A Modified State Refinement GRU for Atrial Fibrillation and Premature Ventricular Contraction Signals Detection
针对房颤和室性早搏两种常见心律失常，提出一种改进的状态细化门控循环单元（MSR-GRU），通过信号点注意力和表示门机制自动检测心电图信号，比传统GRU和现有方法更有效，并首次将心电图内部交互特性融入神经网络以增强可解释性。
Atrial fibrillation (AF) and premature ventricular contraction (PVC) are two typical and common arrhythmias. However, the conventional AF and PVC detection strategy draws support from manual examination of electrocardiogram (ECG) and in turn translates to a labor-intensive procedure. Motivated from state refinement for LSTM, we develop a modified state refinement gated recurrent unit (MSR-GRU) that adaptively refines the current states of signal points in ECG with a signal point-wise attention and a representation gate mechanism for automated AF and PVC detection. Comparative experiments across two public databases confirm the effectiveness of MSR-GRU enjoying a consistent improvement than plain GRU and current methods. In particular, we are the first to incorporate the interaction properties within ECG into neural network thus offering interpretability.