Wearable Sensor-Based Edge Computing Framework for Cardiac Arrhythmia Detection and Acute Stroke Prediction

Abstract

Internet of Things-based smart healthcare systems have gained attention in recent years for improving healthcare services and reducing data management costs. However, there is a requirement for improving the smart healthcare system in terms of speed, accuracy, and cost. An intelligent and secure edge-computing framework with wearable devices and sensors is proposed for cardiac arrhythmia detection and acute stroke prediction. Latency reduction is highly essential in real-time continuous assessment, and classification accuracy has to be improved for acute stroke prediction. In this paper, preprocessing and deep learning-based assessment is performed in the edge-computing layer, and decisions are communicated instantly to the individuals. In this work, acute stroke prediction is performed by a deep learning model using heart rate variability features and physiological data. Classification accuracy is improved in this approach when compared to other machine learning approaches. Cloud servers are utilized for storing the healthcare data of individuals for further analysis. Analyzed data from these servers are shared with hospitals, healthcare centers, family members, and physicians. The proposed edge computing with wearable sensors approach outperforms existing smart healthcare-based approaches in terms of execution speed, latency time, and power consumption. The deep learning method combined with DWT performs better than other similar approaches in the assessment of cardiac arrhythmia and acute stroke prediction. The proposed classifier achieves a sensitivity of 99.4%, specificity of 99.1%, and accuracy of 99.3% when compared with other similar approaches.

This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. Introduction

Edge of Things- (EoT-) based healthcare gained huge attention in the past few years due to the wide usage of wearable devices, 5G communication, and edge computing framework. Wireless healthcare monitoring systems have gained huge attention recently in almost all biomedical applications. Internet of Things- (IoT-) based wearable computing devices generate a huge volume of data that has to be stored and analyzed to get important information in many real-world applications [1]. Though IoT-based healthcare applications are highly essential for accurate and timely medical diagnosis, there are many challenges in the appropriate data collection, required computing facility, and data storage [2].

In most modern data analytics applications, machine learning algorithms are applied to improve the accuracy and efficiency of the systems. In a few applications, these intelligent algorithms even perform the detection of errors and system output. Machine learning algorithms are widely popular in the classification of data into different categories that are highly essential in analyzing biomedical data [3]. Biomedical signal monitoring and processing are highly essential in almost all medical devices and advanced equipment. In recent years, wearable devices are preferred in many healthcare monitoring to provide continuous monitoring, ease of access, flexibility, and disease detection. Wearable devices and sensors are popular recently in the wireless monitoring of infants, elderly persons, differently abled persons, and athletes. Moreover, the increase in the aging population needs special attention and continuous monitoring of elderly people in the age group above 60 [4]. These devices generate a large amount of data that have to be stored and processed. However, mobile devices are not able to store a huge volume of data, and they are not capable of performing complex computations involved in signal processing. To overcome these issues, cloud-based wireless healthcare systems gained huge attention in recent years.

IoT- and cloud computing-based healthcare systems are developed in wireless healthcare monitoring to handle huge volumes of data with variety. The main aim of IoT-based healthcare is to reduce the computing burden and storage requirements of mobile devices and to provide low-cost solutions for healthy living for users [5]. In this paper, an edge-based healthcare system is developed to assess important health parameters like blood pressure, temperature, respiratory rate, electrocardiogram (ECG), and heart rate. Healthcare ECG monitoring is useful in assessing the health condition of the heart.

EoT gained huge popularity in wireless healthcare monitoring to handle huge volumes of data with variety. The main objective of this work is to reduce the computing burden and storage requirements of mobile devices and to provide low-cost solutions with healthy life for the users [6]. In this article, an edge-based healthcare system is developed to assess important health parameters like blood pressure, temperature, respiratory rate, electrocardiogram (ECG), and heart rate. Healthcare ECG monitoring is useful in assessing the health condition of the heart.

