Enhancing atrial fibrillation classification from single-lead electrocardiogram signals using attention-based networks and generative adversarial networks with density-based clustering

초록

Wearable electrocardiogram devices have grown in popularity and are reliable means to record single-lead electrocardiograms continuously and in real-time. However, the challenges still exist with the noise and abnormal rhythms that make it difficult to extract powerful features from these signals. To address this, we developed a deep learning model, that merges multi-streams convolution recurrent neural networks with attention mechanisms. Clustering technique and generative adversarial networks-based data augmentation technique were used to improve this model. The method involves three main steps: data preprocessing for noise removal, quality assessment for outliers removal, and the creation of 9-second electrocardiogram segments. Finally, our newly proposed model processes these refined 9-second segments for accurate atrial fibrillation classification. Our methodology, tested on the 2017 PhysioNet Challenge dataset, initially achieved an F1score of 0.876, which increased to an impressive 0.962 after performing data augmentation using 10-fold cross-validation. This demonstrates the effectiveness of our approach in identifying atrial fibrillation from single-lead electrocardiogram data, surpassing existing state-of-the-art methods.

키워드