Neural Network Provides Real-Time Prediction of Kidney-Injury Risk

By Will Boggs MD

August 06, 2019

NEW YORK (Reuters Health) - A neural-network system can provide continuous prediction of acute kidney injury (AKI) risk using electronic health record (EHR) data, researchers report.

"These findings highlight the potential of AI-enabled technologies and insights to support clinical practice," Dr. Joseph R. Ledsam from DeepMind, in London, told Reuters Health by email. "Combined with innovative ways to surface clinical insights, it could signal a shift in medicine from reactive care to preventive care - giving doctors back time to look after their patients and intervene before they deteriorate."

Current algorithms for detecting AKI rely on changes in serum creatinine and other markers that lag behind renal injury, resulting in delayed access to treatment.

Dr. Ledsam and colleagues developed a recurrent neural network that monitors individual electronic health records, processes the data and maintains an internal memory of relevant information, and outputs the probability of AKI at any stage of severity within the next 48 hours.

This approach predicted 55.8% of inpatient AKI events of any severity within a window of up to 48 hours in advance, with a ratio of two false-positives for every true-positive, the researchers report on July 31 in Nature.

Of the false-positive alerts made by the model, 24.9% were positive predictions made even earlier than the 48-hour window in patients who subsequently developed AKI, and 24.1% were trailing predictions that occurred after an AKI episode appeared to have resolved.

The model correctly predicted 84.3% of episodes in which administration of in-hospital or outpatient dialysis was required within 30 days of the onset of AKI of any stage and 90.2% of cases in which regular outpatient administration of dialysis was scheduled within 90 days of the onset of AKI.

The predictive performance of the model was maintained across time and hospital sites. But the researchers note that the patients on whom the model was trained (from the U.S. Department of Veterans Affairs healthcare system) are not representative of the general population. For example, only 6.38% of patients in the data set were female, and the model performed worse for this demographic.

"It's important to state that, whilst these are really exciting results, this is still early-stage research," Dr. Ledsam said. "The model would need to go through a series of robust and rigorous testing before being applied to other situations or institutions, so we're not in a position to comment on efforts or costs at the moment."

Dr. Michael Simonov from Yale School of Medicine, in New Haven, Connecticut, who recently reported a simple real-time model for predicting AKI in hospitalized patients, told Reuters Health by email, "The feasibility of this model in different settings remains questionable. As the authors note, this model was developed entirely from Veteran's Health Administration data with only 6% female representation and a predominantly white population. Most non-VA settings have a different demographic profile and model performance must be validated in these other settings."

"Implementing models onto the electronic health record for real-time prediction is a timely process which requires careful mapping of variables and extraction of data," he said. "The method of implementation of this complex model with hundreds of thousands of variables on non-VA electronic health record systems remains unobvious and may be a major hindrance to prospective studies of this model."

"This study should be commended in its focus on true and false alerts and the model's use within the clinical workflow," Dr. Simonov added. "Given the growing awareness of alert fatigue and provider burnout, discussing the implementation of the model in a provider-centric way is critical in such models. The authors propose a strategy for clinically meaningful alerts 48 hours prior to AKI with a fair, but not unreasonable, provider burden of two false alerts for every true alert."

Dr. Matthew T. James from Cumming School of Medicine at the University of Calgary, in Canada, who studies AKI, said, "Point-of-care electronic medical record systems are becoming increasingly common and sophisticated, making this approach ever more feasible in our hospitals. This work demonstrates how these systems can move beyond record keeping and be leveraged for meaningful clinical use."

"Like all prognostic tools, these algorithms require validation studies to confirm their accuracy when used in independent groups of patients, as well as clinical impact testing to see if reporting the predictions leads to earlier treatment, lower rates of kidney injury, and better patient outcomes," Dr. James, who also was not involved in the study, told Reuters Health by email.


Nature 2019.