ECG Beat Research Articles

Documenting Spatial Variation of SCG Signals for Optimal Sensor Placement

Introduction Seismocardiography (SCG) (low frequency cardiovascular vibrations acquired on the chest wall) has potential for early noninvasive detection of heart failure (HF) deterioration. However, inaccurate sensor placement may significantly limit the method's utility. Objective To document SCG waveform spatial variability to aid in sensor placement and quantification of sensor position shifts on the signal reproducibility. Methods SCG signals were recorded in 15 healthy subjects using 36 accelerometers placed on the anterior chest wall along with simultaneous ECG and respiratory flow measurements (Figure 1). SCG waveforms were segmented using ECG beats. SCG feature points were identified according to previous studies and chest surface acceleration maps created. Results Figure 2 suggests that SCG may be loudest and positive (i.e., anterior or outward) acceleration at the left lower sternal border (LLSB) during the aortic opening and closure periods. A negative (i.e. posterior or inward) acceleration was observed around the LLSB location during the isovolumic contraction time period. The SCG intra-group variability with change in sensor position is shown in Table 1.Table 1: SCG intra-group variability with sensor position change Conclusions Optimal sensor location was found around the LLSB between 3rd to 5th ICS. Use of this location may help increase robustness of SCG signal interpretation and its utility for HF monitoring. Further, change in sensor locations of 3 cm around 4thICS at LLSB (the target sensor location) resulted in only a ∼7 % change in waveform variability and ∼9 % change in PEP, which if true, would better facilitate ambulatory use and sensor placement by patients.

SVM Based ECG Beat Classification Method for Unsupervised ECG Diagnosis Systems

SVM Based ECG Beat Classification Method for Unsupervised ECG Diagnosis Systems

Comparison Between Non-Linear Autoregressive and Non-Linear Autoregressive with Exogeneous Inputs Models for Predicting Cardiac Ischemic Beats

The prediction accuracy and generalization ability of neural models for forecasting Myocardial Ischemic Beats depends on type and architecture of employed network model. This paper presents the comparison analysis of recurrent neural network (RNN) architectures with embedded memory, Non-linear Autoregressive (NAR) and Non-linear Autoregressive with Exogeneous inputs (NARX) models for forecasting Ischemic Beats in ECG. Numerous architectures of the NAR and NARX models are verified for prediction and the performances are evaluated in terms of MSE. The performances of NAR and NARX models are validated by using ECG signals acquired from MIT-BIH database. The results have depicted that the NARX architecture with 2 neurons in hidden layer and 1 delay line outperformed with least Mean Square Error (MSE) of 0.0001 for detecting the ischemic beats in ECG.

Analysis and classification of ECG beat based on wavelet decomposition and SVM

Objectives: To extract the features of single arrhythmia ECG beat. To develop efficient algorithms for automated detection of arrhythmia based on ECG. Methods/Statistical analysis: The methodology includes pre-processing and segmentation of ECG. Extraction of ECG features are to support the ECG beat classification and analysis of cardiac abnormalities using machine learning techniques. Wavelet decomposition is considered for feature extraction and classification with multiclass support vector machine. Findings: This work evaluates the suitability of the wavelet features of ECG for classifier. The proposed arrhythmia classifier results in an accuracy up to 98% for various classes of arrhythmia considered in this work. Novelty/Applications: This work is an assistive tool for medical practitioners to examine ECG in a limited time with their expertise to make the accurate abnormality diagnosis of the arrhythmia. Keywords: Arrhythmia; classification; feature extraction; support vector machine; wavelet decomposition

Analysis and classification of cardiac arrhythmia based on general sparsed neural network of ECG signals

In medical practices, the ECG plays an important role in diagnosing cardiac arrhythmia. In this paper efficient and most reliable technique is mentioned for the suitable classification of arrhythmia using a general sparsed neural network (GSNN). The sparsed neural network is used to extract the feature of ECG signals and then this feature is used in the neural network for processing to obtain the final classification result. The different class of ECG beat is chosen from MIT-BIH dataset. The signal to noise ratio has been calculated to filter the ECG signal. To evaluate the results, the MATLAB software is used. The main purpose of this paper is to design an efficient neural network and also to implement reliable techniques for the classification of various ECG arrhythmia conditions. The obtained accuracy level of arrhythmia detection is 98%, which is the highest rate of performance. The presented approach by GSNN to classify and predict arrhythmia will provide efficient arrhythmia detection as compared to other techniques. The suggested scheme will improve the prediction and classification efficiency.

