基于Bi‑LSTM和时序注意力的异常心音检测  

Abnormal heart sound detection using Bi‑LSTM and temporal attention mechanism

作　　者：卢官明[1] 蔡亚宁 卢峻禾 戚继荣[3] 王洋 赵宇航 LU Guanming;CAI Yaning;LU Junhe;QI Jirong;WANG Yang;ZHAO Yuhang(School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;Department of Cardiothoracic Surgery,Children’s Hospital of Nanjing Medical University,Nanjing 210008,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学计算机学院,江苏南京210023 [3]南京医科大学附属儿童医院心胸外科,江苏南京210008

出　　处：《南京邮电大学学报(自然科学版)》2025年第1期12-20,共9页Journal of Nanjing University of Posts and Telecommunications:Natural Science Edition

基　　金：国家自然科学基金(72074038);江苏省卫生健康委员会重点项目(K2023036)资助项目。

摘　　要：异常心音检测是对心脏病进行初步诊断的一种有效而方便的方法。为提升异常心音的检测性能,提出了一种基于双向长短时记忆网络(Bi⁃directional Long Short⁃Term Memory,Bi⁃LSTM)和时序注意力的异常心音检测算法。首先对心音片段进行分帧处理,使用平均幅度差函数(Average Magnitude Difference Function,AMDF)和短时过零率(Short⁃Time Zero⁃Crossing Rate,STZCR)提取每帧心音信号的初始特征;然后将它们拼接后作为Bi⁃LSTM的输入,并引入时序注意力机制,挖掘特征的长期依赖关系,提取心音信号的上下文时域特征;最后通过Softmax分类器,实现正常/异常心音的分类。在PhysioNet/CinC Challenge 2016提供的心音公共数据集上对所提出的算法使用10折交叉验证法进行了评估,其准确度、灵敏度、特异性、精度和F1评分分别为0.9579、0.9364、0.9642、0.8838和0.9093,优于已有的其他算法。实验结果表明,该算法在无需进行心音分段的基础上就能有效实现异常心音检测,在心血管疾病的临床辅助诊断中具有潜在的应用前景。Abnormal heart sound detection is an effective and convenient method for the preliminary di⁃agnosis of heart diseases.To improve the effectiveness of abnormal heart sound detection,this paper pro⁃poses an algorithm using the bi-directional long short-term memory(Bi-LSTM)network and the temporal attention.First,the heart sound segments are partitioned into frames,and the initial features of each frame of the heart sound signal are extracted using the average magnitude difference function(AMDF)and short-time zero-crossing rate(STZCR).Second,these initial features are concatenated and input into the Bi-LSTM.The temporal attention mechanism is introduced to explore the long-term dependencies of features and extract the contextual temporal features of heart sound signals.Finally,the Softmax classi⁃fier is used to classify normal/abnormal heart sounds.The proposed algorithm is evaluated on the heart sound public dataset provided by PhysioNet/CinC Challenge 2016.The classification performance from a 10-fold cross-validation indicates that the accuracy,sensitivity,specificity,precision,and F1 score are 0.9579,0.9364,0.9642,0.8838,and 0.9093,respectively,superior to those of the existing algo⁃rithms.The experimental results show that the proposed algorithm can effectively detect abnormal heart sounds without the need for heart sound segmentation,and has potential application prospects in clinical auxiliary diagnosis of cardiovascular diseases.

关 键 词：心音分类 平均幅度差函数 短时过零率 双向长短时记忆网络 时序注意力机制 

分 类 号：TN912.3[电子电信—通信与信息系统]