Arrhythmia detection model using modified DenseNet for comprehensible Grad-CAM visualization

doi:10.1016/j.bspc.2021.103408

Biomedical Signal Processing and Control

Volume 73, March 2022, 103408

https://doi.org/10.1016/j.bspc.2021.103408 Get rights and content

Highlights

•
Deep Learning was used to generate a model that detects arrhythmia using ECG.
•
The basis for judgement was difficult to understand in the basic model structure.
•
This study improve the visualization of Grad-CAM without compromising classification accuracy.
•
This study allows us to visualize irregular intervals or shapes of electrocardiogram.
•
An interpretable model will enable doctors to gain trust in medical deep learning.

Abstract

Diagnosing arrhythmia is difficult, requires significant efforts. Because arrhythmia can be associated with serious diseases, it is important to classify arrhythmia patients with high accuracy, and the basis for the classification model's judgment should be properly demonstrated. Traditional algorithm methods are less accurate, and simply using a high-accuracy image classification deep learning model yields incomprehensible results when the model is visualized with gradient-weighted class activation mapping (Grad-CAM). We want to achieve high-performance deep learning models can also comprehensible visualization. To obtain this, two hypotheses about Grad-CAM were established and the experiment was conducted. As a result, a method that could clearly visualize the response area using Grad-CAM with a higher classification performance of 0.98 accuracy is created.

Introduction

We propose a model structure that can generate a better visualization of arrhythmia classification using the electrocardiogram (ECG) data collected from actual patients (see Table 1).

Electrocardiogram (ECG) is data that record the electrical activity of the heart. By measuring the length of each section of the ECG, it is possible to verify that the electrical signals are transmitted at normal speeds from each part of the heart to the other part. Therefore, ECG data can be used to determine when and how the heart's muscles are activated. An arrhythmia means that the heartbeat of the heart is out of normal. Therefore, arrhythmia can be diagnosed using ECG [1], [2].

An accurate diagnosis of arrhythmia is crucial because it is most likely associated with a disease that can cause major problems in the body. An ECG is a record of the electrical activity in the heart. It is particularly important for diagnosing cardiac dysrhythmia and heart arrhythmia [1], [2].

A Holter's monitoring system is used to measure an ECG for at least 24 h. However, as the ECG data collected through this system increase with time, the amount of data becomes considerably large. Therefore, It takes a great deal of medical staff effort and time resources to read an one person’s electrocardiogram record [1], [2]. Furthermore, even for patients with actual arrhythmias, most of the collected ECG data are normal signals (sinus rhythm). As a result, a judgment error is likely to occur when attempting to detect arrhythmia between normal signals. Therefore, even an experienced specialist requires considerable time to analyze signals, and the accuracy is not high [3].

In this study, a densely connected convolutional network (DenseNet) [4] structure, which is a deep learning structure that shows excellent performance in image classification, is used to create a model for arrhythmia detection. This structure provided efficient arrhythmia detection and classification and minimized human intervention, unlike existing studies that require the assistance of experts in the field at the intermediate stages for characterization and extraction. A experiments is performed to observe the changes in classification performance of the model according to the filter size of DenseNet and found an optimal hyperparameter (see Table 2).

A Gradient-weighted class activation mapping (Grad-CAM) [5], which was developed for the visual interpretation of a convolutional neural network (CNN) model, is used to visualize the basis of model judgment for human confirmation. Because In the medical field, judgment errors are directly related to patients’ health. Therefore, it is difficult for medical staff to trust the deep learning model only because accuracy is high without knowing about the basis of judgment. This decreases the utilization of deep learning in actual clinical practice. To use deep learning models for detecting arrhythmia, it is necessary to inform decision makers about the decision basis of the models and which features of the data entered into the models influence results [6].

When attempting to classify ECG data using the image classification model, visualizations of the Grad-CAM highlighted areas are extremely ambiguous for humans to identify. In order to solve these problems, this study produced a hypothesis for the visualization of an effective Grad-CAM. A model comparison was performed to confirm the hypothesis, which led to a structure that can show higher performance and better visualization for ECG data. Based on this, the fact that the proposed model can be more helpful in actual clinical settings compared to the existing model is turned out. In related work, attention or general CAM is used. This requires major changes to the structure of the deep learning model. Or, to use the existing 2D network as it is, the image of the ECG is used as the input of the model. The contribution of our research is to create an ECG classification deep learning model that can best classify the ECG data acquired by researchers, and to improve the ability to interpret the basis of the judgments performed by the model without changing the main structure of the model.

