Detection of Shockable Ventricular Arrhythmias Using Multiscale Visual Graph Network

Shockable ventricular arrhythmias (SVAs), comprising ventricular tachycardia (VT), ventricular flutter (VFL), and ventricular fibrillation (VF), are frequent contributors to sudden cardiac death (SCD). The timely and accurate automated detection of SVA is paramount for saving lives. This article introduced a method that integrated multiscale analysis of electrocardiogram (ECG) signals with complex network. The method leveraged fixed-frequency range empirical wavelet transform (EWT) filter banks to decompose ECG signals into four subsignals (0.5–32 Hz). These subsignals were subsequently mapped into complex networks using visual graph method. The network features associated with each complex network were then input into the extreme gradient boosting (XGBoost) model to identify SVA. The Creighton University ventricular tachyarrhythmia database (CUDB) and the Massachusetts Institute of Technology-Beth Israel Hospital arrhythmia (MIT-BIH) malignant ventricular arrhythmia database (VFDB) were used for model training and testing. In the classification of 10-s ECG segments, the proposed method achieved a sensitivity of 96.95%, a specificity of 98.26%, and an overall accuracy of 98.01%, demonstrated through a robust ten-fold cross-validation scheme. Similarly, impressive accuracies of 97.37%, 96.85%, and 94.74% were achieved for 8-, 5-, and 2-s segments, respectively. These compelling findings underscored the reliability and versatility of the proposed method in detecting SVA across varying signal lengths, emphasizing its potential in automated cardiac arrhythmia diagnosis.