Two studies have revealed how artificial intelligence can accurately predict kidney complications in hospital patients, helping clinicians to instigate timely treatment or preventive care.
The findings were presented at Kidney Week 2022, the meeting of the American Society of Nephrology currently being held in Orlando, Florida.
In the first study, researchers found that a machine learning risk score combining demographics, inpatient vital signs and laboratory results could predict the risk of future major adverse kidney events among 50,488 patients discharged alive after hospitalization.
“Our work needs to be validated with outside data but it could be used to help prioritize follow-up with nephrology and primary care as well as to determine which patients should (and should not) be sent for transplant or dialysis access evaluation,” commented researcher Jay Koyner, from the University of Chicago Pritzker School of Medicine.
“Similarly, combining our risk score with existing literature that shows acute kidney injury increases the risk of new congestive heart failure, we could potentially determine which patients should be seen by cardiologists.”
All study participants had been treated at the University of Chicago between November 2008 and June 2020, and none had severe, chronic kidney disease at hospital admission.
Nonetheless, 19.7% of patients developed a major adverse kidney event within 90 days of discharge, which was defined as acute kidney injury, chronic kidney disease, need for dialysis, or kidney-related death.
The team developed the machine learning, gradient boosted algorithm in 70 per cent of the admissions and then applied it to the remaining 30 percent of test data.
The model was able to discriminate those test patients that developed post-hospitalization major adverse kidney events, with an area under the receiver operating characteristic curve (AUC) of 0.74.
Among the individual endpoints, it was best at identifying those patients that developed chronic kidney disease after hospitalization, with an AUC of 0.94 at 90 days and 0.92 at 1 year.
In the second study, researchers used machine learning to develop a real-time prediction model for acute kidney injury among patients being treated at intensive care units.
The model was developed from all 16,785 adult admissions to the intensive care unit at Taichung Veterans General Hospital in Taiwan from 2015 to 2000. Participants had a median age of 68 years, 62 per cent were male and the incidence of acute kidney injury was 30.8 per cent.
A total of 60 predictors were included in the model and testing it on four other hospitals revealed AUCs ranging from 0.785 to 0.864. Federated learning using neuron network algorithm improved its prediction performance.
“Early prediction of AKI ahead of 24 hours may help clinicians initiate timely interventions to prevent AKI from happening or alleviate its severity,” said researcher Chun-Te Huang, from the Taichung Veterans General Hospital.
“Our model could be easily shared and integrated to different hospitals to provide a real-time risk prediction in electronic health information systems.”
