부산대학교 도서관

Background: Wearable and portable devices with ECG capabilities have the potential to enable broad AI-based screening for clinical disorders, but models trained on single-lead signals are limited by challenges accessing the raw signal data from proprietary systems. Additionally, wearable ECGs have traditionally been limited to diagnosing only a few rhythm disorders.Methods: From 385,601 clinical ECGs performed at Yale during 2015-2021, we extracted lead I data, and simulated wearable PDF outputs from commercial wearable devices. Gold-standard diagnosis statements from cardiologists accompanying these ECGs were processed and used to generate 41 clinical labels. We developed a multilabel model to predict these labels from wearable-adapted PDF outputs using an EfficientNet-B3 convolutional neural network with self-supervised pretraining. The model deployment pipeline was optimized to segment and process PDF outputs across all commercially available devices as inputs to the model (Fig A & B).Results: In a held-out test set of 11,611 ECGs, the model accurately detected a diverse set of disorders. For the rhythm disorders sinus tachycardia, sinus bradycardia, atrial fibrillation, atrial flutter, premature atrial contraction, and premature ventricular contraction, the model had AUROCs of 0.99, 0.98, 0.98, 0.93, 0.96 and 0.95 respectively. For the conduction disorders left bundle branch blocks, right bundle branch blocks, 1 and 2 degree AV blocks, the model had AUROCs of 0.98, 0.97, 0.96 and 0.95, respectively. Finally, for key disorders such as STEMI, LVH, and low limb-lead voltage, the model had AUROCs of 0.81, 0.86, and 0.77, respectively (Fig C).Conclusions: We developed and validated a deep learning model to identify conduction disorders and arrhythmias from PDF image outputs of wearable and handheld single-lead devices. This approach allows for automated and easily accessible insights into wearable and portable ECGs at the point of care.