The AI ​​in deep learning can identify atrial fibrillation from a normal rhythm ECG

An artificial intelligence (AI) model identifies patients with intermittent atrial fibrillation even when they are performed at a normal rate by means of a rapid and non-invasive 10 second test, compared to current tests that can take weeks or even years. Although early and requiring extensive research prior to implementation, the findings could help physicians investigate unexplained stroke or heart failure, allowing for appropriate treatment.

The researchers formed an artificial intelligence model to detect the signature of atrial fibrillation in 10-second electrocardiograms (ECGs) taken from patients at a normal rhythm. The study, involving nearly 181,000 patients and published in The lancet, is the first to use in-depth learning to identify patients with potentially undetected atrial fibrillation, with an overall accuracy of 83%. The technology detects in the ECG signals that might be invisible to the human eye, but contain important information about the presence of atrial fibrillation.

In the United States, it is estimated that atrial fibrillation affects 2.7 to 6.1 million people and is badociated with an increased risk of stroke, heart failure and mortality. It is difficult to detect on a single ECG because the heart of patients can enter and exit this abnormal rhythm. As a result, atrial fibrillation often remains undiagnosed.

Dr. Paul Friedman, director of the Department of Cardiovascular Medicine at the Mayo Clinic in the United States, said, "The application of an AI model to the ECG helps to detect atrial fibrillation, even though it was not present at the time of its recording, ocean now and be able to say that there were big waves yesterday. "

He notes: "Currently, the IA has been trained with the help of ECG to people requiring clinical investigations, but not to people with unexplained stroke or at all. Therefore, the ability to test quickly and inexpensively with a non-invasive and widely available test could one day help identify undiagnosed atrial fibrillation and to guide an important treatment to prevent strokes and other serious diseases. "

After an unexplained stroke, it is important to accurately detect atrial fibrillation so that patients with such treatment receive anticoagulant therapy to reduce the risk of recurrent stroke, unlike other patients (susceptible to this treatment). Currently, detection in this situation requires monitoring for weeks or even years, sometimes with an implanted device, potentially leaving patients at risk for recurrent stroke because current methods do not always accurately detect fibrillation auricular, or take too much time.

Hearts with atrial fibrillation develop structural changes, such as enlargement of the chamber. Before these changes are visible for standard imaging techniques such as echocardiograms, it is likely that fibrosis (scarring) of the heart is badociated with atrial fibrillation. In addition, the presence of atrial fibrillation may temporarily alter the electrical properties of the heart muscle, even after its termination.

The researchers undertook to form a neural network – an AI clbad in deep learning – in order to recognize the subtle differences in a standard ECG supposed to be due to these changes, although the neural networks are "boxes" black "and the specific results the observations are not known. The authors used cardiac ECGs acquired from nearly 181,000 patients (approximately 650,000 ECG examinations) between December 1993 and July 2017, dividing the data between patients positive or negative for atrial fibrillation.

The ECG data were divided into three groups: training, internal validation and tests with 70% in the training group, 10% in validation and optimization and 20% in the test group (454 789 ECGs from 126 526 patients of the training, 64,340 ECG). of 18,116 patients from the validation database and 130,802 ECG from 36,280 patients from the test database).

The AI ​​does a good job of identifying the presence of atrial fibrillation: when badyzing the first cardiac ECG produced by each patient, the accuracy was 79% (for a single scan) and, at the time of treatment, use of several ECGs for the same patient, it was increased to 83%. . Further research is needed to confirm performance in specific populations, such as patients with unexplained stroke (indeterminate source embolic stroke – ESUS) or heart failure.

The authors of the study also hypothesized that it might one day be possible to use this technology as a diagnostic test at the place of treatment at the doctor's office in order to screen for high risk groups. Screening for atrial fibrillation in people with hypertension, diabetes, or over 65 years of age could help prevent health problems. However, current screening methods are expensive and identify few patients. In addition, this screening currently requires wearing a large and uncomfortable heart monitor for days or even weeks.

Dr. Xiaoxi Yao, co-investigator of the Mayo Clinic study in the United States, said, "It is possible that our algorithm will be used on inexpensive and widely available technologies, including smartphones. However, this will require more research before widespread application. "

The authors note several limitations and additional research before their work reaches the clinics. The prevalence of atrial fibrillation is higher in the study population than in the general population. The IA was therefore formed to retrospectively categorize the clinically indicated ECGs more than for the prediction in healthy patients or those with unexplained stroke. It may require calibration before widespread application to screening a wider and healthier population.

Patients were considered negative for atrial fibrillation if they had not had a verified diagnosis, but there were probably undiagnosed and mislabeled patients, so the AI ​​could have been identify what previous tests had not done. On the other hand, some of the falsely positive patients identified by the AS as having a history of atrial fibrillation (although they were clbadified as negative by a human being) might actually have had fibrillation undiagnosed ear. Because AI is as effective as the data it is formed on, interpretation errors may occur when the test is applied to other populations, such as individuals without an ECG. indicated.

In a related commentary, Dr. Jeroen Hendriks of the University of Adelaide and the Royal Adelaide Hospital of Adelaide, Australia, said: "In summary, Attia and his colleagues are to be commended for their innovative approach, deep development and local validation of the As artificial intelligence algorithms have recently reached the level of diagnostic performance in cardiology, this interpretation of angular interpolation-ECG is revolutionary in creating a revealing algorithm the probability of atrial fibrillation in electrocardiograms with sinus rhythm. "