“Deep Learning” Approach Allows Early Prediction of AKI

Full access

An approach using artificial intelligence enables continuous prediction of acute kidney injury (AKI) in hospitalized patients, according to a research letter in Nature.

The authors developed a deep learning approach for continuous risk prediction of future AKI, based on individual electronic health records. The model was developed by use of retrospective data on more than 700,000 adult patients from the US Department of Veterans Affairs, including 172 inpatient and 1062 outpatient sites. The researchers write, “At every point throughout an admission, the model provides updated estimates of future AKI risk along with an associated degree of uncertainty.” The

An approach using artificial intelligence enables continuous prediction of acute kidney injury (AKI) in hospitalized patients, according to a research letter in Nature.

The authors developed a deep learning approach for continuous risk prediction of future AKI, based on individual electronic health records. The model was developed by use of retrospective data on more than 700,000 adult patients from the US Department of Veterans Affairs, including 172 inpatient and 1062 outpatient sites. The researchers write, “At every point throughout an admission, the model provides updated estimates of future AKI risk along with an associated degree of uncertainty.” The model can output the probability of AKI of any severity within 48 hours, with the possibility of other time windows or severities.

Within the Veterans Affairs dataset, AKI as defined by Kidney Disease: Improving Global Outcomes occurred in 13.4% of hospitalizations. The model correctly predicted 55.8% of AKI episodes. Sensitivity was highest in patients who experienced lasting complications of AKI: the model correctly predicted 84.2% of episodes requiring dialysis in the hospital or within 30 days, and 90.2% of those requiring regular outpatient dialysis within 90 days. For every true alert, there were two false alerts. The model also listed the most relevant clinical features for each prediction, along with predicted future trajectories for important laboratory test results.

Building on work on modeling adverse events from electronic health records, the new study suggests that a machine learning approach can enable prediction of AKI risk within a clinically actionable window. The researchers write, “[O]ur approach may allow for the delivery of potentially preventative treatment—before the physiological insult itself, in a large number of cases” [Tomašev N, et al. A clinically applicable approach to continuous prediction of future acute kidney injury. Nature 2019; 572:116–119].

Save