番茄社区app

Applying artificial intelligence to the ECG may make it possible to screen patients with chronic kidney disease (CKD) for hyperkalemia without making them get a blood test, a study showed.

Researchers trained a deep learning model to predict serum potassium levels of 5.5 mEq/L or more (the usual cutoff for hyperkalemia treatment) and validated it against three regional datasets of patients with stage 3 or greater CKD (n=61,965).

The model appeared to show good results with input from ECG leads I and II (AUC 0.883 for Minnesota; 0.860 for Florida; and 0.853 for Arizona); the addition of V3 and V5 only slightly improved the performance of the model, according to Paul Friedman, MD, of the Mayo Clinic in Rochester, Minnesota, and colleagues, in their study published online in .

A balance of sensitivity and specificity could be achieved with both being in the 75-82% range. But when the model was operating at ~90% sensitivity, its specificity was limited to 63.2% for Minnesota, 54.7% for Florida, and 55.0% for Arizona.

"The ability to noninvasively screen for hyperkalemia using ECG data would represent a major advance in patient care of this life-threatening condition," the authors maintained, noting that they found a negative predictive value of 99.0% to 99.6% for the deep learning model.

The study appeared to be the foundational research by AliveCor in developing its hyperkalemia-detecting for two-lead ECGs from its Kardia wearables.

Friedman’s study didn’t include any data on out-of-hospital ECGs from wearables, although the same Mayo Clinic group has preliminarily reported at a conference that the artificial intelligence platform had 90% to 94% accuracy against that hospital-based data.

The model had been trained from more than 1.5 million ECGs from CKD patients who had at least one serum potassium test within 12 hours before or after the ECG for reference. The investigators excluded ECGs with left bundle branch block.

"Between 50% and 70% of patients in the validation data sets did not have hyperkalemia predicted by the DLM [deep learning model], with less than 1% of all test results being false-negative; on the other hand, up to 42% of all test results were false-positive," Friedman's group noted.

Ultimately, their technique for detecting hyperkalemia is "practically useless in the clinic" as "most doctors would consider almost unacceptable" that 90% sensitivity for such a test, commented Joel Topf, MD, of St. Claire Nephrology in Detroit.

Using the Minnesota dataset as an example, he calculated that the positive predictive value comes out to just 6.1% given the prevalence of hyperkalemia being 2.6% and a specificity of 63%, he told 番茄社区app. "That means 94% of those positive EKGs would actually have a normal potassium. No one is going to care about a screening test that is a false alarm 19 out of 20 times."

"Certainly the performance of the test would improve as the prevalence of hyperkalemia climbs and maybe in a heart failure clinic, where the potassium is increased by commonly used drugs like ACEi [ACE inhibitors] and aldosterone antagonists," Topf continued. "But even if the rate of hyperkalemia was four times as high, the positive predictive value is only 10%, still useless."

Friedman's group argued that false negatives "may not necessarily be false" because the blood test used as reference is less "germane to health and arrhythmia risk" than ECG-derived potassium.

"Another possibility is that there may have been errors in the blood tests in these patients," they suggested. "The ECG-based tests are not susceptible to mechanical, temperature, contamination, or the other potential errors associated with processing blood."

Nevertheless, the investigators acknowledged that additional refinement of the deep learning model is required to bring down the rate of false positives.

The diagnosis of hyperkalemia remains a challenge for now since "patients are often asymptomatic, and guideline-directed blood potassium monitoring is severely underperformed," according to the authors.

"The more interesting angle is asking why they are focused on an intermediate outcome," Topf said. "Patients and cardiologists are not interested in potassium specifically but what the hyperkalemia leads to, which is lethal cardia arrhythmia. Perhaps, the investigators' technique machine could be adapted to predict that outcome."

Comment