Deep Learning for Data Analytics: Foundations, Biomedical Applications, and Challenges

By Nilanjan Dey

Deep learning, a branch of Artificial Intelligence and machine learning, has led to new approaches to solving problems in a variety of domains including data science, data analytics and biomedical engineering. Deep Learning for Data Analytics: Foundations, Biomedical Applications and Challenges provides readers with a focused approach for the design and implementation of deep learning concepts using data analytics techniques in large scale environments. Deep learning algorithms are based on artificial neural network models to cascade multiple layers of nonlinear processing, which aids in feature extraction and learning in supervised and unsupervised ways, including classification and pattern analysis. Deep learning transforms data through a cascade of layers, helping systems analyze and process complex data sets. Deep learning algorithms extract high level complex data and process these complex sets to relatively simpler ideas formulated in the preceding level of the hierarchy. The authors of this book focus on suitable data analytics methods to solve complex real world problems such as medical image recognition, biomedical engineering, and object tracking using deep learning methodologies. The book provides a pragmatic direction for researchers who wish to analyze large volumes of data for business, engineering, and biomedical applications. Deep learning architectures including deep neural networks, recurrent neural networks, and deep belief networks can be used to help resolve problems in applications such as natural language processing, speech recognition, computer vision, bioinoformatics, audio recognition, drug design, and medical image analysis.

• Presents the latest advances in Deep Learning for data analytics and biomedical engineering applications.
• Discusses Deep Learning techniques as they are being applied in the real world of biomedical engineering and data science, including Deep Learning networks, deep feature learning, deep learning toolboxes, performance evaluation, Deep Learning optimization, deep auto-encoders, and deep neural networks
• Provides readers with an introduction to Deep Learning, along with coverage of deep belief networks, convolutional neural networks, Restricted Boltzmann Machines, data analytics basics, enterprise data science, predictive analysis, optimization for Deep Learning, and feature selection using Deep Learning

• Language: English
• Publisher: Academic Press
• Release date: May 29, 2020
• ISBN: 9780128226087

Book preview

Canada

Preface

Himansu Das¹, Chittaranjan Pradhan¹ and Nilanjan Dey², ¹KIIT Deemed to be University, Bhubaneswar, India, ²Techno International New Town (Formerly known as Techno India College of Technology), Kolkata, India

In recent years, deep learning methods have rapidly advanced in their technological development and their practical applications in different diversified research areas. It provides the maximum utilization of computing power of the GPUs to provide better performance. Deep learning encompasses different computational models of several processing layers to gain knowledge and represents the data with numerous levels of abstraction. Deep learning provides more flexible and better performance due to recent advancements in learning algorithms for deep architecture. Deep learning has an edge in contemporary use due to its diversity of applications in different research areas. These include domain-specific information such as medical image recognition, biomedicine, and object tracking that contains useful information about the problems. Different companies analyze these large volumes of data for their business prospects and also to use in decision-making processes that impact the existing and future technologies. Deep learning algorithms extract the high level of complex ideas and process these complex ideas to relatively simpler ideas formulated in the preceding level in the hierarchy.

The main advantage of deep learning is it allows the users to analyze and learn a large volume of data, making it a valuable tool for data science where raw data is largely unlabeled and uncategorized. Specifically, its use in data science and engineering applications is in high demand. Data science deals with the data mining and data analytics approaches of the supervised and unsupervised data sets in different applications. Mostly, it is used for prediction, data analysis, and data visualizations. Similarly, almost all the engineering applications including biomedical engineering, aerospace engineering, thermal engineering, communication engineering, and more use the deep learning approaches for the analysis and visualization purposes. The interest of the research community in deep learning methods is growing fast, and much architecture has been proposed in recent years to address several problems, often with an outstanding performance.

This book will focus on advanced models, architectures, and algorithms in deep learning for data science and engineering applications. Hence, it is expected that the development of deep learning theories and applications would further influence the field of data science and its different application domains.

