An artificial intelligence (AI) system is shown to predict acute kidney injury up to 48 hours before it occurs. The approach, described in a paper published in this week’s Nature, could help to identify patients who are at risk of health deterioration within a time window that may enable earlier treatment.
An estimated 11% of deaths in hospital are attributed to the failure to quickly identify and treat deteriorating patients. To address this issue, Joseph Ledsam and colleagues have developed a deep learning method for assessing patient risk factors. They demonstrate the applicability of this approach for the prediction of acute kidney injury, a potentially life-threatening condition that affects approximately one in five hospital inpatients in the United States.
The system was trained using data from more than 700,000 patients who were treated by the United States Veteran Affairs medical system. 55.8% of acute kidney injury episodes were correctly predicted up to 48 hours before this would be apparent under standard clinical monitoring methods. The system successfully identified up to 90.2% of people with serious forms of the condition that required dialysis treatment. This early warning could enable treatment to be offered before irreversible kidney damage occurs.
The authors note some limitations to their study. For example, for every positive result predicted, two false-positive alarms were raised; however, many of these false-positive cases arose in patients who were already experiencing chronic kidney injury. In addition, only 6.38% of the patients whose data were used to train the AI were female, so it is unclear how applicable this particular approach might be to a broader population. Nevertheless, these results open up the possibilities that AI systems could predict and potentially prevent some adverse events in hospital patients.
Materials: Storing energy in bricksNature Communications
Planetary science: Dawn’s close-up look at CeresNature Astronomy
Engineering: Reducing noise transmitted through an open windowScientific Reports