Optimized time–frequency features and semi-supervised SVM to heartbeat classification

One of the most significant indicator of heart disease is arrhythmia showing heartbeat patterns. Thus, early and accurate detection of arrythmia types by categorization of heartbeats is important. In this paper, we introduce an ECG beat classifier system integrating two main parts: feature extraction and classification. For the first part, we consider the features observed in the time–frequency (t, f) plane where the ECG is projected using a variant of Stockwell transform. For the second part, the framework of semi-supervised SVM with asymmetric costs (AS3VM) has been applied for assessment of the obtained feature sets performance. Notice that four heartbeat types have been considered: normal beats (N), left and right bundle branch blocks (L and R) and premature ventricular contractions (V). The proposed method has been evaluated on PhysionNet’s MIT-BIT arrythmia database. The obtained results show that the suggested approach achieves significant separability of the classes and thus, able to make prediction accuracies of $$99.35\%$$ , $$98.73\%$$ , $$98.57\%$$ and $$99.44\%$$ for, respectively, N, L, R and V beats.

Local feature descriptors based ECG beat classification.

ECG beat type analysis is important in the detection of various heart diseases. The ECG beats give useful information about the status of the monitored heart condition. Up to now, various artificial intelligence-based methods have been proposed for ECG based heart failure detection. These methods were generally based on either time or frequency domain signal processing routines. In this study, we propose a different approach for ECG beat classification. The proposed approach is based on image processing. Thus, the initial step of the proposed work is converting the ECG beat signals to the ECG beat images. To do that, the ECG beat snapshots are initially saved as ECG beat images and then local feature descriptors are considered for feature extraction from ECG beat images. Eight local feature descriptors namely Local Binary Patterns, Frequency Decoded LBP, Quaternionic Local Ranking Binary Pattern, Binary Gabor Pattern, Local Phase Quantization, Binarized Statistical Image Features, CENsus TRansform hISTogram and Pyramid Histogram of Oriented Gradients are considered for feature extraction. The Support Vector Machines (SVM) classifier is used in the classification stage of the study. Linear, Quadratic, Cubic and Gaussian kernel functions are used in the SVM classifier. Five types of ECG beats from the MIT-BIH arrhythmia dataset are considered in experiments and the classification accuracy is used for performance measure. To construct a balanced training and test sets, 5000 and 10,000 ECG beat samples are randomly selected and are used in experiments in tenfold cross-validation fashion. The obtained results show that the proposed method is quite efficient where the calculated accuracy score is 99.9% and the comparisons with the state-of-the-art method show that the proposed method outperforms other methods.

Effects of ECG Summarization Methods on the Characterization of Drug‐Induced QT Prolongation – a Demonstration of Need for Standard Methods of ECG Beat Summarization6.3.6

A vital aspect of drug‐development is the pre‐clinical screening of drug‐induced cardiotoxicity. The effect a drug has on the ECG QT‐interval (QTi) is an especially vital screening due to the positive relationship between QT prolongation and arrhythmic events, such as Torsade de Pointes. Consequently, the accurate recording, measurement, and calculation of QTi and heart rate corrected QTi is necessary for cardiotoxicity studies. However, many aspects of these processes are subjective, and techniques vary greatly between researchers and across studies. Examples of subjective aspects of QT prolongation (QTp) characterization include, but are not limited to: the method of heart rate correction for QT (QTc), the number of data points considered, the number of ECG beats or amount of time averaged per data point, and the specific time points evaluated. While the effects of different QTc methods has previously been explored in the literature, the efficiency and accuracy of various ECG summarization methods still warrants further refinement. We evaluated the effects of different methods in order to begin identifying potential industry standards that can minimize the number of data points needed without compromising model prediction. In order to demonstrate the effect different summarization methods have on the identification of drug‐induced QTp, we used Ponemah software to analyze previously collected ECG telemetry data from non‐human primates. The subjects were administered vehicle, three doses of dofetilide, vehicle again, and three doses of quinidine, with a week between each dose. Data was summarized using various methods, including 1, 15, 30, 60 and 120 minute continuous averages as well as 1‐minute averages taken at 1‐hour intervals. We then quantified the effect size of the doses on QTi using restricted maximum likelihood (REML) estimation. Using these estimates, differences in results between the summarization methods were then compared. After finding dramatic differences, we performed power analyses on each summarization method to identify the most accurate and efficient methods. Results suggest that characterization of the dose response and power varied depending on the ECG summarization method used. Using continuous interval averages was more consistent and had greater power than selecting a 1‐minute average at various intervals. Further improvement is achieved by taking pharmacokinetics into account when selecting time intervals. These results confirm the need for careful consideration of ECG summarization methods during study design in order to avoid missing vital data or introducing researcher bias. It also provides evidence to support the recommendation for the introduction of a uniform industry standard for analyses of cardiovascular data. 6.3.6Support or Funding InformationThis work was funded in part by a research grant from Eli Lily and Co, Indianapolis, IN

