A novel Domain Adaptive Residual Network for automatic Atrial Fibrillation Detection

Abstract Atrial fibrillation (AF) is the most common cardiac arrhythmia and shows a rising trend with the increase of aged. Currently, existing intelligent AF detection methods have achieved good results in massive labeled data. However, it is time-consuming and undesirable to label ECG signals in real applications. Meanwhile, due to distribution discrepancy by different testing conditions, it is unsatisfied for directly applying trained model to other datasets. Inspired by the domain adaptation techniques, this paper proposes a novel Domain Adaptive Residual Network (DARN) to detect AF of unlabeled datasets with the aid of detection knowledge of labeled dataset. Firstly, residual blocks are adopted to extract informative deep features from the ECG signals automatically. Then, deep features are fed into feature classifier to acquire final detection result. Further, the multi-layer multi-kernel maximum mean discrepancy is combined into the training process to reduce distribution discrepancy of different domains, which imposes constraints on network parameters. Finally, the proposed method was evaluated with the data from MIT-BIH Atrial Fibrillation Database (AFDB), MIT-BIH Arrhythmia Database and 2017 Physionet challenge dataset. The experimental results show that the proposed domain adaptive approach improves the accuracy by 4.50% on average and the F1 score by 4.28% on average using the knowledge of AFDB. Additionally, comparison experiment shows that the proposed feature extractor and classifier achieved 98.97%, 98.75%, and 98.84% for the sensitivity, specificity, and accuracy on the AFDB, respectively. Consequently, the proposed method is provided with high application potential as a valuable auxiliary tool for clinical AF detection.