidney issues so serious that they required dialysis. The problems could be detected as far as 48 hours in advance of them taking place.
“Hopefully, in the not-too-distant future, doctors and nurses may start getting a warning a day or two before for these acute causes of patient deterioration,” Dr. Dominic King, DeepMind’s health lead and coauthor of the research paper, told CNN Business.
The work is still in the early stages, and there are some caveats that accompany the results: The researchers noted their system gave two false alerts for every true alert of a kidney injury. And nearly all the patients in the dataset the researchers used (about 94%) were male, so it’s not yet known if the work would be similarly helpful for spotting kidney failure in women.
But it advances what’s known about how deep learning — a form of AI modeled after the way neurons work in the brain, which ingests loads of data and learns to make its own predictions — may be helpful in healthcare.
While deep learning is already known to be useful for identifying health issues such as skin rashes or cancers in scans, slides, and photos, it’s not as advanced when it comes to predicting health problems.
Kidney issues in particular are tricky to identify in advance. These days, doctors and nurses are alerted to acute kidney injury via a patient’s blood test, King said, but by the time that information comes through, the organ may already be damaged.
The research isn’t yet ready for clinical practice but it’s impressive, said Eric Topol, a professor at Scripps Research and author of “Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again.” He believes it could eventually be helpful in part because doctors are simply bad at making predictions about what will happen to individual patients.
“We think someone’s going to die and then they’re like Lazarus, or we think they’re going to not be readmitted and they come back in an hour. We’re just not good at this,” he said. “Applying deep learning for this common and serious issue, kidney injury — that’s really smart.”
Eventually, King envisions alerts about potential kidney problems coming straight to doctors and nurses via a smartphone app. And, in fact, a DeepMind healthcare app called Streams is being used in National Health Service hospitals in the UK; it can alert clinicians if results of a standard blood-test indicate a problem with a patient’s kidneys. Adding an AI-based prediction to this, King said, would let doctors and nurses know far sooner about a potentially serious health issue.
Before rolling an AI system out with any patients, though, King said the researchers would need to make sure their model works accurately with patients of different genders and backgrounds.