Artificial Intelligence From ECGs May ID Asymptomatic LV Dysfunction
The system has the potential to identify people who might benefit from preventive medical or device therapy, researchers say.
Applying artificial intelligence (AI) technology to ECGs may be useful for detecting asymptomatic LV dysfunction, allowing for early intervention to head off overt heart failure, a new study suggests.
Such a system demonstrated “very strong performance” for distinguishing patients with a reduced ejection fraction (≤ 35%) from those with better ventricular function, senior author Paul Friedman, MD (Mayo Clinic, Rochester, MN), told TCTMD.
As reported in the January 2019 issue of Nature Medicine, the area under the receiver operator curve (AUC) was 0.93, which compares favorably with values for other commonly used screening tests like cervical cytology for cervical cancer (0.70) and mammography for breast cancer (0.85), Friedman said.
The AI system used here—called a convolutional neural network—was trained by feeding it raw data from 12-lead ECGs accompanied by information on heart function.
“Because the computer is seeing hundreds of thousands of examples, it’s making these connections without any of the assumptions or biases that a human would apply, and it can detect very fine patterns that are essentially hidden in plain sight,” Friedman explained. “The ECG pattern is there but humans just don’t recognize it. It’s too subtle, and the connection has not been made because no human has looked at that many ECGs to draw the connections.”
If it stands up to further testing, the system might be useful for identifying the 1.4% to 2.2% of the population that has asymptomatic LV dysfunction, which often goes undetected, according to Friedman. He pointed out that there are established therapies—such as ACE inhibitors and beta-blockers—that can boost ventricular function, prevent the onset of symptoms, and improve outcomes if the condition is discovered.
The best studied approach for identifying asymptomatic LV dysfunction is measurement of B-type natriuretic peptide (BNP) levels, “but studies on BNP have been disappointing, and the test requires invasive blood draws,” the investigators note in their paper.
“We hypothesized that if we used a very common, inexpensive, 10-second, noninvasive test—an ECG, which is ubiquitous, it’s everywhere—and applied artificial intelligence to it that the AI algorithm could identify whether a weak heart pump is likely present and then flag those individuals to go on to get the additional testing that would help them,” Friedman said.
For the study, Friedman, lead author Zachi Attia, MSc (Mayo Clinic), and colleagues used paired 12-lead ECG and transthoracic echocardiogram data on 44,959 patients treated at the Mayo Clinic to train the neural network to identify those with an LVEF of 35% or lower based solely on the ECGs. They then tested the system in 52,870 additional patients, of whom 7.8% had LV dysfunction.
The network identified patients with an EF of 35% or lower with 86.3% sensitivity, 85.7% specificity, and 85.7% accuracy. The negative predictive value was 98.7%. It performed well across categories defined by age and sex, but with significant variation in the strength of the associations.
Of note, Friedman said, among 1,335 patients with false-positive results (low EF according to the AI system but a normal echocardiogram), 147 developed LV dysfunction during follow-up, indicating a fourfold higher risk of future ventricular dysfunction compared with people with a negative AI screen.
“It’s like it’s looking into the future, but what I think is actually happening is that the very early manifestation of diseases causing some of those heart cells, those myocytes, to not generate the proper electrical signals [creates] subtle pattern changes showing up on ECG that the computer’s able to pick up,” Friedman said. “And then over time [the dysfunction] becomes more manifest and then the heart pump gets weak.”
The investigators continue to evaluate the AI algorithm and are currently conducting a pilot study at their center. Friedman said he sees the system being integrated into the electronic medical record so it can alert clinicians if anything irregular shows up on an ECG. Physicians would then be able to decide whether further testing or initiation or adjustment of preventive drug or device therapies is necessary.
“Whether this group would benefit from serial screening or medical therapy to prevent the development of ventricular dysfunction, however, is unknown,” the authors acknowledge.
Attia ZI, Kapa S, Lopez-Jimenez F, et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med. 2019;25:70-74.
- Mayo Clinic has licensed the underlying technology described here to EKO, which makes stethoscopes with embedded ECG electrodes, and may receive financial benefit from the use of this technology.
- Friedman and some of his co-authors may also receive financial benefit from this agreement.