Hierarchical Deep Learning with Generative Adversarial Network for Automatic Cardiac Diagnosis from ECG Signals

2025-05-06 0 0 8.46MB 28 页 10玖币

侵权投诉

Hierarchical Deep Learning with Generative Adversarial Network for

Automatic Cardiac Diagnosis from ECG Signals

Zekai Wang, Stavros Stavrakis, Bing Yao∗

Abstract

Cardiac disease is the leading cause of death in the US. Accurate heart disease detection is of critical

importance for timely medical treatment to save patients’ lives. Routine use of electrocardiogram

(ECG) is the most common method for physicians to assess the electrical activities of the heart and

detect possible abnormal cardiac conditions. Fully utilizing the ECG data for reliable heart disease

detection depends on developing eﬀective analytical models. In this paper, we propose a two-level

hierarchical deep learning framework with Generative Adversarial Network (GAN) for automatic

diagnosis of ECG signals. The ﬁrst-level model is composed of a Memory-Augmented Deep auto-

Encoder with GAN (MadeGAN), which aims to diﬀerentiate abnormal signals from normal ECGs

for anomaly detection. The second-level learning aims at robust multi-class classiﬁcation for diﬀer-

ent arrhythmias identiﬁcation, which is achieved by integrating the transfer learning technique to

transfer knowledge from the ﬁrst-level learning with the multi-branching architecture to handle the

data-lacking and imbalanced data issue. We evaluate the performance of the proposed framework

using real-world medical data from the MIT-BIH arrhythmia database. Experimental results show

that our proposed model outperforms existing methods that are commonly used in current practice.

Keywords: Deep learning, Hierarchical Model, Generative Adversarial Network, Multi-branching

Output

1. Introduction

Heart disease is the leading cause of death in the US. It aﬀects about 85.6 million people and

leads to more than $320 billion in annual medical costs [1]. It is of critical importance to develop

accurate and reliable heart disease diagnoses for timely medical treatments to save patients’ lives

∗Corresponding author: byao3@utk.edu;

Zekai Wang and Bing Yao are with the Department of Industrial & Systems Engineering, The University of Tennessee,

Knoxville, TN, 37996 USA.

Stavros Stavrakis is with University of Oklahoma Health Sciences Center, Oklahoma City, OK 73104 USA.

Preprint submitted to arXiv October 21, 2022

arXiv:2210.11408v1 [eess.SP] 19 Oct 2022

[2, 3]. The heart rhythm is generated by the excitation, propagation, and coordination of electrical

signals from the cardiac cells across diﬀerent heart chambers. A normal cardiac cycle starts with

the activation of the sinoatrial node, from where the cardiac electrodynamics spreads out through

the atria. The electrical wave then arrives at the atrio-ventricular node and propagates through

the bundle of His toward Purkinje ﬁbers, leading to the electrical depolarization and repolarization

of the ventricles to complete the cycle. The resulting electrical signals on the body surface are

described by the electrocardiogram (ECG), which consists of a P-wave, QRS-complex, and T-wave

[4]. Changes in electrophysiological properties will vary the propagation pattern of electrodynamics

and lead to diﬀerent types of conduction abnormalities and/or cardiac arrhythmias manifested in

the variation of ECG waveform patterns [5, 6].

In recent years, rapid advancements in wearable sensing and information technology facilitate

the eﬀective monitoring of patients’ heart health conditions [7, 8, 9, 10, 11, 12, 13, 14]. Routine

use of ECG is the most common method for physicians in everyday clinical practice to assess the

electrical activities of the heart and detect possible abnormal cardiac conditions. Physicians gener-

ally identify the cardiac arrhythmia by checking the ECG waveforms with naked eyes. This can be

time-consuming and may require extensive human resources. Additionally, ECG misinterpretation

may happen especially when there exists a large amount of data to inspect, leading to possible

misdiagnosis of fatal heart disease [15]. Auto arrhythmia detection based on machine learning

algorithms can provide important assistance to physicians [16]. However, although ECG signals

contain rich information associated with the electrophysiological condition of the heart, the research

on fully utilizing ECGs for reliable data-driven disease detection poses several challenges including

(1) Nonlinear and nonstationary dynamics: Real-world cardiovascular systems are fea-

tured with nonlinear and nonstationary dynamics from the complicated interactions of many inter-

connected parts such as ion channels and gap junctions to perform cardiac functions, generating

