LAUSR

Ventricular fibrillation (VF) and rapid ventricular tachycardia (VT) of a heart rate over 180 beats min−1, commonly known as shockable rhythms (ShR), are life-threatening ventricular arrhythmias. This study was to assess the feasibility of feeding two-dimensional (2D) time-frequency maps of electrocardiogram (ECG) segment into deep convolutional neural network (CNN) to automatically detect ShR with emphases on optimizing the CNN model and shortening the analysis segment. A total of 115 single-lead surface ECG records collected from four publicly accessible ECG databases were divided into non-overlapping 3-second segments as well as 5-, 8-, and 10-second segments for comparisons. For each one-dimension (1D) ECG segment, its 2D time-frequency maps was obtained by employing a continuous wavelet transform (CWT). Moreover, the performance of automated algorithms with eight different CNN models was evaluated for ShR detection against annotations labeled in the databases. A total of 5,795 3-second maps of ShR and 48,648 maps of non-shockable rhythms (NSR) were analyzed successfully, and a promising performance of the automated algorithm with a twelve-layer CNN model was achieved on this segment length in terms of accuracy (98.82%), sensitivity (95.05%), and specificity (99.43%). Compared with current approaches imposing priority either only on highly real-time or high performance, the presented solution suggested that the deep learning based algorithm together with ECG time-frequency maps may offer the possibility of providing an intelligent and efficient detection of ShR and NSR in short segments especially for public access defibrillation.