Artificial intelligence helps pick up AF in ECG acquired during normal sinus rhythm

Introduction and methods

Screening for atrial fibrillation (AF) is challenging because of the low yield of a single electrocardiograph (ECG) and the cumbersome nature of prolonged monitoring. AF is underdiagnosed, which is a problem because of its association with a higher risk of stroke, heart failure and mortality. A low-cost, widely available and non-invasive test that facilitates identification of patients who are likely to have AF can therefore have important diagnostic and therapeutic implications.

Strokes with no known cause, referred to as embolic stroke of undetermined source (ESUS) are often related to AF, which can remain undetected due to its paroxysmal and often asymptomatic nature [1]. The risk of recurrent stroke is high in ESUS, but treatment with anticoagulation after ESUS to prevent recurrent stroke has only been found to beneficial when AF is indeed documented [2-5]. Thus accurate determination of AF is important to guide therapy.

Evidence is emerging that patients who develop AF have structural changes in the atria, which predispose towards atrial arrhythmias [6]. These changes might be implicated in the pathogenesis of ischemic or embolic stroke. The authors 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 [7].

Now, the authors set out to train a neural network to identify the subtle findings present in a standard 12-lead ECG acquired during normal sinus rhythm that are the consequence of structural changes associated with a history of (or impending) AF, with the aim of developing a diagnostic test. The deep neural network was trained, validated and tested using a large cohort of patients from the Mayo Clinic Digital Data Vault.

All patients of at least 18 years old 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 were included. In patients with ≥1 AF rhythm detected, the first recorded AF ECG was defined as index ECG, and the 31 days preceding the index ECG were considered the window of interest to detect potential structural changes. In patients without AF recorded on ECG, the first available ECG was considered the index ECG. All patients with their ECGs were randomized (7:1:2) to one of three groups: training, internal validation and testing dataset. Performance of the artificial intelligence (AI)-enhanced ECG to identify patients with AF was mathematically assessed by the area under the curve (AUC) of the receiver operating characteristic (ROC) curve.

Main results

• 180,922 Patients were included, with 649,931 normal sinus rhythm ECGs. The model was trained based on 454,789 ECGs, with a mean of 3.6 (SD: 4.8) ECGs per patients. Validation was done based on 64,340 ECGs, with a mean of 3.6 ECGs (SD: 4.8) per patient, and testing was done on 130,802 ECGs, with a mean of 3.6 (SD: 4.9) ECGs per patient.
• A majority (55.7%) of patients with ≥1 AF rhythm recorded in the testing dataset, had their first normal sinus rhythm ECG within 1 week of the index AF ECG.
• The ROC UAC for detection of AF was 0.87 (95%CI: 0.86-0.88) when testing the model on the first sinus rhythm ECG for each patient in the internal validation set. In the testing dataset, ROC UAC was also 0.87 (95%CI: 0.86-0.88).
• Performance was also assessed based on the sensitivity, specificity, accuracy, and F1 score of the model. An overall accuracy of the testing set of 79.4% (95%CI: 79.0-79.9) was found.
• When using multiple sinus rhythm ECGs from the same patient (in the first 31 days from the study start date), the AUC improved to 0.89 (95%CI: 0.89-0.90) using the average probability of AF scores on the test dataset. When applying a more sensitive approach of using the score of the ECG with the highest risk, AUC increased to 0.90 (95%CI: 0.90-0.91). Similar improvements were seen in the validation set.
• When using the maximum score, the F1 score, sensitivity and specificity all improved, yielding an overall accuracy of 83.3% (95%CI: 83.0-83.7) on the testing dataset.

Conclusion

References

Find this article online at The Lancet