ECG signals with nonlinear waveforms. Traditional statistical and machine learning methods de-

pend heavily on manual feature engineering of such waveform data, which generally consists of

two stages [17]: human experts extract useful features from raw ECGs at the ﬁrst stage and then

employ machine learning algorithms on the handcrafted features to generate predictive results at

the second stage. However, this procedure is restricted by the data quality and human expert

knowledge [18], and may result in information loss, which lacks the potential for real clinical im-

plementation. Thus, new algorithms that are able to eﬀectively and automatically extract useful

features are urgently needed for reliable heart disease identiﬁcation.

(2) Lack of training labels and imbalanced data issue: Most existing data-driven models

for ECG analysis are achieved through supervised learning, which requires a large volume of anno-

tated ECG cycles (with diagnostic labels such as normal, abnormal, or speciﬁc types of arrhythmia).

However, the annotation process requires cardiologists to manually inspect the ECG signals and

assign a label to each diﬀerent pattern, which is time-consuming and labor-intensive. Additionally,

it is impractical to collect enough data for each type of disease-altered signals in order to meet

the requirement for suﬃcient supervised training. This is due to the fact that data associated

with abnormal heart conditions is signiﬁcantly less than data from healthy people. Moreover, the

occurrence rate of diﬀerent arrhythmias is highly diverse. Data-driven predictive modeling based

on such imbalanced data tends to ignore the minority classes, leading to unsatisfactory detection

performance. As such, new methods that can eﬀectively model the ECGs and account for the

data-lacking and imbalanced data issues are needed for reliable disease identiﬁcation.

This paper proposes a hierarchical deep learning framework with Generative Adversarial Net-

work (GAN) to investigate ECG signals for automatic identiﬁcation of diﬀerent types of arrhyth-

mias. We ﬁrst propose a Memory-Augmented Deep auto-Encoder with Generative Adversarial

Network (MadeGAN) to achieve the ﬁrst-level anomaly detection (i.e., binary classiﬁcation for nor-

mal and abnormal signals). Second, we employ the transfer learning technique to transfer knowledge

learned from the ﬁrst-level training for second-level multi-class classiﬁcation to identify diﬀerent

types of arrhythmias. In addition, in the second-level network, we adapt the multi-branching archi-

tecture developed in our prior work [19] to solve the imbalanced data issue among diﬀerent types

of heart diseases. We evaluate our proposed hierarchical deep learning framework using the data

from the MIT-BIH arrhythmia database [20]. Experimental results show that our proposed method

signiﬁcantly outperforms existing approaches that are commonly used in current practice.

2. Research Background

A variety of statistical and machine learning algorithms have been developed for ECG data

analysis and pattern recognition [21]. For example, Yang et al [22] developed a dynamic spatiotem-

poral warping algorithm to measure dissimilarities between ECG signals and further employed the

spatial embedding to transform the warping dissimilarity matrix into feature vectors for myocardial

infarction classiﬁcation. Bertsimas et al [23] utilized the XGBoost algorithm to capture disease-

altered patterns in ECG cycles for heart disease prediction. Wavelet-based and recurrence analysis

approaches have also been widely implemented to learn waveform features for ECG classiﬁcation

[24, 25, 26]. Lyon et al [27] investigated the linear and quadratic discriminants, support vector

machine, random forest, and Bayesian network for heartbeat classiﬁcation from ECG signals. A

comprehensive review of statistical and machine learning methods in ECG detection can be found

in [17]. However, most existing traditional data-driven methods depend heavily on manual feature

engineering, which is a labor-intensive trial-and-error process and is generally limited by human

expert knowledge [18, 28].

Deep Neural Network (DNN) is another powerful tool that has achieved promising results in

the area of data-driven disease detection [29]. Unlike conventional statistical and machine learning

methods, the main advantage of DNNs is that they do not require explicit feature engineering.

Instead, feature extraction is automatically achieved by intermediate layers of the network. It

has been demonstrated that DNN-based features are more informative than handcrafted features

for arrhythmia detection [30, 31]. As such, a variety of DNN models including convolutional

neural networks (CNNs) [32] have been designed for arrhythmia detection and have outperformed

conventional statistical methods [33, 34]. For example, Hannun et al [35] employed 1D CNN to

classify 12 rhythm classes and achieved high performance that is comparable to the diagnostic

results provided by cardiology experts. Li et al [36] combined a 2D CNN and a distance matrix to

classify congestive heart failure. Shashikumar et al [37] developed an attention-based model with

a 2d CNN as the feature extractor and a bidirectional recurrent neural network to capture the

temporal pattern in ECG signals.

