Artificial Intelligence Detects 'Hidden' AFib During Normal Sinus Rhythm

Batya Swift Yasgur MA, LSW

August 07, 2019

An artificial intelligence (AI) model may identify patients with intermittent atrial fibrillation (AF), even when performed during normal sinus rhythm, in a little as 10 seconds, a new study suggests.

Investigators analyzed data from almost 650,000 sinus rhythm electrocardiograms (ECGs) in more than 180,000 adults between December 1993 and July 2017, dividing the ECGs into three groups: training (70%), internal validation (10%), and testing (20%).

The testing group used an AI-enabled ECG designed to recognize subtle changes using specially trained neural network technology.

The primary outcome was whether the AI-programmed ECG could accurately detect AF in the testing dataset in patients with confirmed AF prior to being tested with the AI device.

AI accurately identified the presence of AF, with 79% accuracy for a single scan and 83% accuracy when analyzing multiple ECGs for the same patient.

"We found that the AI ECG [was] powerful in its ability to detect recent or impending AF with an AUC of 0.90," Paul Friedman, MD, professor of medicine, Mayo Clinic, Rochester, Minnesota, told | Medscape Cardiology.

"In the future, this might facilitate point-of-care diagnosis by allowing application of the algorithm on low-cost, widely available technologies. For instance, we have previously shown the translation of neural networks created using 12-lead ECGs to mobile, smartphone-based electrodes that typically use a single lead," said Friedman, who is also Normal Blane and Billie Jean Harty Chair, Mayo Clinic Department of Cardiovascular Medicine Honoring Robert L. Frye, MD.

The study was published online August 1 in the Lancet.

Subtle Patterns

"Screening for AF can be challenging, due to the low diagnostic yield of a single ECG to detect an often fleeting arrhythmia," the authors write.

AF is "difficult to detect and often goes undiagnosed," Friedman observed, with detection "requiring monitoring for weeks to years, sometimes with an implanted device, potentially leaving patients at risk of recurrent stroke, as current methods do not always accurately detect AF or take too long."

"An AI-enabled ECG uses machine learning in the form of deep neural networks to identify subtle patterns in the standard 12-lead ECG to identify disease that is silent or impending," he explained.

The previous experience of the AI-ECG in detecting silent disease "motivated us to create a new network that would focus on AF, given its important association with heart failure, stroke, and death, and the significant role its presence plays in determining treatment," he added.

The researchers analyzed data from all adult patients (≥18 years) at the Mayo Clinic ECG laboratory between December 31, 1993 and July 21, 2017.

To be considered positive for AF, patients were required to have at least one digital, sinus rhythm, standard 10-second, 12-lead ECG with atrial fibrillation. To be considered negative for AF, patients were required to have no AF recorded on ECG and, additionally, no reference to AF in the diagnostic codes in their electronic medical record.

Any ECG with a rhythm of AF or atrial flutter was classified as AF.

For patients with multiple ECGs, the researchers defined a "window of interest" for the purpose of analysis.

For those who had had at least one AF rhythm recorded, the researchers defined the first recorded AF ECG as the index ECG and the first date of the window of interest as 31 days before that date.

For patients with multiple ECGs but no AF recorded, the index ECG was considered the first ECG available.

The primary outcome was the ability of the AI-enhanced ECG to identify patients with AF, assessed by area under the curve (AUC) of the receiver operating characteristic (ROC) curve, as well as sensitivity, specificity, accuracy, and F1 score.

The researchers also performed a secondary analysis to determine whether use of additional sinus rhythm ECGs per patient might improve the AUC of the AI-enabled ECG in detecting a history of AF.

Another secondary analysis was performed that included only the first normal sinus rhythm after the index AF ECG.

Of the 1 million ECGs available, the researchers analyzed 649,931 normal sinus rhythm ECGs (180,922 patients; mean [SD] age, 60.3 [16.5] years at first visit; 49.6% male; 8.5% having ≥1 recorded AF).

The model was trained using 454,789 ECGs recorded from 126,526 patients, (mean, 3.6 [4.8] ECGs per patient).

