Editor’s note: Find the latest COVID-19 news and guidance in Medscape’s Coronavirus Resource Center.
There is something special about the electrocardiograms (ECG) of people hospitalized with COVID-19 that ECG-processing algorithms, developed using artificial intelligence (AI), can “recognize” well enough to discern when its absent, suggests a proof-of-concept study.
TThe research involving about 48,000 tracings from patients in 12 countries spanning the world points to one pathway toward development of ECG systems that can potentially be used almost anywhere to screen for SARS-CoV-2 infection, proposes a report published June 15 in Mayo Clinic Proceedings.
Trained on about 26,000 ECGs from patients with and without COVID-19, a deep neural network (DNN) achieved a negative predictive value (NPV) for the illness in excess of 99% in a sample cohort with a 5% prevalence of patients hospitalized with the infection.
The NPV exceeded 90% even in a cohort with a SARS-CoV-2 prevalence of 35%, suggesting the tool might reliably rule out the infection under real-world conditions even during a pandemic, conclude the report’s authors, led by Zachi I. Attia, PhD, and Suraj Kapa, MD, Mayo Clinic College of Medicine, Rochester, Minnesota.
Overwhelmingly, most of the sample COVID-related 12-lead ECGs on which the DNN was trained and tested were from hospitalized patients, and the remainder were at least symptomatic, observed Kapa for theheart.org | Medscape Cardiology.
“We can only make conclusions to that extent,” he said. The high NPVs in the study would apply — hypothetically in practice — only to sicker people who present to the hospital.
“To extend it to asymptomatic individuals would require further study,” Kapa said. That’s what would be needed to transfer the approach from 12-lead ECGs to handheld devices for screening in the community. “That’s something we’re looking at,” he said, in a large multicenter cohort study of whether the DNN expressed through a mobile or wearable ECG system “can have a similar accuracy to what we got with the 12-lead ECG.”
The DNN learned to recognize ECGs from patients with COVID-19 by isolating subclinical ECG features and any patterns of their distribution within the cohort that might signal or rule out the infection, a process Kapa called a “learned synthesis” of tracing characteristics and their association with the infection.
The patterns uncovered by the DNN might not involve conventional ECG features, but can be based on mathematical relationships that can be discerned by the algorithms but are unknowable from a standard reading by a clinician, Kapa explained.
The study used ECGs obtained from patients at 28 centers in 12 countries at about the time of their confirmed COVID-19 diagnosis and from age- and sex-matched SARS-CoV-2-free control patients at the same sites, the report notes.
A DNN was trained on about 26,000 ECGs to recognize those from patients with COVID-19 and was validated on about 7,900 others obtained from a cohort with a 32.7% prevalence SARS-CoV-2 cases to which it had not been previously exposed.
The area under the curve (AUC) was 0.767 (95% CI, 0.756 – 0.778) for discernment of acute COVID-19 infection, with a sensitivity of 98%, specificity of 10%, positive predictive value (PPV) of 37%, and NPV of 91%.
A second validation process was carried out after 51,000 ECGs obtained prior to 2019 were added to the mix, bringing the COVID-19 prevalence to about 5% and making it more reflective of “a real-world population,” the report says. That led to an AUC of 0.780 (95% CI, 0.771 – 0.790), with a specificity of 12.1% and NPV of 99.2%.
Kapa said and he and his Mayo colleagues working on AI solutions focus on proof of concept for their products, “and if something works, it would be made freely available to anybody and everybody.” That means a successful AI-ECG innovation could conceivably be used broadly in products from mobile ECG companies like AliveCor, wearable makers such as Apple or Fitbit, or hospital ECG-monitor manufacturers, he said.
“The study was designed and conceived by Mayo Clinic investigators with no financial support from industry,” notes the report, which acknowledges “the generous contribution of data, time, human resources, and intellectual capital from medical centers from around the world (authors and Discover Consortium) invited to participate.” In addition, “General Electric (Marquette, WI), SHL (Tel Aviv, Israel), Philips (Amsterdam, Netherlands), and Epiphany Healthcare (Midlothian, VA) donated resources, expertise, and in some cases, equipment to aggregate electrocardiographic data into a central research server for analysis.”
Mayo Clin Proc. Published online June 15, 2021. Full text
Need Your Help Today. Your $1 can change life.