A deep learning model applied to handheld single-lead ECGs combined with age and sex discriminated 2-year incident AF risk (AUROC 0.695; 95% CI 0.637-0.742) comparably to the CHARGE-AF score.
Observational (n=15,694)
Does a deep learning model applied to handheld single-lead ECGs predict 2-year incident AF in primary care patients aged ≥65 years?
An AI algorithm applied to handheld single-lead ECGs can predict 2-year incident AF risk with performance comparable to clinical risk scores, potentially improving AF screening efficiency.
Effect estimate: AUROC 0.695 (95% CI 0.637-0.742)
Absolute Event Rate: 0.695% vs 0.679%
Background: Deep learning of 12-lead electrocardiograms (ECGs) can estimate future atrial fibrillation (AF) risk. Whether these models can be applied to single-lead ECGs (1L ECGs) from wearable or handheld devices is unknown. Aim: To apply a previously published deep learning model to estimate 2-year incident AF risk using handheld 1L ECGs obtained prospectively in a large AF screening trial. Methods: In the VITAL-AF trial (NCT03515057), over 16,000 primary care patients aged ≥65 years underwent AF screening with handheld 1L ECG. We applied our published deep learning model (ECG-AI) derived using single leads of over 450,000 standard 12-lead ECGs external to the VITAL-AF study (1L ECG-AI). We compared the performance of 1L ECG-AI to the validated CHARGE-AF clinical risk score and calculated net reclassification indices. Results: Among 15,694 individuals enrolled in VITAL-AF and without prevalent AF (age 74±7 years, 9,104 58% women), 1L ECG-AI discriminated 2-year incident AF (area under receiver operating characteristic curve AUROC 0.666 95% CI 0.603-0.721). With addition of age and sex (1L ECG-AI AS), 2-year AF discrimination (0.695 0.637-0.742) was comparable to the 11-component CHARGE-AF clinical risk score (0.679 0.625-0.729). Two-year AF incidence was markedly higher with 1L ECG-AI AS in the top 5% (8.0% 5.7-10.2) vs bottom 5% (0.89% 0.23-0.55). At a threshold of ≥3% estimated 2-year AF risk, AF incidence was progressively higher among patients at high risk according to neither model (1.5% 1.2-1.8), one model (1L ECG-AI: 3.3% 2.3-4.2; CHARGE-AF 3.4% 2.2-4.6), and both models (5.8% 5.0-6.5), implying that clinical risk and ECG-AI signals are complementary ( Figure ). Compared to screening all individuals at the guideline-based threshold of ≥65 years, 1L ECG-AI AS resulted in favorable net reclassification improvement (0.27 0.22-0.32). Conclusion: An ECG-AI algorithm developed using single leads of a 12-lead ECG combined with age and sex can discriminate AF risk using real-world handheld 1L ECGs with comparable performance to the CHARGE-AF score. 1L ECG-AI signals complement clinical risk factors. ECG-AI applied to 1L ECG may increase the reach and efficiency of AF screening.
Khurshid et al. (Tue,) conducted a observational in Incident atrial fibrillation (n=15,694). 1L ECG-AI combined with age and sex (1L ECG-AI AS) vs. CHARGE-AF clinical risk score was evaluated on 2-year incident AF discrimination (AUROC) (AUROC 0.695, 95% CI 0.637-0.742). A deep learning model applied to handheld single-lead ECGs combined with age and sex discriminated 2-year incident AF risk (AUROC 0.695; 95% CI 0.637-0.742) comparably to the CHARGE-AF score.