Listen to this article
Artificial intelligence can now warn critical care doctors that their patients are at risk of developing severe kidney damage up to two days early, with the potential to save hundreds of thousands of lives every year.
The machine learning model was developed jointly by DeepMind Health, a division of Google, and the US Department of Veterans Affairs (VA), a federal agency that provides healthcare services to military veterans across the US.
Acute kidney injury affects up to one in five patients admitted to hospitals in both the UK and US, and is responsible for an estimated 1.4m deaths per year. Patients with AKI are unable to process and remove waste as a result of sudden kidney failure, which occurs as a common hospital complication of anything from surgery to infection.
“It’s notoriously hard to pick up on, and if you don’t, many thousands end up dying, or needing [kidney] transplants,” said Dominic King, clinical lead at DeepMind Health. Experts believe that up to 30 per cent of cases of the disease, which is traditionally identified only two or three days after the kidney starts deteriorating, could be prevented if a doctor intervened early enough.
The future of diagnosis?
Healthcare is one of the most promising areas likely to be transformed by machine learning systems, which are able to sift through large amounts of data quickly and find meaningful patterns.
AI is already being tested as a way to more quickly diagnose everything from breast cancer to diabetic retinopathy — and has been found to be significantly more accurate and speedy than human experts, allowing doctors to treat patients before they deteriorate.
Estimated number of deaths per year from acute kidney injury
In a paper published in Nature on Wednesday, London-based DeepMind described its two-year analysis of more than 700,000 medical records from US military veterans across the country, creating a research model to predict severe progression of AKI, which could result in the patient requiring dialysis or a kidney transplant.
It used millions of metrics taken from 10 years of medical history for each patient to create effectively a new formula for diagnosing the illness. It was then able to predict the disease correctly in nine out of 10 patients up to two days ahead of time, almost doubling performance compared with standard AI models for the disease.
“The improvement over existing [machine learning] approaches is substantial,” said Chris Russell, a machine learning scientist at the University of Surrey.
“However, there is no comparison against current medical practice in the [Nature] paper . . . We would need to see an improvement over human clinicians in order to be confident that it would actually make a difference to people’s lives,” he added.
DeepMind, which is also testing machine learning to diagnose breast cancer and eye diseases in the UK, founded its health division in 2016 with the intention of developing and deploying AI technologies in real-world settings, including GP surgeries and hospitals, but so far it has no commercial AI products and the division has yet to generate any revenues.
In November, the company announced that it would transfer control of its health unit to a new Google Health division in California, an indication of its plans to expand and commercialise its efforts.
No money changed hands between the two organisations during the kidney injury project, and both are free to make use of any technology developed. The VA said that, together with Google, it was planning a clinical trial of the DeepMind algorithm at the VA hospital in Palo Alto, California.
“If we can predict AKI 48 hours in advance, we can treat for it . . . so our chances of preventing kidney injury, and therefore dialysis and even death, are much higher,” said Chris Nielson, a critical care physician at the VA, who worked on the DeepMind AI research.
“What’s really interesting is that . . . it’s applicable to a wide range of diseases from sepsis, to heart disease, liver failure, COPD,” he added. “Pretty much every disease we deal with, if we can identify it early, we have a much better chance of treating and preventing progression.”
The privacy problem
One of the inherent challenges of using AI in healthcare is that it requires large sets of extremely personal data, and poses a threat to patient privacy if it isn’t properly de-identified.
To protect the privacy of the veterans whose records were used to develop the algorithm, their sensitive medical data were de-identified by the VA before sharing with DeepMind. Details obscured included patient dates of birth, most of the measurement categories and dates on which they were taken, and even the exact value of each medical measurement, according to Nenad Tomasev, a DeepMind researcher and the lead author of the paper.
But privacy concerns may threaten DeepMind’s hopes of applying its algorithms to real-life patients in the UK. The company, which works with several National Health Service trusts on research and clinical projects, fielded controversy in 2017 regarding its use of 1.6m patient medical records from the NHS’s Royal Free Trust, which resulted in the UK’s Information Commissioner ruling it was shared illegally.
In particular, DeepMind hopes to apply this type of AI model to Streams, an app that helps hospital staff at NHS trusts monitor patients’ test results to spot AKI without the help of AI technology.
“No one has deployed this type of cutting-edge AI for direct patient care yet, so we are working through how that would look. We are having discussions with our Streams partners at the moment,” DeepMind Health’s Mr King said.
Get alerts on Artificial intelligence when a new story is published