1. Froelich, Warren

Article Content

Researchers from the United Kingdom are using a machine learning model to help identify patients at high risk for acute kidney injury (AKI) up to 30 days before it may occur, according to findings presented during the AACR Virtual Special Conference: Artificial Intelligence, Diagnosis, and Imaging.

kidney; renal. kidne... - Click to enlarge in new windowkidney; renal. kidney; renal

Today, most cases of acute kidney injury (AKI) aren't detected before toxins appear in the patient's blood, usually in the form of increased levels of creatinine, produced by muscles and filtered by the kidneys into urine. Unfortunately, by the time creatinine rises, the damage to kidneys has already been done, creating life-threatening conditions.


So, the ability to predict AKI days or weeks before it occurs not only could reduce lengthy hospital stays and medical costs, it would also save thousands of lives.


"AKI causes huge disruption for treatment and distress for the patient," said Lauren Scanlon, PhD, a data scientist with The Christie NHS Foundation Trust, who presented results during the AACR meeting. "So, it would be amazing if we could predict AKI before it occurs and prevent it from ever happening."


Assessing Acute Kidney Injury

AKI, also known as acute renal failure, is a sudden onset of kidney failure or kidney damage over a few hours or days, resulting in the build-up of dangerous waste products in the blood. Patients that experience AKI are at increased risk for adverse health outcomes such as end-stage renal disease, pulmonary complications, cardiovascular events, and death. Critically ill patients, including those undergoing chemotherapy for cancer, are particularly vulnerable to AKI.


According to Scanlon, the estimated number of deaths associated with AKI in England is above 40,000. Various reports indicate that AKI occurs anywhere from 5 to 20 percent of hospitalizations in the United States, with the prevalence of AKI in cancer patients about 7.5-9.6 percent.


"It presents a large problem to the health service in the UK," Scanlon said, "increasing inpatient stay length, inpatient care costs, and increasing mortality."


When AKI is detected, physicians employ a range of measures called the AKI care bundle, that includes monitoring of fluid balance, drug dosing, and reviewing medications. Given that AKI is detected from blood tests, the U.K. researchers hoped to use patient's own routinely collected blood to create an ongoing risk prediction model that would warn of a potential case of AKI up to 30 days before it occurs. Once alerted, physicians could put the AKI care bundle into practice.


"Using routine collected blood test results will ensure that it's applicable to all our patients and can be implemented in an automated manner," Scanlon said.


Machine Learning Markers

Between January 2017 and May 2020, the research team collected data from 597,403 blood tests from 48,865 cancer patients treated at The Christie NHS Foundation Trust. Each sample contained a marker capable of identifying future AKI events.


Using this data, the team trained a machine learning algorithm called a Random Forest Model to predict, from the blood sample marker, if a patient would develop AKI within 30 days. Output from this model were assigned a potential AKI risk factor: very low, low, medium, high and very high.


The trained model gave an area under the receiver operating characteristic curve (ROC AUC) of 88.1 percent (95% CI 87.8-88.3%) when assessing predictions per blood test for AKI occurrences within 30 days. The area under a ROC curve is a measure of the accuracy of a quantitative test. A score of 50 percent is the same as a coin flip and is considered useless. A score of 100 percent is perfect.


Satisfied with their model's training performance, the team sought to validate its potential through a prospective analysis of blood tests collected from 9,913 patients at their hospital over a 3-month period of time, between June 1, 2020 and August 31, 2020. The model, with a risk alert of medium or higher, predicted that 484 patients would come down with AKI during this time. Results showed that 330 were correctly predicted, while 117 were not predicted, yielding a success rate of 73.8 percent of patients detected with an AKI within 30 days.


"This is pretty amazing and shows our model really is working and can correctly detect AKI up to 30 days before they occur, giving a huge window to put in preventative strategies," Scanlon noted.


For the 154 patients incorrectly identified as having a medium or higher risk by the model, some 17 did not have a follow-up blood test and nine only had one blood test in the 30 days following the risk alert. This means that after the AKI prediction was made, no other blood test was taken.


"These patients could have been suffering from an undiagnosed AKI, so that potential unintended benefit of our prediction model would be to identify patients at risk and arrange for them to have a follow-up blood test if one was not scheduled, allowing for intervention or treatment of potentially undiagnosed AKI," Scanlon said in an interview.


As for next steps, Scanlon said the team plans to test their model through a clinical trial.


"We've done our testing and we can see our model is predicting and doing exactly what we intended," she said. "But we now need to go through a clinical trial to see if putting intervention strategies in place does prevent these AKI from taking place."


Warren Froelich is a contributing writer.