The rest of this paper is organized as follows: Chapter 2 introduces previous studies on arrhythmia detection and classification and states our goal. Chapter 3 introduces the models and techniques used in this study. Chapter 4 presents datasets and preprocessing. Chapter 5 describes the learning process and its evaluation. Chapter 6 describes the experimental results, and Chapter 7 provides the conclusion and the direction of future work.

Section snippets

Literature review and goal

Various commercial programs with rule-based algorithms have been developed to help detect and diagnose arrhythmia. However, these programs are not reliable for use by practitioners because of their low judgment accuracy. There have been attempts to solve this problem by utilizing various machine learning techniques. However, the performance of these techniques is insufficient for practical use. Recently developed models based on deep learning have shown higher accuracy compared to even

Classifier model

We developed an ECG classification model using the DenseNet architecture. As DenseNet has a skip-connection structure, the information entered into the model can be passed to the end without loss. In addition, even when performing the back-propagation operation, the operation at the end of the model is transmitted to the front part without the gradient vanishing problem, unlike the other models. This model shows excellent performance by mitigating the problems of gradient vanishing. We found

Dataset

The data used in this study were the ECGs of approximately 52,000 patients collected from university hospitals in South Korea. The ECG data were obtained and labeled by Cardiologist of the university hospitals. ECGs are measured by attaching electrodes to various parts of the body, and various types of ECGs can be obtained according to the positions of the electrodes. Lead II ECG data collected from the university hospitals were used. The data consisted of 12 categories: normal (sinus), atrial

Training & evaluation metrics

For model training and verification, 20% of 52,043 data samples were separated into test data; further, 80% of the remaining samples were used as training data and 20% as verification data. Therefore, 64% of the total data were used as the training data, 16% as the verification data, and the remaining 20% as the test data. The number of training, verification, and test data samples was 33,308, 8,327, and 10,408, respectively. Each data sample was randomly selected. For a fair model comparison,

Comparison of classification performance

Table 3 compares the performance of each classification model for test data. The actual classification of test data for each model can be found in the confusion matrix in Table 4. In Table 3, AlexNet for ECG [23] had a relatively low accuracy (Eq. (3)) value, while all other models had a high accuracy of 0.98 or higher and did not show much difference. The sensitivity (Eq. (2)) of Table 3 is the result of how abnormal the abnormalities of the arrhythmia were classified and is a particularly

Conclusion and discussion

ECG data were used to generate a high-performance model for determining the presence of heart arrhythmia. In the medical field, when diagnosing mistakes are made, the severity and cost required to rectify these mistakes are high [2]. In order to solve this problem, we collected about 52,000 ECG data from university hospitals in Korea and created a model to classify the ECG. The proposed model is highly accurate, and it may contribute to reducing the risk of decision making [3]. This study used

CRediT authorship contribution statement

Jin-Kook Kim: Conceptualization, Investigation, Methodology, Writing - original draft. Sunghoon Jung: Conceptualization. Jinwon Park: Visualization. Sung Won Han: Conceptualization, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This work was supported by Korea University Grant (K1915041, K1920081). This research was also supported by National Research Foundation of Korea (NRF-2019R1F1A1060250).

References (23)

Y. Hagiwara et al.
Computer-aided diagnosis of atrial fibrillation based on ECG signals: a review
Inf. Sci.
(2018)
C. Antzelevitch et al.
Overview of basic mechanisms of cardiac arrhythmia
Cardiac Electrophysiol. Clin.
(2011)
R.J. Martis et al.
Current methods in electrocardiogram characterization
Comput. Biol. Med.
(2014)
S. Pal et al.
Empirical mode decomposition based ECG enhancement and QRS detection
Comput. Biol. Med.
(2012)
U.R. Acharya et al.
Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network
Inf. Sci.
(2017)
Y. Xia et al.
Detecting atrial fibrillation by deep convolutional neural networks
Comput. Biol. Med.
(2018)
A. Isin et al.
Cardiac arrhythmia detection using deep learning
Procedia Comput. Sci.
(2017)
P. Rajpurkar, A.Y. Hannun, M. Haghpanahi, C. Bourn, A.Y. Ng, Cardiologist-level arrhythmia detection with convolutional...
G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected convolutional networks. In: Proc. 2017 IEEE...
R.R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-cam: Visual explanations from deep networks...

A. Holzinger, C. Biemann, C.S. Pattichis, D.B. Kell, What do we need to build explainable AI systems for the medical...

Cited by (22)

A comprehensive review on efficient artificial intelligence models for classification of abnormal cardiac rhythms using electrocardiograms
2024, Heliyon
Deep learning has made many advances in data classification using electrocardiogram (ECG) waveforms. Over the past decade, data science research has focused on developing artificial intelligence (AI) based models that can analyze ECG waveforms to identify and classify abnormal cardiac rhythms accurately. However, the primary drawback of the current AI models is that most of these models are heavy, computationally intensive, and inefficient in terms of cost for real-time implementation. In this review, we first discuss the current state-of-the-art AI models utilized for ECG-based cardiac rhythm classification. Next, we present some of the upcoming modeling methodologies which have the potential to perform real-time implementation of AI-based heart rhythm diagnosis. These models hold significant promise in being lightweight and computationally efficient without compromising the accuracy. Contemporary models predominantly utilize 12-lead ECG for cardiac rhythm classification and cardiovascular status prediction, increasing the computational burden and making real-time implementation challenging. We also summarize research studies evaluating the potential of efficient data setups to reduce the number of ECG leads without affecting classification accuracy. Lastly, we present future perspectives on AI's utility in precision medicine by providing opportunities for accurate prediction and diagnostics of cardiovascular status in patients.
Toward explainable artificial intelligence: A survey and overview on their intrinsic properties
2024, Neurocomputing
Artificial intelligence and its derivative technologies are not only playing a role in the fields of medicine, economy, policing, transportation, and natural science computing today but also in future industries such as electric vehicles and meta-universe. However, because of the black-box nature of most common artificial neural networks (DNNs), there needs to be more understanding of what is happening behind these astounding performances. The abstract reasoning process in the networks raises concerns about the security of AI systems. Because of this, more and more researchers are turning their attention to the explainability of black-box neural networks to find some attributions that explain the reasoning performed by the networks by studying the black-box nature of the networks. Deep neural networks and their explainability act as a mutually beneficial symbiosis, facilitating each other’s development. This survey reviews the deep network explainable methods applicable for the field of computing vision proposed within the last decade and categorizes these methods in terms of their starting point to explain deep neural networks. Focusing on their intrinsics, we review the methods that are applicable for object classification, object detection, and looked forward that of object tracking. In each cluster, we show that the methods share some similarities but have their own highlight. Furthermore, we shed light on the future development of the field of explainable AI.
Interpretation of lung disease classification with light attention connected module
2023, Biomedical Signal Processing and Control
Lung diseases lead to complications from obstructive diseases, and the COVID-19 pandemic has increased lung disease-related deaths. Medical practitioners use stethoscopes to diagnose lung disease. However, an artificial intelligence model capable of objective judgment is required since the experience and diagnosis of respiratory sounds differ. Therefore, in this study, we propose a lung disease classification model that uses an attention module and deep learning. Respiratory sounds were extracted using log-Mel spectrogram MFCC. Normal and five types of adventitious sounds were effectively classified by improving VGGish and adding a light attention connected module to which the efficient channel attention module (ECA-Net) was applied. The performance of the model was evaluated for accuracy, precision, sensitivity, specificity, f1-score, and balanced accuracy, which were 92.56%, 92.81%, 92.22%, 98.50%, 92.29%, and 95.4%, respectively. We confirmed high performance according to the attention effect. The classification causes of lung diseases were analyzed using gradient-weighted class activation mapping (Grad-CAM), and the performances of their models were compared using open lung sounds measured using a Littmann 3200 stethoscope. The experts’ opinions were also included. Our results will contribute to the early diagnosis and interpretation of diseases in patients with lung disease by utilizing algorithms in smart medical stethoscopes.
Machine learning in metastatic cancer research: Potentials, possibilities, and prospects
2023, Computational and Structural Biotechnology Journal
Cancer has received extensive recognition for its high mortality rate, with metastatic cancer being the top cause of cancer-related deaths. Metastatic cancer involves the spread of the primary tumor to other body organs. As much as the early detection of cancer is essential, the timely detection of metastasis, the identification of biomarkers, and treatment choice are valuable for improving the quality of life for metastatic cancer patients. This study reviews the existing studies on classical machine learning (ML) and deep learning (DL) in metastatic cancer research. Since the majority of metastatic cancer research data are collected in the formats of PET/CT and MRI image data, deep learning techniques are heavily involved. However, its black-box nature and expensive computational cost are notable concerns. Furthermore, existing models could be overestimated for their generality due to the non-diverse population in clinical trial datasets. Therefore, research gaps are itemized; follow-up studies should be carried out on metastatic cancer using machine learning and deep learning tools with data in a symmetric manner.
A deep learning refinement strategy based on efficient channel attention for atrial fibrillation and atrial flutter signals identification
2022, Applied Soft Computing
Citation Excerpt :
Arrhythmia is currently considered as an essential group of cardiovascular diseases, which has endangered human health and received considerable attention [1]. Clinically, atrial fibrillation (AF) and atrial flutter (AFL) are the most frequent arrhythmias with an estimated prevalence of 1.5% to 2% and 0.09% in the individual worldwide [2–5]. More than half of AFL patients often have AF, both have similar physiological features and are also closely associated with various cardiovascular diseases such as stroke and myocardial infarction [6,7].
Atrial fibrillation (AF) and atrial flutter (AFL) are the most frequent arrhythmias recently. However, given both similar physiological features, visually evaluating electrocardiogram as the most traditional diagnosis scheme is taxing and error-prone. In this work, we specifically design two network modules based on bidirectional long short term memory (BiLSTM) and gate recurrent unit (BiGRU) for automatic AF and AFL detection. Motivated from Efficient Channel Attention network, we aim to reformulate BiLSTM and BiGRU with a feature recalibration approach that enables the model to adaptively focus more on the relevant feature representations and suppress irrelevant parts while appropriately capturing cross-channel interaction for learning effective channel attention. The results lead to consistent performance gains than several published researches with an accuracy of 99.2% and 99.3% across the two publicly available data sets while demonstrating the effectiveness of both modules. In particular, various derivative gradient values of sample ECG segments are visualized to improve interpretability. To our knowledge, this work offers the first empirical investigation of existing BiLSTM and BiGRU refinements for a better performance gain, showing great potential for many computer vision tasks.
A novel myocardial infarction localization method using multi-branch DenseNet and spatial matching-based active semi-supervised learning
2022, Information Sciences
Citation Excerpt :
The proposed multi-branch densely connected convolutional network (MB-DenseNet) can automatically extract ECG heartbeat features from different leads. Although there are several deep network architectures, we apply DenseNet to the multi-branch architecture since its efficiency in detection and recognition of features has been shown in different recent works [12,27–29]. Moreover, to solve the individual differences, we propose a novel active semi-supervised learning (ASSL) strategy to update the model.
Individual differences among patients and the high cost of manual labeling are major challenges for electrocardiogram (ECG) diagnosis algorithms. To tackle this problem, we develop a novel deep active semi-supervised learning framework for myocardial infarction localization based on 12-lead ECG signals. First, a new deep learning model named multi-branch densely connected convolutional network (MB-DenseNet) is designed to automatically extract and fuse heartbeat features from 12-lead ECG signals. Moreover, to improve the classification results for new patients, we propose a novel active semi-supervised learning (ASSL) mechanism to update the model. Active learning (AL) is employed to improve the classification ability of the initial model firstly, and then a new semi-supervised learning method named self-training with spatial matching (STSM) is designed to update the model further. Based on the artificial knowledge from AL, STSM combines spatial matching algorithm and the trained model to label valuable unlabeled samples automatically. We conduct experiments based on intra-patient and patient-specific schemes using the PTB database. The MB-DenseNet yields an accuracy of 99.87% under the intra-patient scheme. For the patient-specific scheme, the updated model achieves an accuracy of 96.09%. Compared with state-of-the-art methods, our method can effectively reduce manual labeling while achieving comparable performance.

View all citing articles on Scopus

View full text

Arrhythmia detection model using modified DenseNet for comprehensible Grad-CAM visualization

Highlights

Abstract

Introduction

Section snippets

Literature review and goal

Classifier model

Dataset

Training & evaluation metrics

Comparison of classification performance

Conclusion and discussion

CRediT authorship contribution statement

Declaration of Competing Interest

Acknowledgements

Inf. Sci.

Cardiac Electrophysiol. Clin.

Comput. Biol. Med.

Comput. Biol. Med.

Inf. Sci.

Comput. Biol. Med.

Procedia Comput. Sci.