Elsevier

The Lancet

Volume 394, Issue 10201, 7–13 September 2019, Pages 861-867
The Lancet

Articles
An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction

https://doi.org/10.1016/S0140-6736(19)31721-0Get rights and content

Summary

Background

Atrial fibrillation is frequently asymptomatic and thus underdetected but is associated with stroke, heart failure, and death. Existing screening methods require prolonged monitoring and are limited by cost and low yield. We aimed to develop a rapid, inexpensive, point-of-care means of identifying patients with atrial fibrillation using machine learning.

Methods

We developed an artificial intelligence (AI)-enabled electrocardiograph (ECG) using a convolutional neural network to detect the electrocardiographic signature of atrial fibrillation present during normal sinus rhythm using standard 10-second, 12-lead ECGs. We included all patients aged 18 years or older with at least one digital, normal sinus rhythm, standard 10-second, 12-lead ECG acquired in the supine position at the Mayo Clinic ECG laboratory between Dec 31, 1993, and July 21, 2017, with rhythm labels validated by trained personnel under cardiologist supervision. We classified patients with at least one ECG with a rhythm of atrial fibrillation or atrial flutter as positive for atrial fibrillation. We allocated ECGs to the training, internal validation, and testing datasets in a 7:1:2 ratio. We calculated the area under the curve (AUC) of the receiver operatoring characteristic curve for the internal validation dataset to select a probability threshold, which we applied to the testing dataset. We evaluated model performance on the testing dataset by calculating the AUC and the accuracy, sensitivity, specificity, and F1 score with two-sided 95% CIs.

Findings

We included 180 922 patients with 649 931 normal sinus rhythm ECGs for analysis: 454 789 ECGs recorded from 126 526 patients in the training dataset, 64 340 ECGs from 18 116 patients in the internal validation dataset, and 130 802 ECGs from 36 280 patients in the testing dataset. 3051 (8·4%) patients in the testing dataset had verified atrial fibrillation before the normal sinus rhythm ECG tested by the model. A single AI-enabled ECG identified atrial fibrillation with an AUC of 0·87 (95% CI 0·86–0·88), sensitivity of 79·0% (77·5–80·4), specificity of 79·5% (79·0–79·9), F1 score of 39·2% (38·1–40·3), and overall accuracy of 79·4% (79·0–79·9). Including all ECGs acquired during the first month of each patient's window of interest (ie, the study start date or 31 days before the first recorded atrial fibrillation ECG) increased the AUC to 0·90 (0·90–0·91), sensitivity to 82·3% (80·9–83·6), specificity to 83·4% (83·0–83·8), F1 score to 45·4% (44·2–46·5), and overall accuracy to 83·3% (83·0–83·7).

Interpretation

An AI-enabled ECG acquired during normal sinus rhythm permits identification at point of care of individuals with atrial fibrillation.

Funding

None.

Introduction

Atrial fibrillation is common, underdiagnosed, and associated with an increased risk of stroke, heart failure, and mortality.1, 2 Screening for atrial fibrillation can be challenging due to the low diagnostic yield of a single electrocardiograph (ECG) to detect an often fleeting arrhythmia and the cumbersome nature of prolonged monitoring. Clinical risk scores can be used to identify patients at risk but have only modest performance. Due to these limitations, major medical societies have issued inconsistent guidelines on atrial fibrillation screening.

A low-cost, widely available, and non-invasive test that facilitates identification of patients who are likely to have atrial fibrillation would have important diagnostic and therapeutic implications. For instance, up to a third of strokes have no known cause—so-called embolic stroke of undetermined source (ESUS).3 Many of these strokes are related to atrial fibrillation, which can be under-detected due to its paroxysmal and often asymptomatic nature.4 Patients with ESUS are at high risk of a recurrent stroke, and when atrial fibrillation is documented, anticoagulation reduces the risk of recurrent stroke and might reduce mortality.5, 6 However, empirical use of anticoagulants following ESUS, whether with warfarin or a direct oral anticoagulant, has not been shown to be beneficial7, 8 and increases risk of bleeding;7, 8, 9 therefore, determination of whether atrial fibrillation is present is crucial to guide therapy.

Research in context

Evidence before this study

There is a robust literature on screening for atrial fibrillation in the general population; we did not do a formal systematic review. However, most efforts have focused on either one-time screening with a single electrocardiograph (ECG) or the use of various implantable or wearable monitors to capture infrequent atrial fibrillation episodes over time. Some studies have evaluated discrete ECG features—often P wave characteristics—as predictors of atrial fibrillation, but no individual feature has high enough predictive value to offer clinical utility using routine statistical modelling. The intensive evaluation of the ECG afforded by a convolutional neural network might be able to detect subtle, multifaceted perturbations in the ECG. We have previously shown convolution neural networks can evaluate the resting ECG for detection of antiarrhythmic drug levels, abnormal electrolytes levels, and detection of asymptomatic left ventricular dysfunction, providing proof of concept that clinically important phenomena can be detected with artificial intelligence (AI) applications to the ECG.