In Chapter 1, the authors have been proposed a modified preprocessing and unique classification technique based on deep learning for the Electrocardiogram signal. Chapter 2 aims to reduce the computational complexity of the deep learning architecture for cardiac disease classification by using the feature extracted data. The aim of Chapter 3 is to provide an overview of how the training time and generalization capability for deep learning algorithms including deep belief networks and extreme learning machine autoencoder kernels. In Chapter 4, the authors present a brief review of the diagnosis of Alzheimer’s disease based on magnetic resonance imaging analysis using deep learning. They also present a deep architecture for early diagnosis of Alzheimer’s disease based on implicit feature extraction for the classification of magnetic resonance images. This model aims to classify Alzheimer’s disease patients against a group of patients without the disease. In Chapter 5, the effectiveness of various CNN-based pre-trained models for the detection of abnormalities in radiographic images has experimented and their performances are compared using standard statistical measures. In Chapter 6, the authors introduced the deep-wavelet neural network (DWNN) as a feature extraction method for image representation. DWNN is a deep learning tool based on the Mallat algorithm for wavelet decomposition at multiple levels. The authors applied the DWNN to the task of breast lesion detection and classification in thermographic images. In Chapter 7, a novel way of classifying the existing information retrieval models are introduced, along with their recent improvements and developments. The approach is the first one to classify the existing work according to how they generate the features and ranking functions. In Chapter 8, the authors proposed a new approach: the use of Autoencoders, a deep neural network with unsupervised training, to work like an intelligent filter, denoising the electrical potentials data. In Chapter 9, the authors explored architecture for evaluating the accuracy in disease classification using the PlantVillage dataset. Also, an overview of the deep learning architecture with basic building layers are briefly discussed. The SqueezeNet resulted in the best classification accuracy of 98.49% with original color images.

Chapter one

Short and noisy electrocardiogram classification based on deep learning

Sinam Ajitkumar Singh¹ and Swanirbhar Majumder²,    ¹Department of ECE, NERIST, Nirjuli, India,    ²Department of IT, Tripura University, Agartala, India

Abstract

Electrocardiogram (ECG) contains valuable data that assist in the initial investigation of cardiovascular diseases. Hence, the study of such electrical signals becomes a beneficial issue for many researchers. In this chapter, we shall propose a modified preprocessing and unique classification technique based on deep learning. A set of modified preprocessing steps has been implemented with the delineation of ECG signals using the wavelet transform (WT) followed by elimination of noise based on the Pan and Tompkins algorithm. Preprocessed signals have been converted to scalogram images based on continuous wavelet transform (CWT). Finally, a unique approach using deep learning algorithm for classification of the preprocessed scalogram images has been proposed here. The proposed model in this chapter shall be analytically verified using publicly available data sets A of PhysioNet 2016/ CinC challenge. The results show that deep learning based on a convolutional neural network (CNN) efficiently can be used for predicting the cardiovascular anomalies. The chapter begins with a discussion on short and noisy ECG classification and its importance with a brief overview on ECG signal processing. This is followed up with a basic literature survey and analysis of the deep learning technique used in this particular area as well as a general discussion of deep learning in the field of cardiological signals. This chapter also introduces a novel approach based on decision fusion for predicting the heart abnormality and compares the validated results with other existing methods.

Keywords

ECG; PCG; wavelet transform; CNN; scalogram; deep learning

1.1 Introduction

Cardiac abnormalities are the signs of disorder of the heart. Abnormalities include arrhythmias, coronary artery disease, mitral valve prolapse, and congenital heart disease. The study of heart characteristics is one of the necessary measures in assessing the cardiovascular system. Electrocardiogram (ECG) and phonocardiogram (PCG) are two existing problems in which the former produces due to the electrical movements of the heart while the latter produces due to the routine motion of the heart sounds. In the available literature, several researchers have employed different approaches that assist in evaluating the morphological characteristics of the ECG signals; they signify various cardiovascular abnormalities by analyzing the ECG [1–4]. The different approaches include support vector machines (SVM) [5], multilayer perceptron (MLP) [6], learning vector quantization (LVQ) [7], high order statistic [8], and K nearest neighbors (KNN) [9].

Arrhythmia is the most common of the diseases associated with heart abnormalities. As a result, most of the literature has dealt with arrhythmia classification [1,2,5,7–9]. The most efficient approach for predicting arrhythmia is the exploration of ECG signals [10]. The study of specific characteristics of ECG recording like beats, morphological and statistical features gives meaningfully correlated clinical data that further helps in predicting ECG pattern. Automated ECG classification is a complicated task as the features associated with morphological and temporal characteristics differ for different subjects under various conditions. Diagnosis of cardiovascular abnormalities using ECG recording has a definite drawback as the ECG signal varies from person to person, and an abnormal ECG signal has different morphological characteristics for related disorders. However, two distinct sets of disorders may have a similar characteristic on an ECG signal. These can cause a problem in the diagnosis of heart abnormalities using the ECG signal [10–12]. The anomalies of the heartbeat have to be detected after carefully analyzing the ECG signal. Consequently, the steps for analyzing the ECG signal, which mainly include wearable healthcare devices and bedside monitoring, take longer duration and involve a complicated procedure.

