Conference Coverage

AI algorithm finds diagnostic AFib signatures in normal ECGs



– Researchers have created an artificial intelligence algorithm that can evaluate a 10-second ECG recording of a person in normal sinus rhythm and tell with a sensitivity and specificity of almost 80% whether or not that person ever had atrial fibrillation episodes some time in the past or will have a first arrhythmia episode in the near future.

Dr. Paul A. Friedman, professor of medicine and chair of the department of cardiovascular medicine, Mayo Clinic, Rochester, Minn. Mitchel L. Zoler/MDedge News

Dr. Paul A. Friedman

Although this algorithm – derived from and then validated with a dataset of nearly 650,000 ECG recordings from more than 180,000 patients – still needs prospective validation, it offers the prospect for a potential revolution in screening for atrial fibrillation (AFib), Paul A. Friedman, MD, cautioned at the annual International AF Symposium. If initial clinical findings are confirmed, it would show that a 10-second, 12-lead ECG recording can provide the same screening scope as what otherwise takes weeks of ambulatory ECG recording with a Holter monitor or an implanted device, explained Dr. Friedman, professor of medicine and chair of the department of cardiovascular medicine at the Mayo Clinic in Rochester, Minn.

This finding “could have important implications for atrial fibrillation screening and for the management of patients with unexplained stroke,” Dr. Friedman and his associates noted in the published report of their study (Lancet. 2019 Sep 7;394[10201]:861-7). “We’re still working to define the window of ECG” recording time that provides the optimal assessment for a history of asymptomatic AFib, but the “possibilities this opens are huge,” Dr. Friedman said in his talk at the symposium. This work sprang from the premise that “subtle signatures” in a brief, apparently normal sinus rhythm ECG tracing can harbor reliable clues about AFib history or an imminent episode.

The 2019 report by Dr. Friedman and associates documented that in the validation phase of their study, the trained artificial intelligence (AI) program identified patients with a history of AFib or an impending arrhythmia event from a single, 10-second ECG that to the naked eye seemed to show normal sinus rhythm with a sensitivity of 79.0%, a specificity of 79.5%, and an accuracy of 79.4%. It also showed an area under a receiver operating characteristic curve of 0.87, meaning that screening for AFib by this method compared favorably with the area-under-the-curve (AUC) results tallied by several widely accepted screening tools, including Pap smears for cervical cancer (AUC of 0.70), mammograms for breast cancer (AUC of 0.85), and CHA2DS2-VASc scoring for estimating stroke risk in AFib patients (AUC of 0.57-0.72), Dr. Friedman said.

The researchers developed the AI algorithm with more than 450,000 10-second ECG tracings collected from roughly 126,000 patients who underwent at least one ECG recording as part of their routine care at the Mayo Clinic during 1993-2017. The goal was for the program to find and validate recurring characteristics in the ECG that consistently linked with a history of or an impending AFib episode and that did not appear in ECG recordings from people without any AFib history. The program this effort produced then underwent further adjustment with the use of more than 64,340 ECGs from an additional 18,116 patients, and then the final product underwent validation testing with a further 130,802 ECGs collected from an additional 36,280 people, the study phase that resulted in the reported sensitivity and specificity estimates.

It’s currently unclear to Dr. Friedman and associates what specific features the program uses to classify patients. It’s an important question, but if the results are reproducible and reliable, this uncertainty shouldn’t slow clinical adoption, he said in an interview.

While “this particular algorithm needs prospective vetting,” a similar algorithm developed by Dr. Friedman and the same research team that uses a 10-second ECG to identify patients with a left ventricular ejection fraction of 35% or less is further advanced in development, and a device that uses this algorithm will soon receive Food and Drug Administration review under a fast track designation that the agency approved in late 2019.

The researchers developed this algorithm for estimating left ventricular function using a strategy similar to their development of a tool for diagnosing AFib (Nat Med. 2019 Jan 7;25[1]:70-4), and results from 100 patients prospectively studied with this approach to ECG analysis and reported at the American Heart Association scientific sessions in November 2019 showed that the algorithm identified substantial left ventricular dysfunction with an AUC of 0.906 (Circulation. 2019 Nov 19;140[suppl 1]:A13447). The same team of investigators has developed an AI algorithm that can calculate a person’s physiologic age based on the ECG recording (Circ Arrhythm Electrophysiol. 2019 Sep;12[9]: 10.1161/CIRCEP.119.007284).

The study received no commercial funding, and Dr. Friedman and coauthors had no relevant disclosures. The Mayo Clinic has licensed a related artificial intelligence algorithm to EKO, and Dr. Friedman may benefit financially from this arrangement.

Recommended Reading

Infective endocarditis isn’t what it used to be
The Hospitalist
Single-center study outlines stroke risk, DOAC type in nonvalvular AFib patients
The Hospitalist
AFib screening cuts hospitalizations and ED visits
The Hospitalist
AFib on the rise in end-stage COPD patients hospitalized for exacerbations
The Hospitalist
DOACs surpass warfarin in low-weight AFib patients
The Hospitalist
NOACs benefit early stage chronic kidney disease patients
The Hospitalist
AI technology meets AFib detection
The Hospitalist
Post-TAVR anticoagulation alone fails to cut stroke risk in AFib
The Hospitalist
Dual therapy best for AFib with ACS no matter the treatment strategy
The Hospitalist
Genetic test stratified AFib patients with low CHA2DS2-VASc scores
The Hospitalist
   Comments ()