1 Two-stream Network for ECG Signal Classiﬁcation

2025-04-27 0 0 650.57KB 9 页 10玖币

侵权投诉

Two-stream Network for ECG Signal

Classiﬁcation

Xinyao Hou, Shengmei Qin, Jianbo Su, Senior Member, IEEE

Abstract—Electrocardiogram (ECG), a technique for medical monitor-

ing of cardiac activity, is an important method for identifying cardiovas-

cular disease. However, analyzing the increasing quantity of ECG data

consumes a lot of medical resources. This paper explores an effec-

tive algorithm for automatic classiﬁcations of multi-classes of heartbeat

types based on ECG. Most neural network based methods target the

individual heartbeats, ignoring the secrets embedded in the temporal

sequence. And the ECG signal has temporal variation and unique indi-

vidual characteristics, which means that the same type of ECG signal

varies among patients under different physical conditions. A two-stream

architecture is used in this paper and presents an enhanced version

of ECG recognition based on this. The architecture achieves classiﬁ-

cation of holistic ECG signal and individual heartbeat and incorporates

identiﬁed and temporal stream networks. Identiﬁed networks are used

to extract features of individual heartbeats, while temporal networks

aim to extract temporal correlations between heartbeats. Results on the

MIT-BIH Arrhythmia Database demonstrate that the proposed algorithm

performs an accuracy of 99.38%. In addition, the proposed algorithm

reaches an 88.07% positive accuracy on massive data in real life,

showing that the proposed algorithm can efﬁciently categorize different

classes of heartbeat with high diagnostic performance.

Index Terms—Heartbeat, Electrocardiogram, Pattern recognition, Deep

learning

1 INTRODUCTION

THE cardiovascular disease (CVD) is a major threat to

human health and the leading cause of death interna-

tionally. According to the World Health Organization, is-

chaemic heart disease and all forms of stroke are responsible

for a quarter of the total number of deaths globally [1].

Electrocardiogram (ECG) records electrical activity gener-

ated by the heart. ECG is used for emotion recognition [2]

and biometric identiﬁcation [3] due to its unique individual

characteristics. ECG analysis is also an effective tool for

assessing cardiac health, as it is easy to use as a non-invasive

method. And ECG is one of the most effective way to study

the classiﬁcation of cardiac arrhythmias, which is of great

interest to biomedical researchers. Therefore, accurate and

effective analysis of ECG signals is critical in the diagnosis

•X. Hou and J. Su are with the Key Laboratory of System Control

and Information Processing, Department of Automation, Ministry of

Education, Shanghai Jiao Tong University, Shanghai 200240, China

(e-mail: hxy0928@sjtu.edu.cn; jbsu@sjtu.edu.cn).

•Qin is with Zhong Shan Hospital, No 180, Fenglin Road, Shanghai;

200032, China (e-mail: Qin.shengmei@zs-hospital.sh.cn).

and treatment of cardiovascular disease [4]. However, there

is a wide variety of heart diseases and a sheer volume of

ECG data, which means that long-term reliance on a manual

diagnosis not only easily introduces errors and leads to un-

stable diagnosis results, but also consumes a lot of medical

resources. Therefore, there is an urgent need to realize a

quick, accurate, and robust analysis of ECG data. Due to

the time variation and unique individual characteristics of

the ECG, the feature of the same kind of ECG signal varies

in different patients under different physical conditions,

which poses considerable difﬁculties for recognition of ECG

patterns [5]. Many works have been carried out to advance

the automatic analysis of ECG signals. In recent years,

with the maturity and popularity of artiﬁcial intelligence

technology, machine learning has become widely used in

ECG signal analysis. The typical method is to divide the task

into extraction and classiﬁcation of ECG signal features.

The correct representation of the ECG signal plays an im-

portant role in the diagnosis of heart disease, and different

types of feature extraction techniques are used. In general,

the P-QRS-T complex is the primary unit of the ECG signal

[6], as shown in Fig. 1. As the attributes of the P-QRS-T

complex of healthy people are relatively ﬁxed, it is often

used as a reference by doctors when diagnosing the ECG

signal, by comparing the normal values in the sinus rhythm

for a healthy male adult with those of the patients [7]. Based

on this approach, some researchers adopt a similar approach

by using computer methods instead of manually extracting

features of the QRS complex. There are many feature de-

tection algorithms for the QRS complex, such as derivative-

based [8], wavelet variation [9] and digital ﬁlters [10]. In

addition, high-order moments are utlized to extract features

of QRS complexes in [11]. Some studies make use of statisti-

cal attributes of ECG signals, which provide the complexity

and distribution of the ECG to achieve improved recogni-

tion performance [12]. Frequency-domain-based techniques

are also popular for classifying ECG signals [13]. And the

ECG classiﬁcation is achieved by various machine learning

algorithms, represented by K-nearest neighbor (KNN) [9],

artiﬁcial neural network (ANN) [14], linear discriminant

analysis (LDA) [15] and support vector machine (SVM) [16].

In [17], a ﬁltered feature selection method and SVM are com-

bined to achieve early detection of ventricular ﬁbrillation

(VF) and rapid ventricular tachycardia (VT) in the MIT-BIH

dataset. A block-based neural network (BbNN) is proposed

in [18], which is created by a group of bi-dimensional block

arXiv:2210.06293v1 [eess.SP] 5 Oct 2022

networks with ﬂexible structure and internal conﬁguration.

