Detecting Atrial Fibrillation from Single-Lead ECG Using Unbalanced Multi-classification Support Vector Machine

Atrial fibrillation (AF) is an common arrhythmia. The incidence of AF has been increasing with the acceleration of urbanization and social aging. Therefore, the wearable ECG acquisition devices with single-lead ECG came out for early diagnosis, monitoring and management of AF. However, it is still great challenge to accurately detect AF from massive ECG data. This study proposed a method detecting AF from single-lead ECG signals based on unbalanced multi-classification support vector machine(SVM). The novel method first screened 73 effective features by correlation analysis from 110 candidate features, which have been confirmed to be associated with AF in literature. Then, an unbalanced four-class SVM classifier was designed to detect four types of ECG signals (including AF, other arrhythmia, artifactual and normal) based on the distribution of different types of ECG data. Finally, the data provided by the PhysioNet/Computing in Cardiology Challenge 2017 confirmed that the proposed method had a overall good performance compared with five other related methods. Also, the data from MIT Arrhythmia Database and the MIT Atrial Fibrillation Database confirmed the robustness of proposed method with AF detection score of > 0.97 and with the scores of > 0.9 in other arrhythmia, artifactual and normal. The proposed method has a good application prospect in AF aided diagnosis, monitoring and management of AF.