基于多频段多任务编解码模型的心电图基准点联合检测

Joint detection of ECG Fiducial points based on multi-band and multi-task encoding and decoding model

中文关键词  心电图基准点检测  编解码模型  时域卷积神经网络

英文关键词  ECG fiducial point detection  encoding-decoding model  temporal convolutional network

中文摘要

      基准点检测是心电图(ECG)诊断分析的基础. 但是, ECG具有波形变异性, 且经常受到各种伪迹和噪声的干 扰, 使得基准点检测精度受限. 针对该问题, 本文首先构建概率图模型, 分析各频带ECG成分与基准点检测任务之间 的推断关系. 然后, 在此概率图模型的启发下提出了一种多频段多任务编解码网络. 该网络先分别对不同频段的 ECG成分进行一维卷积变换提取特征. 然后, 通过时域卷积模组学习各频段特征的注意力掩码以抵御噪声. 最后, 利用多分支关联结构, 实现多个ECG基准点的联合检测. 在MIT-BIH QT和LUDB数据集上的五重交叉验证实验结 果表明: 所提方法能够有效地提高ECG基准点的检测精度, 达到了当前最优的水平.

英文摘要

      Fiducial point detection is the basis of the electrocardiogram (ECG) diagnostic analysis. However, the ECG has waveform variability and is often disturbed by various artifacts and noises, limiting the detection accuracies of fiducial points. This paper first builds a probability graph model to analyze the inference relationships between different band ECG components and fiducial point detection tasks. Then, we propose a multi-band multi-task encoding-decoding network inspired by this probability graph model. The network first performs 1-D convolutions on each ECG component to extract features, then learns the attention masks to resist noise through temporal convolutional modules, and finally adopts the dependent multi-branch structure to realize the joint detection of ECG fiducial points. The experimental results with fivefold cross-validation on the MIT-BIH QT and LUDB databases show that the proposed method can effectively improve the detection accuracy of ECG fiducial points, comparable to the state-of-the-art level