This work overcomes the possibility of varying ECG signals

over time and individual differences, achieving 97% classi-

ﬁcation accuracy on MIT-BIH.

Fig. 1. Standard ﬁducial points in the ECG. The ECG signal consists of

ﬁve major deﬂections, including P, Q, R, S and T, plus a small deﬂection,

known as the U wave [19].

Although these methods work well, they are difﬁcult to

implement and use because most are based on manually de-

signed features. Different feature extraction techniques are

used to cope with different diseases due to the importance of

proper representation of the ECG signal. But there are many

diseases and it is impossible to design features for each of

them. Compared with traditional methods, deep learning

can automatically extract features and perform complex

data preprocessing. In recent years, deep learning has made

impressive achievements in computer-aided medical diag-

nosis [20] and is used in ECG classiﬁcation [21]. In [22], ECG

signals are transformed into grayscale images and computer

vision techniques are used. An end-to-end deep learning

algorithm for ECG analysis is proposed, which employs a

deep neural network (DNN) to classify the class of ECG

[23]. In practice, many diseases need to be diagnosed on

the holistic ECG signal, as some of them appear over time.

And most of the previous works focus on the classiﬁcation

of individual heartbeats instead of the holistic ECG signal.

There is very little effort devoted to classify the holistic

ECG signal. Hence it motivates us to achieve the classiﬁ-

cation of the holistic ECG signal. As the ECG signal has

temporal variation and unique individual characteristics,

which means that the same type of ECG signal varies

among patients under different physical conditions, a two-

stream architecture is proposed in this paper. The archi-

tecture incorporates identiﬁed and temporal networks and

accurately classify the holistic ECG over a long period.

Speciﬁcally, individualized networks are used to extract fea-

tures of individual heartbeats while temporal networks are

employed to extract temporal correlations between heart-

beats, taking into account temporal variation and unique

individual characteristics of the ECG. And with the pur-

pose of demonstrating the generalization and excellence of

our architecture, seven detailed categories of heartbeat are

collected, each containing data from a thousand different

adults, also expending the ECG study with deep learning.

The work operates as an inter-patient paradigm rather than

an intra-patient paradigm.

Section 2 mainly introduces the proposed architecture

for ECG recognition and provides a detailed description of

its internal structure. In Section 3, the datasets uesd and

the preprocessing are illustrated and the implementation

details of the experiments are outlined. In Section 4, the

performance of the architecture is evaluated in the MIT-BIH

dataset and real life. The architecture is discussed in Section

5 and this paper is summarized in Section 6.

2 TWO-STREAM ARCHITECTURE FOR ECG

RECOGNITION

The ECG signal is inﬂuenced by the object and time, which

means that the same type of ECG signal varies among pa-

tients under different physical conditions. And inspired by

the Two-Stream networks of action recognition [24], where

the action is composed of spatial and temporal stream, the

two-stream architecture of ECG recognition is proposed.

In this architecture, the ECG signal can be decomposed

into identiﬁed and temporal components. The identiﬁed

part, in the form of an individual heartbeat (represented

as the P-QRS-T complex), contains information about the

unique individual characteristics represented in the ECG

signal. The temporal part, in the form of a holistic ECG

signal (represented as the combination of multiple P-QRS-

T complexes), transmits the symptoms of the ECG and

its changes over time. The corresponding ECG recognition

architecture accordingly is designed and divided it into two

streams, as shown in Fig. 2. Each stream is implemented

with a neural network, whose result is combined by late

fusion. Two fusion methods are considered: averaging the

output scores and training a fully connected layer on the

stacked features extracted from each stream.

Most studies in the ﬁeld of heartbeat classiﬁcation focus

on individual heartbeats and use an intra-patient paradigm.

In the scheme, the heartbeats of the same patient used in

both training and testing subsets make the evaluation result

overly optimistic [25]. Moreover, this scheme does not take

into account the time variation and unique individual char-

acteristics of the ECG. Our architecture takes these aspects

into consideration and operates as an inter-patient paradigm

rather than an intra-patient paradigm.

2.1 Identiﬁed stream network

The ECG signal contains unique individual characteristics

and disease symptoms. The identiﬁed stream network op-

erates at an individual heartbeat, efﬁciently extracting the

identity features and static features of the heartbeat. The

appearance of the individual heartbeat (static characteristics

of the heartbeat) is a useful clue as many diseases can

be identiﬁed from a single heartbeat without the holistic

ECG signal. The classiﬁcation of individual heartbeats can

be achieved by the identiﬁed stream network designed.

And the classiﬁcation of individual heartbeats is pretty

competitive for some speciﬁc classes. On the other hand,

the symptoms vary among patients, identity features matter.

This network can be pre-trained for the purpose of identiﬁ-

cation in order to realize the identiﬁcation capability on its

own.

The identiﬁed stream network is constructed by an 11-

layer 1D convolution neural network [26]. It consists of

seven alternating convolutions and average-pooling layers,

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1Two-streamNetworkforECGSignalClassicationXinyaoHou,ShengmeiQin,JianboSu,SeniorMember,IEEEFAbstractElectrocardiogram(ECG),atechniqueformedicalmonitor-ingofcardiacactivity,isanimportantmethodforidentifyingcardiovas-culardisease.However,analyzingtheincreasingquantityofECGdataconsumesalotofmedicalres...

展开>> 收起<<

1 Two-stream Network for ECG Signal Classiﬁcation.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Two-stream Network for ECG Signal Classiﬁcation

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: