PerC-SimAM-BLSTM: Position Perception Circular Convolution With Simple Attention Mechanism Based on BLSTM for Bundle Branch Block Detection

Left bundle branch block (LBBB) and right bundle branch block (RBBB) are two typical types of arrhythmias. Due to their similar electrocardiogram (ECG) features, the conventional visual inspection ECG strategy has always been challenging. Motivated by MetaFormer, we investigate a novel MetaFormer-like model that is composed of both position perception circular convolution (PerC) as a new Token Mixer and simple attention mechanism with bidirectional long short-term memory (SimAM-BLSTM) as a new Channel MLP called PerC-SimAM-BLSTM for automatic LBBB and RBBB detection. This framework not only inherits the global feature extraction ability of Transformer but also enhances global-local feature acquisition capability of PerC. Experiment results demonstrate better improvements than the existing strategies and also strike a better balance between performance and complexity from an empirical experiment perspective. Mathematical convergence proof ensures stable convergence of model training to the theoretical optimal solution from a theoretical perspective. Our PerC-SimAM-BLSTM achieves an accuracy of over 99.1% across the China Physiological Signal Challenge 2018 and MIT-BIH arrhythmia databases (MADBs), demonstrating great potential in lightweight devices.