基于深度学习可分析标签相关性的心电信号分类

ECG Signals Classification by Analyzing Label Correlation Based on Deep Learning

心血管疾病是全球主要死因之一,及早发现心脏病的风险因素对于预防猝死至关重要,基于心电图的心律失常自动检测是心脏病筛查的重要手段,对降低心脏病死亡率至关重要。临床上同一患者可能同时出现多个心电异常类型,因此,解决心电图数据中多标签问题且考虑标签关系显得尤为关键。本文提出了一个可分析心电图多标签相关性的卷积神经网络分类模型,该模型包含三个模块:空间特征提取模块,卷积神经网络作为基础架构,用来提取心电图的特征;标签与特征相关性嵌入学习模块,用于探索标签之间的相关性,并将标签与特征相关性嵌入信息融合到特征空间中;解码器模块,用于预测相应心律失常类别。为解决样本不均衡,采用非对称损失函数平衡正负标签。通过对两个心电数据集进行实验,实验结果显示,本文提出的模型相较于现有心电图分类方法,表现更优异。

Cardiovascular disease is one of the leading causes of death in the world.Early detection of risk factors for heart disease is crucial for the prevention of sudden death.Clinically,multiple types of ECG abnormalities may occur in the same patient at the same time.Therefore,it is particularly important to solve the problem of multiple labels in ECG data and consider the label relationship.In this paper,a convolutional neural network classification model is proposed to analyze the correlation between ECG multi-labels.The model includes three modules.In the spatial feature extraction module,convolutional neural network is used as the infrastructure to extract ECG features.The label and feature correlation embedding learning module is used to explore the correlation between labels and fuse the label and feature correlation embedding information into the feature space.Decoder module,which predicts the probability of the corresponding arrhythmia class.Besides,to solve the data imbalance,asymmetric loss was used to balance the positive and negative labels.Experiments on two electrocardiogram datasets show that the proposed model has better performance than the existing ECG classification methods.