Log in to MyACC
Artificial Intelligence-Enabled ECG Algorithm for AF
Study Questions:
Can artificial intelligence (AI) applied to a sinus rhythm electrocardiogram (ECG) help identify structural changes associated with development of atrial fibrillation (AF)?
Methods:
Included patients were 18 years or older with at least one digital, sinus rhythm, standard 12-lead ECG acquired at the Mayo Clinic between December 31, 1993, and July 21, 2017. Patients were deemed positive for AF if they had had at least one ECG with AF or atrial flutter. ECGs were assigned to the training, internal validation, and testing datasets in a 7:1:2 ratio. The area under the curve (AUC) of the receiver operator characteristic curve (ROC) was calculated for the internal validation dataset to select a probability threshold, which was then applied to the testing dataset. Performance of the AI model was evaluated by calculating the AUC, accuracy, sensitivity, specificity, and two-sided 95% confidence intervals (CIs). The authors implemented a convolutional neural network and utilized the eight nonaugmented leads (I, II, V1-6) to create an 8 x 5,000 matrix.
Results:
A total of 180,922 patients were included with 649,931 sinus rhythm ECGs for analysis. Of these, 454,789 ECGs from 126,526 patients comprised the training dataset; 64,340 ECGs from 18,116 patients comprised the internal validation dataset; and 130,802 ECGs from 36,280 patients comprised the testing dataset. 3,051 (8.4%) patients in the testing dataset had verified AF (or atrial flutter). The “window of interest” for patients without AF began at the study start date; for those with verified AF, the window of interest began 31 days prior to the index ECG that showed AF. AI modeling on the first sinus rhythm ECG identified patients with history of AF 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), and overall accuracy of 79.4% (79.0–79.9). AI modeling on all sinus ECGs acquired during the window of interest (i.e., in patients with AF, on the sinus ECGs within 31 days before the first ECG with AF) 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), and overall accuracy to 83.3% (83.0–83.7). In a secondary analysis using the first sinus ECGs after the index AF episode, the AUC of the network was also 0.90 (0.89-0.91).
Conclusions:
AI modeling of ECGs obtained in sinus rhythm can identify patients with a history of or impending AF.
Perspective:
- AF is associated with increased risk of stroke, heart failure, and mortality. Screening for AF, especially subclinical arrhythmia burden, remains a challenge due to the low yield from a single ECG and even ambulatory ECG monitoring. The authors hypothesized they could identify subtle electrical changes (e.g., wavelets within P wave due to fibrosis) on a sinus rhythm ECG indicative of structural changes associated with AF. Their novel findings support great promise for a simple single ECG in sinus rhythm to guide clinical decision-making in patients at risk for AF. The authors have previously used AI to demonstrate the capability of the resting ECG to detect abnormal electrolyte levels, antiarrhythmic drug levels, and left ventricular dysfunction.
- Further validation is needed to factor in prospective/predictive accuracy for developing AF, potential influences on structural changes (age, sex, race, comorbidities), and relative health of the population studied.
- AI modeling will greatly enhance the screening capabilities of wearable and smartphone ECG (single-lead, six-lead) technologies to detect numerous conditions, including AF.
Keywords: Anti-Arrhythmia Agents, Arrhythmias, Cardiac, Arrhythmia, Sinus, Artificial Intelligence, Atrial Fibrillation, Atrial Flutter, Electrocardiography, Electrocardiography, Ambulatory, Electrolytes, Heart Failure, ROC Curve, Secondary Prevention, Stroke, Ventricular Dysfunction, Left
< Back to Listings