Many researchers have applied wavelet transform for the preprocessing and extraction of features. Yu and Chen [13] used statistical features by decomposing the recorded ECG signal based on discrete wavelet transform (DWT) and classified six classes based on probabilistic neural network (PNN). Thomas et al. [14] used dual-tree complex wavelet transform (DTCWT) for extracting the features from the QRS complex and had classified the features with an artificial neural network based on multilayer back propagation. Kumari and Devi [15] obtained a morphological feature vector based on the wavelet coefficients and independent component analysis (ICA) followed by arrhythmia classification using an SVM classifier.

As per available literature on the research, some of the researchers have used single-lead ECG recordings for detecting sleep apnea based on PhysioNet 2000 challenge. Based on the database provided, some of the researchers had employed a decision fusion method that helps in achieving a high classification performance. For instance, Li et al. [16] applied the decision fusion method by combining two binary classifiers [SVM and artificial neural network (ANN)] and achieved a classification accuracy of 85% in per-minute segments and 100% in per-recording segments for detecting obstructive sleep apnea (OSA) using a single-lead ECG recording. The authors in Ref. [17] analyzed OSA detection by comparing the CNN-based deep learning model with a decision fusion model and reported that the CNN model based on a deep learning algorithm was found to be superior in performance to the decision fusion model. Using the above approaches, a deep learning algorithm based on the transfer learning approach has been employed for predicting heart abnormalities from an ECG recording.

Hence, this chapter proposes an alternative approach for classifying the heart abnormalities using an ECG recording based on a deep learning algorithm that helps to improve the performance of the model. The preprocessing steps have been implemented by applying wavelet transform followed by noise removal algorithm. The filtered ECG signal has been further preprocessed by applying a continuous wavelet transform (CWT) that results in the conversion of the data to 2D scalogram images. The scalogram images are used to train the different convolutional neural networks (CNNs) based on deep learning for the classification. In this chapter, the detailed information about the basic idea correlated to the cardiovascular system and ECG signal has been discussed in Section 1.2. Section 1.3 deals with the study of various approaches associated with the heart abnormality analysis and details about the database. The study of the proposed method of using deep learning is discussed in Section 1.4, followed by a result and discussion in Section 1.5.

1.2 Basic concepts

1.2.1 Cardiac cycle

The heart, as we know, is a muscular organ found in the thoracic cavity. It has four chambers with the purpose of delivering oxygen to tissues. The left ventricle has a distinct, thick muscular wall as it helps to pump the blood across the body and into the circulation system, whereas the right ventricle has a thin muscular wall as it helps to pump the blood across the lungs. The cardiac cycle is split into four different periods: two relaxation phases and two contraction phases. The oxygenated blood moves through the left atria, past the mitral valves, and into the left ventricles during the ventricular filling phase. The atrial contraction (recorded as a P wave) starts during the firing of the S-A node and results in the filling of ventricles. During the starting period of ventricular contraction, the pressure is generated across the ventricles. This phase is the second cardiac phase, also known as the isovolumetric contraction period. Blood is discharged from the heart only when the ventricular pressure surpasses the aortic pressure. Blood will flow across the semilunar valves until the pressure gradient in the arteries surpasses the contracting ventricles pressure during the third phase (ventricular ejection period). The QRS complex wave represents an electrical activity for both the isovolumetric contraction period and the ventricular ejection period. The final phase is also known as an isovolumetric relaxation period that is marked as the resting phase during ventricular repolarization. In the ECG signal, an isovolumetric relaxation period is represented by the T wave.

1.2.2 Electrocardiogram

An ECG signifies the electrical impulse generated due to cardiac activities. The ECG signal carries valuable information for the cardiologist to achieve a comprehensive analysis of the subject. ECG is the typical conventional tool concerning the diagnosis of cardiac abnormality. The cardiologist extracts the ECG recording by placing the electrode to the patient’s body. The most popular tools for recording the ECG signal are the Holter machine. A cardiologist employs a Holter device on the cardiac patient that requires consistent monitoring to determine the abnormal heartbeat for a day. The beat of the cardiac cycle may be easily computed by computing the waves. The sinoatrial (SA) node is a collection of cells found at the right atrium. It helps to generate electrical impulses that regulate the flow of blood in the body.

1.2.3 The QRS wave

Due to the limited information provided by the PhysioNet 2016 challenge about the ECG database collection, this chapter centers on characterizing the features based on the QRS complex. The QRS complex indicates the ventricular depolarization [18] of the heart. The physical characteristic of the QRS wave has been expressed by the sequence of ventricular contractions. Ventricular contraction can be interpreted in two steps: septum contraction followed by contraction of the ventricular wall since the Purkinje fibers are positioned just underneath the endocardium, activation further expanded to the epicardium. Ventricular contraction occurs first at the septum. Routine septal contractions activate from the left to the right side. Hence this results in the generation of small septal R waves in the lead V1 and Q wave in lead V6. Contraction develops simultaneously both to the right and the left of the ventricular