Background: Atrial fibrillation (AF) is the most prevalent cardiac arrhythmia, and its early detection is critical for mitigating risks such as stroke and heart failure. While wearable single-lead ECG patches provide a significant potential for early AF due to its portability and user-friendliness for long-term continuous monitoring, there have been limited studies in leveraging advanced AI models for AF detection using data from these wearable ECG patches in patients with paroxysmal AF.
Objective: This study aims to evaluate the effectiveness of convolutional neural networks (CNN) in detecting AF from single-lead 14-day ECG wearable patches.
Method: We considered continuous 14-day single-lead ECGs collected with a ZioXT ECG patch from 89 AF patients prescribed with catheter ablation, collected during the post-procedure blanking period (2-weeks post-procedure). Manual adjudication of cardiac rhythm was performed by ECG technicians with 5% of the data reviewed by an independent cardiologist for accuracy. AF annotation was performed manually by ECG experts. The signals were converted into 10 second 200Hz segments. A CNN was developed with two layers of convolutions with 16 and 32 filters respectively, followed by two fully connected layers with a sigmoid activation function for AF detection. Drop out layers are used to mitigate overfitting.
Results: ECG segments were split into 3 disjoint sets of 4.3 million segments (65 subjects), 1.1 million (12 subjects), and 1.1 million (12 subjects) for training, validation, and test. On ECG segments from the held-out subjects, the model offered an accuracy of 89%, specificity of 91%, and sensitivity of 72% for AF detection.
Conclusion: This study demonstrates the feasibility of using CNNs for detecting AF in very short ECG epochs extracted from single-lead long-term ECG wearable patches. The model could be used for on-board patch notification system and enables rapid intervention for patients with recurrent AF episodes.