The researchers divided ECGs into two additional categories:

  • Internal validation dataset (n = 64,340 ECGs from 18,116 patients)

  • Testing dataset (n = 130,802 ECGs from 36,280 patients)

Of the testing dataset, 8.4% of patients had verified AF before the normal sinus rhythm ECG tested by the model.

Hidden in Plain Sight

In patients in the testing dataset with at least one AF recorded, 55.7% of the 3051 first normal sinus rhythm ECGs in the window of interest took place within 1 week of the index AF ECG.

In the internal validation dataset, as well as the testing dataset, a single AI-enabled ECG had the same level of success in identifying AF (AUC of 0.87; 95% CI, 0.86 - 0.88).

The probability value, which was found to have similar sensitivity, specificity, and accuracy (79.2%) on the internal validation set, was then applied to the testing set, yielding an F1 score of 39.2% (38.1 - 40.4), with a sensitivity of 79.0% (77.5 - 80.4), specificity of 79.5% (79.0 - 79.9), and an overall accuracy of 79.4% (79.0 - 79.9).

The researchers tested the model on all sinus rhythm ECGs in the first 31 days from the study start date, selecting the average and maximum probability of AF scores.

The AUC improved to 0.89 (0.89 - 0.90) when the average score on the test dataset was used, and to 0.90 (0.90 - 0.91) when the researchers applied a more sensitive approach of using the score of the ECG with the highest risk.

Similar improvements were seen with the same analysis on the internal validation set.

The researchers also conducted a secondary analysis on the testing dataset, including only the first normal sinus rhythm after the onset of AF, and found that the AUC of the network improved to 0.90 (0.89 - 0.91).

As in the primary analysis, the probability threshold yielded a similar sensitivity and specificity on the internal validation set and was used to classify the patients in the testing dataset.

When the researchers used the maximum score with the calculated threshold, they found improvements in all measures (the F1 score, 45.4% [44.2 - 46.5]; sensitivity, 82.3% [80.9 - 83.6]; specificity, 83.4% [83.0 - 83.8]; and overall accuracy, 83.3% [83.0 - 83.7]) on the testing dataset.

"The addition of AI to a standard, widely available, inexpensive, noninvasive test — the electrocardiogram — transforms it into a powerful [tool] that can detect recent or impending AF, which may be useful to guide treatment to prevent strokes, heart failure, and death," Friedman commented.

"It does this due to the ability of the network to see subtle patterns in the network that are hidden in plain sight," he added.

Needle in a Haystack

Commenting on the study for | Medscape Cardiology, Jeroen Hendriks, MSc, PhD, Center for Heart Rhythm Disorders, University of Adelaide and Royal Adelaide Hospital, Australia, who coauthored an accompanying editorial, called the research "robust."

"Rather than trying to observe atrial fibrillation by prolonged monitoring of sinus rhythm, the authors suggest that AI can avoid this needle-in-a-haystack scenario and instead identify from as few as one normal sinus rhythm ECG if there is indeed a needle hidden within."

In the future, "AI may support the treatment team in clinical decision-making" and may also be useful, for example, in patients with ESUS [embolic stroke with undetermined source] to identify if AF is the underlying cause of the stroke," Hendriks suggested.

"This particular model is specific to AF; however, we have created models for a number of other conditions that use the same fundamental principles, so that a large number of conditions can be detected using the AI-ECG," Friedman added.

The study and authors received no external or commercial support; funding was via internal Mayo Clinic resources. The authors declare no relevant financial relationships. Hendriks reports that the University of Adelaide has received on his behalf lecture or consulting fees from Medtronic and Pfizer/Bristol-Myers Squibb. He is also supported by a fellowship from the National Heart Foundation of Australia . The Institute of Cardiovascular Research has received an award from the British Heart Foundation. His coauthor's disclosures are listed on the original editorial.

Lancet. Published online August 1, 2019. Abstract, Editorial

For more Medscape Cardiology news, join us on Facebook and Twitter.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.