A deep learning model trained on more than 1.5 million electrocardiograms and developed by a team at the Mayo Clinic improved detection rates for hyperkalemia in patients with chronic kidney disease (CKD), according to a study published April 3 in JAMA Cardiology.
Hyperkalemia—the presence of abnormally high potassium levels in a patient’s blood—is common in sufferers of CKD, senior author Paul A. Friedman, MD, of the Department of Cardiovascular Medicine at the Mayo Clinic, and colleagues wrote in JAMA. Serum potassium monitoring can reduce the risk of hyperkalemia by more than 70%, but guideline-directed monitoring is “severely underused” in clinical practice.
“When present, hyperkalemia is often asymptomatic and associated with cardiac arrhythmias and death,” the authors said. “We sought to improve detection of hyperkalemia in patients with CKD by developing and validating a noninvasive screening test using the electrocardiogram (ECG).”
Hyperkalemia has been linked to a host of ECG abnormalities, including peaked T waves, QRS prolongation and PR shortening, but physicians still struggle to diagnose hyperkalemia from ECGs, achieving a sensitivity of just 34% to 43%. To improve those rates, Friedman and co-authors trained a deep convolutional neural network (DNN) to detect serum potassium levels of 5.5 mmol/L or less, keeping in mind that anything above that threshold was considered hyperkalemia.
The researchers trained their model using 1,576,581 ECGs from 449,380 patients seen at the Mayo Clinic in Rochester, Minnesota, between 1994 and 2017. The DNN was trained using two and four ECG leads.
Validation of the model, which involved pulling retrospective data from Mayo Clinic facilities in Minnesota, Florida and Arizona, included 61,965 patients with stage 3 or worse CKD. All patients had a serum potassium count drawn within four hours of their initial ECG.
Friedman et al. found the prevalence of hyperkalemia in the three validation sets ranged from 2.6% in Minnesota to 4.8% in Florida. Using just two ECG leads, the area under the curve (AUC) for the deep learning model worked out to 0.883 for Minnesota, 0.860 for Florida and 0.853 for Arizona. Using a 90% sensitivity operating point, sensitivities were 90.2%, 91.3% and 88.9% for Minnesota, Florida and Arizona, respectively; specificities were 63.2%, 54.7% and 55%, respectively.
“The model was robust across diverse patients, geography and year,” the researchers wrote. “At a high-sensitivity operating point, the deep learning model performed well as a potential screening tool to rule out hyperkalemia, with a negative predictive value greater than 99%.”
The team wrote their model performance was better than that of other common screening tests, noting mammography has achieved an AUC of 0.78 and stool DNA testing for colorectal cancer has an AUC of 0.73.
“For patients with CKD with a clinical indication for serum potassium evaluation, such as with RAAS (renin-angiotensin-aldosterone system) inhibitor medical management, the application of artificial intelligence to the ECG may enable noninvasive screening for hyperkalemia,” they said.