In wireless monitoring, each patient is considered a node of the wireless sensor network, and these nodes are connected to the central node at the hospital. Wireless networking, mobile computing, and cluster computing solutions have been utilized in the performance improvement of healthcare networks. Pervasive computing and wireless sensor networks are used for different possible ways of sending data in medical applications. In that work, the authors used various wireless technologies in the biomedical domain for monitoring physiological signals. Many sensors have been utilized for signal acquisition, and signal processors are used for preprocessing the acquired ECG signals. In this work, the healthcare system is improvised with the application of a cloud framework in healthcare monitoring.

In wireless healthcare monitoring, sensors are linked with various hardware and software components used for the effective monitoring of patients. Figure 1 shows the simple wearable computing framework with a cloud or edge computing layer. The collected data from wearable devices are preprocessed to provide relevant data to the computing layer. Preprocessed data are assessed in the computing layer using edge computing devices, and the results are sent back to the end user. To perform in-depth analysis and prediction, various machine learning and deep learning algorithms may be utilized for feature extraction and classification.

[figure(s) omitted; refer to PDF]

Wireless healthcare monitoring is highly essential in assessing chronic diseases such as cardiac arrhythmias, diabetes, hypertension, and cardiovascular disease. Since wireless healthcare monitoring faces some serious challenges, wearable sensors with an edge computing approach are presented in this paper. Edge computing-based smart healthcare systems are developed by researchers using wearable devices [7–9]. These approaches focus on the physiological parameters and history of patients in the assessment of individuals. A few issues in the existing works are as follows: (i) wearable sensors and devices are not capable of performing prediction tasks and critical analysis. (ii) Cloud computing frameworks are utilized in existing approaches for deep learning-based detection and prediction tasks that require intensive computing. (iii) The classification accuracy needs to be improved in life-threatening heart disease risk analysis and acute stroke prediction.

A deep learning-based CNN model is developed to assess heart disease in a fog environment [10]. However, there is a requirement for improving accuracy in cardiac arrhythmia detection and acute stroke prediction. An edge computing framework with wearable devices is presented in this paper to overcome the drawbacks of existing edge computing-based approaches and to improve the accuracy of cardiac arrhythmia detection. The simple preprocessing tasks are carried out on the user’s mobile device, whereas feature extraction and deep learning-based assessment are carried out on edge devices. A cloud server is used to store and analyze a large amount of collected data. Healthcare centers, emergency services, patients, family members, and physicians are communicated through the cloud server.

In the existing works, cardiac arrhythmia detection has been carried out with different methodologies and various risk factors. A few of the risk factors are type-II diabetes, high blood pressure, tobacco use, and high cholesterol. Feature selection techniques are highly required for improving accuracy and effective utilization of computing resources in the assessment of heart diseases using machine learning approaches. An artificial neural network (ANN) with multilayer perceptron is utilized for cardiac arrhythmia detection and acute stroke prediction. Convolutional neural network (CNN) and support vector machine (SVM) are combined with spectrogram for classifying heart sounds in the heart diseases diagnosis. It is important to segregate abnormal heart sounds and lung sounds in a phonocardiogram (PCG) based on cardiac arrhythmia analysis [11].

This paper comprises five sections. Section 2 reviews the literature related to edge computing framework, wearable computing, preprocessing of ECG signals, and machine learning-based assessment. Section 3 proposes the edge computing framework with the necessary block diagram and explanations. The obtained results are discussed in Section 4. Finally, the work is concluded in Section 5.

2. Literature Review

In wireless ECG monitoring, wearable devices and sensors collect various physiological parameters of individuals. Sensors are connected through the wireless sensor network for monitoring a greater number of patients. The high-quality signal can be observed through the microcontroller-based preprocessing circuit. ECG signal processing is performed using wavelets for better spectral feature extraction and to improve the SNR. The combination of online monitoring and high-quality processing makes this system a powerful technology for wireless monitoring systems.

CVD risk detection is performed in many works by applying machine learning approaches to identify heart abnormality detection and heartbeat classification. An artificial neural network is used for CVD risk detection in the mobile cloud approach [12]. In their work, the author used virtual machines in the cloud where ANN-based training and learning are done, and the processed data is sent back to the mobile devices. However, the training and learning are performed on the entire data which leads to numerous computations, and cloud computing devices are burdened in case of continuous monitoring of a greater number of users.

Machine learning is also used by many authors in heart disease detection. ANFIS has been used in coronary heart disease risk detection in the cloud-based monitoring system [13]. In addition, decision trees, Naïve Bayes, $K$-nearest neighbor (KNN), and support vectors (SVM) have been utilized in many works in developing better healthcare monitoring systems. Machine learning techniques are applied in data analysis and complex design applications where it is difficult to apply conventional algorithms and techniques [14]. In biomedical applications, many expert systems are used in smart healthcare monitoring devices, hence the huge amount of data to be processed and classified appropriately using machine learning techniques [15].

A review of ECG signal monitoring techniques indicated that there is a requirement for developing a better-quality assessment and preprocessing technique [16]. Mathematical morphology-based processing introduces distortions to the QRS waves, which cause difficulty in the extraction of $R$-$R$ intervals [17]. Hence, morphology-based analysis is not effective for ECG signal preprocessing and feature extraction. To overcome the difficulties in morphology analysis, $R$-peak detection techniques are required. $R$ wave slope detection is vital in locating the QRS complex in ECG signal feature extraction applications. The low $R$ wave slopes may mislead the physician in ECG signal diagnosis. Hence, the noise removal of ECG signals is carried out using filtering algorithms for better feature extraction [18]. In addition, while monitoring the heart rate of athletes, elderly persons and infants require some wearable devices; the conventional monitoring systems have to be improved with wireless monitoring and processing capability.

Cloud computing and edge computing are widely deployed in smart healthcare systems. In cloud-based systems, signal quality analysis is performed on the mobile device. The complex computations are transferred to the cloud server, and the results are sent back to the mobile devices with a display. Mobile and wearable devices are used for the collection of data, and a cloud server is used for storage and computation. Since mobile devices and wearable devices are connected to the hospital through the internet and router directly, the transfer of data can be accomplished easily [19]. The qualified and processed data are sent to edge devices which may be in either a healthcare center or diagnostic laboratory. The preprocessed data are sent to edge devices, and KNN-based CVD risk detection is carried out. KNN-based ECG beat classification is performed for cardiac arrhythmia detection. Many machine-learning algorithms are utilized for biomedical data analysis [20].

3. Proposed Edge Computing Framework

Cardiac arrhythmia detection and acute stroke prediction are focused on in this work using wearable sensors and an edge computing framework. Figure 2 depicts the process flow of the edge computing framework with sensors and cloud servers. In the proposed framework, preprocessing and feature extraction are performed in the edge computing layer, and decisions can be communicated instantly to the individuals. A huge amount of collected data are sent to the cloud server for storage and further analysis in improving the accuracy of acute stroke risk prediction. Since few heart arrhythmia conditions may be life-threatening, the accuracy needs to be improved in acute stroke prediction from the data acquired by wearable devices. To improve the accuracy of prediction, a deep learning-based CNN algorithm is utilized for feature extraction and classification.

[figure(s) omitted; refer to PDF]

A few significant contributions of the proposed work are as follows:

(i) Time domain and frequency domain HRV features are used in cardiac arrhythmia detection

(ii) Various physiological data and HRV features are applied to the CNN for acute stroke prediction

(iii) Both MIT-BIH and real-time recorded ECG signals are used to predict acute stroke using a deep learning model

(iv) A deep CNN-based multiclass classification is performed in acute stroke prediction to classify the subjects into normal, less stroke-risky, and high stroke-risk

3.1. Preprocessing and Feature Extraction

The collected ECG data are preprocessed in the wearable device, and they are sent to the graphics processing unit- (GPU-) based edge computing devices for deep learning-based acute stroke prediction. Blood pressure, glucose level, and cholesterol level can be sent without any preprocessing. However, preprocessing and feature extraction play a major role in ECG signal analysis. A complete data analysis of the ECG signal is highly essential for heart rate feature extraction [21]. Wavelet transform is utilized for ECG signal preprocessing and feature extraction. Statistical and frequency domain analysis is applied to extract the heart rate variability (HRV) of the ECG signal. Statistical features and frequency domain features are utilized for obtaining experimental results. A few statistical features are the mean of $RR$ intervals (RRmean), mean of heart rate (HRmean), the standard deviation of heart rate (HR Std), root mean square standard deviation, and frequency domain features which are VLF Power, LF Power (ms2), HF Power (ms2), and LF norm.

3.2. Assessment of Cardiac Arrhythmia and Acute Stroke Prediction

The collected data are stored as training data and testing data. Usually, 80% of the data can be trained to obtain good classification results. Cardiac arrhythmia detection is performed by preprocessing, model development, testing of the model, and prediction. CNN-based deep learning-based stroke prediction and risk classification are carried out in this stage by the following steps: (i) feature extraction from the quality assessed ECG signal, (ii) applying various physiological data and feature extracted ECG signals to deep learning model, and (iii) cardiac arrhythmia detection is performed by testing the developed model. Figure 3 describes the deep learning model-based cardiac arrhythmia detection.

[figure(s) omitted; refer to PDF]

In this work, deep CNN is utilized for the study with one input layer, two convolution layers, two pooling layers, one fully connected layer, and a classification layer. In this work, two convolution layers and two max-pooling layers are utilized for better localization and feature mapping. Feature maps at the convolution layer are evaluated by $$\begin{array}{}\text{(1)}& X_j^{m,l}=f\sum _{i\in M_j}{X}_j^{m,l-1}\ast W_{ij}^{m,l}+b_j^{m,l}.\end{array}$$

$X_j^{m,l}$ is the one-dimensional kernel to produce $j$^th feature map in the convolution layer. Bias values for the $j$^th filter are $b_j$ and $M_j$.

The feature map for the pooler layer can be computed by $$\begin{array}{}\text{(2)}& X_j^{m,l}={\max }_{v\in V}{X}_j^{m,l-1}r\ast s+v,\end{array}$$where $r$, $s$, and $V$ are the elements of each feature map.

Acute stroke detection is also focused in the proposed edge computing framework by processing physiological signals. Heart-related chronic diseases are mainly responsible for cardiac arrest, heart attack, and sudden cardiac death. Hence, acute stroke or ischemic stroke detection is highly essential in heart-related chronic disease assessment. Real-time testing of the product enables us to make it better to use the wearable device for stroke prediction. The classification accuracy of chronic disease risk and acute stroke prediction is carried out using specificity, sensitivity, and total classification accuracy. Figure 4 depicts the procedure involved in acute stroke prediction.

[figure(s) omitted; refer to PDF]

A deep learning model is developed for acute stroke prediction using a multiclass CNN algorithm. The training phase and testing phase of the deep learning model is illustrated in Figure 5. The deep learning model is provided with several sample data collected from MIT-BIH and real-time data. Sample data comprises both the normal subject data and heart patient data with stroke symptoms. The input data comprises many physiological data, heart rate features, and corresponding results. Training is performed on the input data to obtain a trained deep-learning model. The obtained trained model is used to test the incoming new data from wearable devices. Predicted results from testing are communicated to physicians, healthcare centers, relatives, and emergency services. In case of high stroke risk, it is required to pay immediate attention, and information has to be sent to emergency healthcare services.

[figure(s) omitted; refer to PDF]

4. Results and Discussion

Experimental results are obtained using collected physiological data from wearable devices such as temperature, blood pressure, ECG, respiration rate, and cholesterol level. Heart rate variability features are extracted from both the MIT-BIH data and real-time recorded data. Simulation experiments are performed using Matlab software (version 2019a) with a laptop (8 GB DDR4 RAM and Intel i7 processor with a processor speed of up to 4.6 GHz). Feature extraction from the ECG signal is carried out using National Instruments (NI) Biomedical kit application software. The system performance and battery performance of the mobile are observed by transmitting data to edge devices and servers.

The simulation experiments were performed using both the MIT-BIH database and real-time data collected from wearable devices. From the MIT-BIH database, both the cardiac disease patient data and healthy people records have been applied. Fifty heart patients and thirty healthy individuals were recruited for the real-time wearable computing-based data collection. Though the ECG signal amplitudes are in the $mV$ range, the obtained amplitudes are due to the ECG amplifiers. The acquired ECG signal is processed using wavelets for measuring physiological parameters. After identifying the QRS complexes in the ECG signal using DWT, the heart rate can be measured using the interval between R peaks [22].

Heart rate variability features and physiological data such as blood pressure, respiration rate, temperature, and cholesterol level have been fed into the CNN classifier. These details are included in the first paragraph of the results and discussion section. In addition, the parameters of CNN data are discussed in Figure 6. In the existing works, machine learning algorithms are utilized with preprocessing and feature extraction for heart arrhythmia analysis and stroke prediction; however, they could not obtain the required classification accuracy. In this paper, the CNN classifier is used to improve classification accuracy with three output layers for providing normal, low stroke risky, and high stroke risky.

[figure(s) omitted; refer to PDF]

Statistical features such as the mean of $RR$ intervals (RRmean), mean of heart rate (HRmean), the standard deviation of heart rate (HR Std), root mean square standard deviation, and frequency domain parameters such as VLF (ms2), LF (ms2), HF (ms2), frequency ratio LF/HF, LF norm, and HF norm are noted for all the subjects. Power spectral density plots with frequency domain features and spectrogram plots with wavelet coefficients are observed using NI biomedical kit. These features are highly essential to improve the accuracy of the proposed deep learning model. Figure 6(a) depicts the power spectral density plot with all the frequency domain features. Figure 6(b) shows the spectrogram plot of the ECG data to obtain the wavelet coefficients and their distribution in terms of scale values. The scale of subjects is varying from 0 to 59 from 7.1 to 36.6 sec. Assessment of the observed values is used to analyze heart arrhythmia and acute stroke prediction.

Time domain features such as mean and standard deviation of heart rate (HR) are considered. Threshold values are assigned to the HR mean and HR standard deviation while categorizing the considered data into normal and disease risk. A total of 512 training data (80%) with a signal length of 60 seconds have been considered which includes MIT-BIH database samples. After training, 128 testing data (20%) were used to validate the accuracy of the classifier. The $R$-peak detected signal is used to calculate beats per minute. To evaluate the performance of the proposed approach, sensitivity, specificity, and accuracy are calculated. These parameters are defined as follows: $$\begin{array}{}\text{(3)}& \text{Sensitivity}=\frac{\text{TP}}{\text{TP}+\text{FN}}\times 100\%,\text{(4)}& \text{Specificity}=\frac{\text{TN}}{\text{TN}+\text{FP}}\times 100\%,\text{(5)}& \text{Accuracy}=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{FN}+\text{TN}+\text{FP}}\times 100\%.\end{array}$$

In the proposed work, a CNN classifier with a learning rate of 0.1, epochs of 25, and batch size of 640 has been selected for training and learning. Training accuracy, testing accuracy, and training error are measured and depicted in Figure 7.

[figure(s) omitted; refer to PDF]

In the edge computing-based approach, the battery power requirement is ranging from 1 to 1.5% of the total battery power. It is worth noting that mobile approaches consume 12 to 14% of the battery. The execution speed improves by six to ten times in the proposed edge computing approach with three-edge devices when compared with single-edge devices. Latency analysis is also carried out to study the effective performance of the proposed edge framework. Figure 8 shows the latency analysis and comparison of the edge computing framework and the computing framework without any edge devices. The latency time is greatly reduced by the edge computing framework due to the less dependency on cloud services.

[figure(s) omitted; refer to PDF]

Execution speed and latency time are calculated to assess the performance of the proposed edge computing framework. Execution speed improves in the framework due to the local processing and analysis through edge devices. The speed is further improved by increasing the number of edge devices by minimizing the dependency on cloud-based services. Table 1 compares the execution speed of the proposed edge computing approach by deploying a different number of edge devices. In addition to performing accuracy in terms of specificity and sensitivity, the battery is not burdened much while intensive computing operations in the edge computing-based approach.

Table 1

Comparison of execution speed while deploying different edge devices.

Data	Execution speed
	Without edge (sec)	Single-edge device (sec)	Two-edge devices (sec)	Three-edge devices (sec)	Speed increase
MIT-BIH # 101	713	453	324	112	6.4×
MIT-BIH # 102	709	511	355	106	6.9×
MIT-BIH # 103	813	422	311	97	8.4×
MIT-BIH # 104	754	504	297	84	8.9×
MIT-BIH # 105	823	421	324	86	9.6×
MIT-BIH # 200	765	435	231	90	8.5×
MIT-BIH # 205	912	401	341	105	8.7×
NI lab #101	854	398	375	125	6.8×
NI lab 2 #102	939	431	392	132	7.1×

In most classification approaches, it is necessary to compare the sensitivity, specificity, and accuracy of different approaches while applying them for risk detection. In this work, extracted heart rate features using DWT and physiological parameters are applied to the deep CNN approach. Here, multilead ECG with multiple instance learning (ML-ECG & MIL), morphology features with KNN (MF-KNN), energy features with SVM (EF-SVM), and multiple feature-based CNN (MF-CNN) methods are compared to proposed DWT with deep CNN approach. Table 2 compares the obtained values of specificity, sensitivity, and accuracy while considering 250 patient records. Figures 9(a) and 9(b) depict the comparative sensitivity analysis and comparative specificity analysis, respectively, which are performed on 50, 100, 150, 200, and 250 patient records. It is observed from these figures that there is a slight variation in the performance. However, the proposed DWT with a deep CNN approach is superior to other existing methods due to its feature extracted and well-trained data.

Table 2

Comparison of specificity, sensitivity, and accuracy.

Classifier	MIT-BIH
	Specificity (%)	Sensitivity (%)	Accuracy (%)
ML-ECG & MIL [23]	84.7	85.4	84.9
MF-KNN [24]	86.8	87.4	87.1
EF-SVM [25]	92.2	92.4	92.2
MF-CNN [26]	94.9	95.4	95.1
Proposed (DWT + CNN)	99.1	99.4	99.3

[figure(s) omitted; refer to PDF]

From Table 2, it is observed that specificity, sensitivity, and accuracy are high in the approach compared with similar methods. Sensitivity, specificity, and accuracy analyses are carried out with a different number of records.

5. Conclusion

In this work, an edge computing approach is proposed to improve the accuracy and speed of assessing cardiac arrhythmia and acute stroke prediction. A deep learning-based CNN model is developed for detecting cardiac arrhythmia and predicting acute stroke. The physiological data and heart rate features of both the MIT-BIH and real-time data are applied to the deep learning model. The proposed DWT-based feature extraction and deep CNN-based multiclass classification provide more accuracy than many existing feature extraction and classification approaches. The proposed classifier achieves a sensitivity of 99.4%, specificity of 99.1%, and accuracy of 99.3% when compared with other similar approaches. The execution speed improves by six to ten times in the proposed edge computing approach with three-edge devices when compared with the single-edge device.

References

[1] D. He, S. Zeadally, "An analysis of RFID authentication schemes for Internet of Things in healthcare environment using elliptic curve cryptography," IEEE Internet of Things Journal, vol. 2 no. 1, pp. 72-83, DOI: 10.1109/JIOT.2014.2360121, 2015.

[2] A. R. M. Forkan, I. Khalil, "PEACE-home: probabilistic estimation of abnormal clinical events using vital sign correlations for reliable home-based monitoring," Pervasive and Mobile Computing, vol. 38, pp. 296-311, DOI: 10.1016/j.pmcj.2016.12.009, 2017.

[3] S. Aziz, M. U. Khan, M. Alhaisoni, T. Akram, M. Altaf, "Phonocardiogram signal processing for automatic diagnosis of congenital heart disorders through fusion of temporal and cepstral features," Sensors, vol. 20 no. 13,DOI: 10.3390/s20133790, 2020.

[4] L. García, J. Tomás, L. Parra, J. Lloret, "An m-health application for cerebral stroke detection and monitoring using cloud services," International Journal of Information Management, vol. 45, pp. 319-327, DOI: 10.1016/j.ijinfomgt.2018.06.004, 2019.

[5] L. Catarinucci, D. D. Donno, L. Mainetti, L. Palano, L. Patrono, M. L. Stefanizzi, L. Tarricone, "An IoT-aware architecture for smart healthcare systems," IEEE Internet of Things Journal, vol. 2 no. 6, pp. 515-526, DOI: 10.1109/jiot.2015.2417684, 2015.

[6] M. S. Hossain, G. Muhammad, A. Alamri, "Smart healthcare monitoring: a voice pathology detection paradigm for smart cities," Multimedia Systems, vol. 25, 2017.

[7] X. Li, X. Huang, C. Li, R. Yu, L. Shu, "EdgeCare: leveraging edge computing for collaborative data management in mobile healthcare systems," IEEE Access, vol. 7, pp. 22011-22025, DOI: 10.1109/ACCESS.2019.2898265, 2019.

[8] I. V. Pustokhina, D. A. Pustokhin, D. Gupta, A. Khanna, K. Shankar, G. N. Nguyen, "An effective training scheme for deep neural network in edge computing enabled Internet of medical things (IoMT) systems," IEEE Access, vol. 8, pp. 107112-107123, DOI: 10.1109/ACCESS.2020.3000322, 2020.

[9] A. Alabdulatif, I. Khalil, X. Yi, M. Guizani, "Secure edge of things for smart healthcare surveillance framework," IEEE Access, vol. 7, pp. 31010-31021, DOI: 10.1109/ACCESS.2019.2899323, 2019.

[10] S. Tuli, N. Basumatary, S. S. Gill, M. Kahani, R. C. Arya, G. S. Wander, R. Buyya, "Healthfog: an ensemble deep learning based smart healthcare system for automatic diagnosis of heart diseases in integrated iot and fog computing environments," Future Generation Computer Systems, vol. 104, pp. 187-200, DOI: 10.1016/j.future.2019.10.043, 2020.

[11] Y. Karaca, M. Moonis, Y. D. Zhang, C. Gezgez, "Mobile cloud computing based stroke healthcare system," International Journal of Information Management, vol. 45, pp. 250-261, DOI: 10.1016/j.ijinfomgt.2018.09.012, 2019.

[12] X. Wang, Q. Gui, B. Liu, Z. Jin, Y. Chen, "Enabling smart personalized healthcare: a hybrid mobile-cloud approach for ECG telemonitoring," IEEE Journal of Biomedical and Health Informatics, vol. 18 no. 3, pp. 739-745, DOI: 10.1109/jbhi.2013.2286157, 2014.

[13] H. Yang, Z. Wei, "Arrhythmia recognition and classification using combined parametric and visual pattern features of ECG morphology," IEEE Access, vol. 8, pp. 47103-47117, DOI: 10.1109/ACCESS.2020.2979256, 2020.

[14] D. M. Farid, L. Zhang, C. M. Rahman, M. A. Hossain, R. Strachan, "Hybrid decision tree and naive Bayes classifiers for multi-class classification tasks," Expert Systems with Applications, vol. 41 no. 4, pp. 1937-1946, DOI: 10.1016/j.eswa.2013.08.089, 2014.

[15] A. Ullah, S. U. Rehman, S. Tu, R. M. Mehmood, M. Ehatisham-Ul-Haq, "A hybrid deep CNN model for abnormal arrhythmia detection based on cardiac ECG signal," Sensors, vol. 21 no. 3,DOI: 10.3390/s21030951, 2021.

[16] N. Dey, A. S. Ashour, F. Shi, S. J. Fong, R. S. Sherratt, "Developing residential wireless sensor networks for ECG healthcare monitoring," IEEE Transactions on Consumer Electronics, vol. 63 no. 4, pp. 442-449, DOI: 10.1109/TCE.2017.015063, 2017.

[17] S. Yazdani, J. M. Vesin, "Extraction of QRS fiducial points from the ECG using adaptive mathematical morphology," Digital Signal Processing, vol. 56, pp. 100-109, DOI: 10.1016/j.dsp.2016.06.010, 2016.

[18] N. Wang, J. Zhou, G. Dai, J. Huang, Y. Xie, "Energy-efficient intelligent ECG monitoring for wearable devices," IEEE Transactions on Biomedical Circuits and Systems, vol. 13 no. 5, pp. 1112-1121, DOI: 10.1109/TBCAS.2019.2930215, 2019.

[19] S. J. Kang, S. Y. Lee, H. I. Cho, H. Park, "ECG authentication system design based on signal analysis in mobile and wearable devices," IEEE Signal Processing Letters, vol. 23 no. 6, pp. 805-808, DOI: 10.1109/LSP.2016.2531996, 2016.

[20] J. W. Lee, H. S. Lim, D. W. Kim, S. A. Shin, J. Kim, B. Yoo, K. H. Cho, "The development and implementation of stroke risk prediction model in National Health Insurance Service’s personal health record," Computer Methods and Programs in Biomedicine, vol. 153, pp. 253-257, DOI: 10.1016/j.cmpb.2017.10.007, 2018.

[21] R. Gutiérrez-Rivas, J. J. García, W. P. Marnane, A. Hernández, "Novel real-time low-complexity QRS complex detector based on adaptive thresholding," IEEE Sensors Journal, vol. 15 no. 10, pp. 6036-6043, DOI: 10.1109/JSEN.2015.2450773, 2015.

[22] M. Li, R. Guo, K. Zhang, Z. Lin, F. Yang, S. Xu, A. Abubakar, "Machine learning in electromagnetics with applications to biomedical imaging: a review," IEEE Antennas and Propagation Magazine, vol. 63 no. 3, pp. 39-51, DOI: 10.1109/MAP.2020.3043469, 2021.

[23] L. Sun, Y. Lu, K. Yang, S. Li, "ECG analysis using multiple instance learning for myocardial infarction detection," IEEE Transactions on Biomedical Engineering, vol. 59 no. 12, pp. 3348-3356, DOI: 10.1109/TBME.2012.2213597, 2012.

[24] M. Arif, I. A. Malagore, F. A. Afsar, "Detection and localization of myocardial infarction using k-nearest neighbor classifier," Journal of Medical Systems, vol. 36 no. 1, pp. 279-289, DOI: 10.1007/s10916-010-9474-3, 2012.

[25] L. N. Sharma, R. K. Tripathy, S. Dandapat, "Multiscale energy and eigenspace approach to detection and localization of myocardial infarction," IEEE Transactions on Biomedical Engineering, vol. 62 no. 7, pp. 1827-1837, DOI: 10.1109/TBME.2015.2405134, 2015.

[26] W. Liu, Q. Huang, S. Chang, H. Wang, J. He, "Multiple-feature-branch convolutional neural network for myocardial infarction diagnosis using electrocardiogram," Biomedical Signal Processing and Control, vol. 45, pp. 22-32, DOI: 10.1016/j.bspc.2018.05.013, 2018.