摘要
心房颤动(atrial fibrillation,AF)作为最常见的心律失常疾病,其早期诊断依赖于对心电图(electrocardiogram,ECG)中形态学与节律特征的精准识别。现有房颤检测方法难以兼顾ECG信号的形态特征与多尺度节律特征,导致关键病理特征建模不完整。针对这一问题,结合残差卷积网络和多尺度时序建模提出一种注意力增强双流通道模型(attention-augmented dual-stream channel model,AADM),通过协同特征提取与动态融合机制实现多维信息互补。采用残差卷积模块(residual convolutional unit,RCU)提取ECG细粒度形态特征以增强局部表征;通过双通道双向循环神经网络网络分别捕获单心跳内QRS-T节律细节与跨心跳RR间期全局时序模式;结合时空注意力机制动态融合形态-节律异质特征。通过8个公开数据集验证,该方法在房颤检测的多项指标上均表现出显著优势。
Atrial fibrillation(AF)is the most common arrhythmia disease,and its early diagnosis depends on the accurate recognition of morphological and rhythmic features in the electrocardiogram(ECG).Existing AF detection methods are difficult to take into account both the morphological and multi-scale rhythmic features of ECG signals,resulting in incomplete modeling of key pathological features.To address this bottleneck,this paper proposes an attention-augmented dual-stream channel model(AADM)by combining residual convolutional networks and multi-scale temporal modeling,and realizes multi-dimensional information complementation through collaborative feature extraction and dynamic fusion mechanism.First,the residual convolution module(RCU)is used to extract ECG fine-grained morphological features to enhance local representation;secondly,a dual-channel BiLSTM network is used to capture the QRS-T rhythm details within a single heartbeat and the global temporal pattern of the RR interval across heartbeats;finally,the spatiotemporal attention mechanism is combined to dynamically fuse the morphological-rhythm heterogeneous features.Verified by 8 public datasets,this method shows significant advantages in multiple indicators of AF detection.
作者
杨福松
刘运胜
胡峰
于洪
YANG Fusong;LIU Yunsheng;HU Feng;YU Hong(School of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Department of Military Preventive Medicine,Army Medical University,Chongqing 400038,P.R.China)
出处
《重庆邮电大学学报(自然科学版)》
北大核心
2025年第6期817-829,共13页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家重点研发项目计划(2021YFF0704103)。