However, most existing deep learning algorithms for ECG analysis are based on supervised

learning, which requires a large volume of annotated ECG signals and also suﬀers from the problem

of extremely imbalanced data. Thus, the application of unsupervised and semi-supervised learning

in ECG analysis has been increasingly investigated. For example, Auto-Encoder (AE), a semi-

supervised deep learning technique, has been widely used to study ECG data by extracting critical

low-dimension representation of the raw signals for disease prediction [38, 39]. Furthermore, GAN-

based framework, another semi-supervised learning technique to capture inherent data distributions

[40, 41], has been applied in ECG analysis. For example, Zhou et al [42] developed a BeatGAN

structure to model ECG signals for anomaly detection. Wang et al [43] employed an auxiliary

classiﬁer GAN for data augmentation to handle the imbalanced issue. Shin et al [44] integrated

the AnoGAN framework [45] with a decision boundary-based model for ECG anomaly detection.

However, most existing semi-supervised deep learning methods mainly focus on diﬀerentiating the

abnormal ECGs from normal ones (i.e., binary classiﬁcation) and they are not able to perform

Encoder

Decoder

Discriminator

Real

Fake

Input ECG signal 𝒙Memory module Reconstructed signal: 𝒙"

𝒛 = 𝑓

&(𝒗)

𝒛* = 𝛀𝐓𝒘

Anomaly score:

𝒙 − 𝒙"

1st level MadeGAN

Feature

extractor

Normal Abnormal

Abnormal type

2nd level classification

Branching Output

(a)

(b)

Discriminator

Shallow

classifier

1D CNN

Output

(c)

Figure 1: The proposed two-level hierarchical deep learning framework: (a) ﬁrst-level MadeGAN for anomaly detec-

tion; (b) second-level classiﬁcation for arrhythmia type identiﬁcation; (c) Multi-branching output.

multi-class classiﬁcation to identify diﬀerent types of cardiac arrhythmia. Thus, novel analytical

models are urgently needed to eﬃciently handle the imbalanced data issue and the data lacking

problem for both robust anomaly detection and accurate disease identiﬁcation from ECG signals.

3. Research Methodology

As shown in Fig. 1, this section presents the proposed hierarchical deep learning framework

for automatic ECG diagnosis. We denote a single ECG cycle as x∈Rdx×1, where dxdenotes

the dimensionality of x. Each ECG cycle is associated with a multiclass label y. As such, each

training data point can be described by the tuple (x, y) with y= 0 indicating normal signal

and y= 1,2, . . . , M corresponding to other diﬀerent types of arrhythmias. Our objective is to

ﬁrst diﬀerentiate abnormal ECG signals from normal ones (i.e., ﬁrst-level anomaly detection) and

then classify the abnormal signals into diﬀerent types of arrhythmias (i.e., second-level multi-class

classiﬁcation). Speciﬁcally, we propose a Memory-Augmented Deep auto-Encoder with Generative

Adversarial Network (MadeGAN) to achieve the ﬁrst-level anomaly detection. The second-level

classiﬁcation network is constructed by integrating a shallow classiﬁer with the part of trained

discriminator from the ﬁrst-level learning (i.e., transfer learning to handle the data-lacking problem)

and a multi-branching layer (to handle the imbalanced data issue).

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：

HierarchicalDeepLearningwithGenerativeAdversarialNetworkforAutomaticCardiacDiagnosisfromECGSignalsZekaiWang,StavrosStavrakis,BingYaoAbstractCardiacdiseaseistheleadingcauseofdeathintheUS.Accurateheartdiseasedetectionisofcriticalimportancefortimelymedicaltreatmenttosavepatients'lives.Routineuseofelec...

展开>> 收起<<

Hierarchical Deep Learning with Generative Adversarial Network for Automatic Cardiac Diagnosis from ECG Signals.pdf

共28页,预览5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！

Hierarchical Deep Learning with Generative Adversarial Network for Automatic Cardiac Diagnosis from ECG Signals

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: