AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Translating Single-Lead Artificial Intelligence-Enhanced Electrocardiography from Hospital to Wearable Recordings: Multistage Validation Across International Clinical and Consumer Cohorts
By El-Medany A, Patlatzoglou K, Aowidah R et al.
Introduction Single-lead AI-ECG may enable scalable cardiovascular screening using wearable devices. However, clinical Lead I performance may not translate directly to wearable recordings due to population, acquisition, noise, and proprietary processing differences. Research Questions We assessed transportability of single-lead AI-ECG from retrospective clinical Lead I ECGs to real wearable recordings and prospective cohorts, using biometric prediction, signal robustness, and transthoracic echocardiographic (TTE) structural heart disease (SHD) detection. Methods A Contrastive ECG-Language Alignment model, pretrained on 10.6M Massachusetts General Hospital ECG-report pairs, was fine-tuned using 77K Beth Israel Deaconess Medical Center (BIDMC) paired Lead I ECG-TTE data and externally validated in EchoNext-Mini for left ventricular ejection fraction (LVEF) ≤45%, and moderate/severe aortic stenosis (AS) and mitral regurgitation (MR). Transportability to wearables was assessed in retrospective Taiwan Apple Watch ECGs using the same SHD endpoints and BIDMC-trained Lead I ResNet models for sex and BMI prediction. Biometric models were evaluated on volunteer Apple Watch recordings. Signal-domain analyses compared raw/enhanced Apple Watch recordings and no-noise, synthetic-noise, and extracted-noise augmentation. Apple Watch and clinical Lead I validation were performed in TTE patients. Performance was assessed using AUROC, sensitivity, specificity, predictive values, and mean absolute error (MAE). Results In retrospective Lead I validation, AUROCs were preserved between BIDMC internal and EchoNext-Mini external validation for LVEF ≤45% (0.89/0.89), AS (0.82/0.84), and MR (0.81/0.83). In the BIDMC biometric benchmark, BMI MAE was 4.69 kg/m ² and sex AUROC 0.83. In Taiwan Apple Watch ECGs, SHD AUROCs attenuated to 0.76/0.70/0.67, respectively; biometrics were sex AUROC 0.68 and BMI MAE 5.12 kg/m ² . Healthy-volunteer Apple Watch recordings showed BMI MAE 5.44 kg/m ² and sex AUROC 0.72. Extracted-noise augmentation improved simulated-noise robustness, but did not significantly outperform no-noise training. Signal enhancement reduced sex AUROC from 0.72 to 0.65, while improving BMI MAE from 5.44 to 4.71 kg/m ² . In prospective ECG-TTE validation, Apple Watch AI-ECG detected LVEF ≤45% with AUROC 0.85 (95% CI 0.77-0.92) and sensitivity 92.9%, with discrimination for AS and MR of 0.80 (0.58-0.95) and 0.80 (0.66-0.92). Clinical Lead I achieved similar LVEF ≤45% discrimination (AUROC 0.88; DeLong p=0.49 vs Apple Watch). Conclusion Single-lead AI-ECG showed preserved performance across clinical ECG datasets, but translation to wearables was task-specific and influenced by device processing, signal quality and population shift. Wearable AI-ECG shows clinical promise, but device-acquired prospective validation is essential before deployment.