An algorithm developed by DeepMind Health, the healthtech unit of Google-parent Alphabet’s DeepMind AI research company, has accurately predicted impending kidney failure with an accuracy of 90.2%. The AI is still in a relatively early development phase and the algorithm needs to be further refined, and the patient data available to it more complete, before it becomes viable in a clinical setting. But the researchers behind the project believe it could be a first step towards preventing a condition that contributes to the loss of hundreds of thousands of lives a year in the USA alone.

Those who survive AKI, acute kidney injury, often face the prospect of months to years of dialysis treatment as part of their recovery process. Dialysis is a time consuming procedure to have to go through regularly and is also extremely expensive. In the USA that cost is either picked up by the health insurer, if the patient is lucky enough to be covered. In the UK, it’s a cost burden that usually falls upon the stretched budget of the NHS. As such, the prospect of AI being able to accurately detect the patients most at risk, allowing preventative or alleviative steps to be taken, could potentially save a fortune as well as the obvious health benefit to patients.

AKI often strikes patients already in hospital. While it can be caused by a wide range of different factors, it often occurs as a result of complications during surgery, as a result of sepsis or side effects or immune system reactions to medicines. Other triggers of AKI that often occur while a patient is being treated in a hospital include urinary-tract blockages, burn complications and heart attacks.

AKI is such a problem because until now the main way of detecting it has been a spike in the creatinine levels in the blood. However, higher concentrations of creatinine occur after the fact, meaning medics are reacting to rather than preventing AKI.

The DeepMind Health research was recently published in the scientific journal Nature. The algorithm was developed, or trained, on the basis of 703,782 anonymised health records obtained from the U.S. Department of Veteran Affairs. These electronic medical records are particularly useful as they are far richer than standard patient records. 600,000 different possible signals including blood-test results, vital signs, prescriptions and past procedures, ward transfers and admissions to an intensive care unit had to first be narrowed down to those the researchers found could be connected to the risk of AKI developing. Of the original 600,000 data points, it was decided 4000 could influence impending kidney failure.

Across all cases of AKI, the algorithm showed 55.8% accuracy in predicting that it would develop within 48 hours. In cases where the results were severe enough to subsequently require dialysis treatment, accuracy increased to 90.2%.

Being tipped off up to 48 hours in advance means medics can take preventative action such as providing more diuretics, intravenous fluids or adjusting medication that could be toxic to the patient’s kidneys. If the KPI cannot be prevented entirely, medics can quickly move to treat the patient and alleviate potential repercussions, reducing the level of ongoing treatment subsequently required.

Further research needs to be conducted with an even broader set of medical records, including more of women which were a small percentage of the original database 703,782. For future use in a general clinical setting, the fact that most patient data files are less complete than that of U.S. veterans must also be overcome. The general trend towards greater digitalistion of health records should eventually naturally overcome that issue. But in the meanwhile it is hoped the accuracy of the algorithm can still be improved for cases where patient files are less complete.

DeepMind Health’s product lead Dominic Knight also hopes to develop predictive algorithms able to give advance warning of other common medical conditions in danger of developing:

“Sepsis, liver failure, diabetes complications. We see huge potential that [this algorithm] could be applied to other preventable conditions”.