Abstract Body (Do not enter title and authors here): Background
Artificial intelligence-enhanced electrocardiograms (AI-ECGs) based on single-lead recordings have demonstrated potential in detecting left ventricular systolic dysfunction (LVSD). However, validation studies remain limited due to a lack of sufficient data from self-recorded smartwatch ECGs. To address this gap, we established a real-world cohort of smartwatch ECG recordings.
Hypothesis
This study aims to determine whether AI-ECG scores derived from self-recorded smartwatch ECGs can reliably screen for and monitor LVSD in practical settings.
Methods
Between July and December 2024, we recruited participants who had recently completed or were scheduled for echocardiography (Echo) and instructed them to record smartwatch ECGs (Samsung Galaxy or Apple Watch) at least twice daily for over a week, ensuring an Echo within 14 days. Our previously developed convolutional neural network-based AI-ECG model was fine-tuned for smartwatch ECGs using a foundation model with preprocessing to address smartwatch-specific signal noise. The model outputs a score ranging from 0 to 100, where higher values indicate an increased likelihood of LVSD. We assessed model performance using three approaches: (1) Comprehensive analysis of all available ECGs, (2) Global sampling using three random ECGs from each participant’s entire dataset, and (3) Daily sampling with three random ECGs per day per participant.
Results
A total of 30 participants were enrolled in the study, with 78.1% using Samsung Galaxy Watches and 21.9% using Apple Watches. Echo identified LVSD in 7 participants (23.3%). The median AI-ECG score was 57.2 among those with LVSD and 4.7 in participants without LVSD. In total, 1,540 ECGs were collected, including 906 from the LVSD group. The AUROC was 0.918 (95% CI: 0.905–0.931) when all available ECGs were analyzed, 0.910 when three random ECGs per participant were analyzed, and 0.900 when three random ECGs per day per participant were analyzed. When the repeated measurements for all participants were plotted using a Generalized Additive Model (GAM), some degree of variability was observed, but the score distribution for each participant remained relatively consistent.
Conclusion
Our findings show that an AI-driven single-lead ECG method can reliably track LVSD using self-recorded smartwatch data in real-world environments. This evidence strongly supports expanding the application of our AI-ECG model to interpret smartwatch ECG recordings.