Adapting 12-Lead ECG AI Model to 1-Lead Smart Watches for Diagnosis in Clinical Heart Failure Patients

Smart Watches offer unobtrusive acquisition of 1-lead electrocardiogram (ECG); however, ECG AI models typically rely on large-scale 12-lead datasets. This study explores the feasibility of adapting an ECG AI model, originally devised for 12-lead ECGs, to 1-lead ECGs acquired with an Apple Watch from clinical heart failure patients. For evaluation, we implemented a state-of-the-art 1D-ResNet architecture and trained it on 1,676,384 ECGs from the Brazilian CODE dataset in both, a 12-lead and a 1-lead (lead I) configuration. We validate both models on the German PTB-XL 12-lead dataset (N=21,799) and demonstrate that restriction to a single lead maintained competitive performance; e.g. the F1 score for left bundle branch block (LBBB) decreased rather slightly. To test real-world application, we use 502 Apple Watch ECGs from 29 decompensated heart failure (HF) patients, focusing on four arrhythmias. As the Apple Watch ECGs are unitless, we derive a calibration factor by comparing R-peak amplitudes in simultaneous 12-lead and Apple Watch ECGs. In the Apple Watch data, we achieved a maximum accuracy of 87.10%, 93.55%, 74.19%, and 90.32% for first-degree AV block (n=9), right bundle branch block (n=3), left bundle branch block (n=12), and atrial fibrillation (n=5). These findings confirm that 1-lead recordings from consumer smart watches can be effectively integrated with ECG AI models trained on large multi-lead clinical datasets.Clinical relevance- We adapt and validate a 1D-ResNet for single-lead ECG data and test our pipeline on heart failure patients from an ongoing study conducted at the University Medical Center Göttingen. This approach bridges the gap between large clinical datasets and consumer-based smart watches, enhancing arrhythmia detection in vulnerable patient populations.