Added value of this study

This is the first study to our knowledge to use a convolution neural network to identify the electrocardiographic signature of atrial fibrillation present during sinus rhythm. We used an AI model to find signals in the ECG that might be invisible to the human eye but contain important information about the presence of atrial fibrillation. The AI model was trained using the standard 10-second, 12-lead ECG alone and does not require any other inputs for atrial fibrillation risk assessment. Importantly, the detection of the atrial fibrillation signal in the ECG relies on this easily obtained 10-second recording as opposed to the more invasive loop recording or cumbersome Holter monitoring. We found that an AI model can differentiate between patients with a history of (or impending) atrial fibrillation with a high degree of accuracy using a single routine ECG. Addition of multiple ECGs within an individual patient improved the model accuracy and suggests repeated measures might yield even better performance.

Implications of all the available evidence

Our study supports the hypothesis that subtle patterns on the normal sinus rhythm ECG can suggest the presence of atrial fibrillation. The ability to identify patients with potentially undetected atrial fibrillation using an inexpensive, non-invasive, widely available, point-of-care test has important practical implications for atrial fibrillation screening and potentially for the management of patients with prior stroke of unknown cause.

Prolonged ambulatory cardiac rhythm monitoring is frequently used to screen for atrial fibrillation, particularly after ESUS. Approaches include insertion of implantable loop recorders and wearable patches.1, 2 These strategies are invasive or inconvenient, expensive, require a monitoring infrastructure, and have a low yield.10

There is growing evidence that patients who develop atrial fibrillation—even in an apparently normal heart—have structural changes in the atria that predispose towards atrial arrhythmias;11 these changes might be important for the pathogenesis of ischaemic or embolic stroke. We have previously used machine learning in the form of deep neural networks to identify subtle patterns in the standard 12-lead ECG to identify the presence of asymptomatic ventricular dysfunction.12

We hypothesised that we could train a neural network to identify the subtle findings present in a standard 12-lead ECG acquired during normal sinus rhythm that are due to structural changes associated with a history of (or impending) atrial fibrillation. Such a diagnostic test could be inexpensive, widely available, and immensely useful following ESUS to guide therapy. To test this hypothesis, we trained, validated, and tested a deep neural network using a large cohort of patients from the Mayo Clinic Digital Data Vault.

Section snippets

Data sources and study population

We included all patients aged 18 years or older with at least one digital, normal sinus rhythm, standard 10-second, 12-lead ECG acquired in the supine position at the Mayo Clinic ECG laboratory between Dec 31, 1993, and July 21, 2017. All ECGs were acquired at a sampling rate of 500 Hz using a GE-Marquette ECG machine (Marquette, WI, USA) and the raw data were stored using the MUSE data management system. ECGs in our laboratory are intially read by the GE-Marquette ECG system and then over-read

Results

We identified 210 414 patients with 1 000 000 ECGs and, after applying exclusion criteria, included 180 922 patients with 649 931 normal sinus rhythm ECGs for analysis (figure 2). We trained the model using 454 789 ECGs recorded from 126 526 patients, with a mean of 3·6 ECGs (SD 4·8) per patient. In patients with at least one atrial fibrillation recorded in the testing dataset, 1698 (55·7%) of the 3051 first normal sinus rhythm ECGs in the window of interest were within 1 week of the index

Discussion

In this study, we found that the AI-enabled ECG recorded during normal sinus rhythm performed well (AUC 0·87 for a single ECG and 0·90 for multiple) in identifying the presence of atrial fibrillation. This compares favourably with other medical screening tests such as B-type natriuretic peptide for heart failure (AUC 0·60–0·70),20 Papanicolaou smear for cervical cancer (AUC 0·70),21 and the CHA2DS2-VASc Score for stroke risk (AUC 0·57–0·72). 22 The ability to identify undetected atrial

References (30)

  • RB Schnabel et al.

    50 year trends in atrial fibrillation prevalence, incidence, risk factors, and mortality in the Framingham Heart Study: a cohort study

    Lancet

    (2015)
  • T Sanna et al.

    Cryptogenic stroke and underlying atrial fibrillation

    N Engl J Med

    (2014)
  • DJ Gladstone et al.

    Atrial fibrillation in patients with cryptogenic stroke

    N Engl J Med

    (2014)
  • DT Martin et al.

    Randomized trial of atrial arrhythmia monitoring to guide anticoagulation in patients with implanted defibrillator and cardiac resynchronization devices

    Eur Heart J

    (2015)
  • RG Hart et al.

    Rivaroxaban for stroke prevention after embolic stroke of undetermined source

    N Engl J Med

    (2018)
  • Cited by (760)

    View all citing articles on Scopus

    Contributed equally

    View full text