A Lightweight Deep Learning Model for Fast Electrocardiographic Beats Classification With a Wearable Cardiac Monitor: Development and Validation Study.

BackgroundElectrocardiographic (ECG) monitors have been widely used for diagnosing cardiac arrhythmias for decades. However, accurate analysis of ECG signals is difficult and time-consuming work because large amounts of beats need to be inspected. In order to enhance ECG beat classification, machine learning and deep learning methods have been studied. However, existing studies have limitations in model rigidity, model complexity, and inference speed.ObjectiveTo classify ECG beats effectively and efficiently, we propose a baseline model with recurrent neural networks (RNNs). Furthermore, we also propose a lightweight model with fused RNN for speeding up the prediction time on central processing units (CPUs).MethodsWe used 48 ECGs from the MIT-BIH (Massachusetts Institute of Technology-Beth Israel Hospital) Arrhythmia Database, and 76 ECGs were collected with S-Patch devices developed by Samsung SDS. We developed both baseline and lightweight models on the MXNet framework. We trained both models on graphics processing units and measured both models’ inference times on CPUs.ResultsOur models achieved overall beat classification accuracies of 99.72% for the baseline model with RNN and 99.80% for the lightweight model with fused RNN. Moreover, our lightweight model reduced the inference time on CPUs without any loss of accuracy. The inference time for the lightweight model for 24-hour ECGs was 3 minutes, which is 5 times faster than the baseline model.ConclusionsBoth our baseline and lightweight models achieved cardiologist-level accuracies. Furthermore, our lightweight model is competitive on CPU-based wearable hardware.

ECG Beat Classification using a Sliding Window and Correlation of the Three-bit Linear Prediction Error Signal

Sudden cardiac arrest (SCA) is responsible for half of all deaths due to heart disease. Most SCAs could be avoided by obtaining an early diagnosis from ECG recordings. The long-term monitoring systems record a large number of beats and require automatic detection and classification of the premature ventricular contraction (PVC) beats. Several ECG beat classification algorithms based on different methodologies have been developed and implemented. This paper presents a novel algorithm for automatic recognition of a premature ventricular contraction (PVC) beat based on a three-bit linear prediction error signal (LPES). The algorithm is composed of three main stages: signal denoising and QRS detection; nonlinear transformation of the linear prediction error signal e(n); and a sliding window. The proposed algorithm was tested using ECG signals from two recognized arrhythmia databases, MIT-BIH and AHA. The selected signals contained normal beats as well as abnormal beats. Sensitivity and specificity parameters were used to measure the accuracy of the proposed classifier. The sensitivity achieved using the proposed algorithm was 96.3% and the specificity was 99.0%. In addition to its accuracy, the main advantages of using the proposed algorithm are its simplicity and robustness.

False Alarm Reduction in Self-Care by Personalized Automatic Detection of ECG Electrode Cable Interchanges.

Introduction False alarm reduction is an important challenge in self-care, whereas one of the most important false alarm causes in the cardiology domain is electrodes misplacements in ECG recordings, the main investigations to perform for early and pervasive detection of cardiovascular diseases. In this context, we present and assess a new method for electrode reversals identification for Mason-Likar based 3D ECG recording systems which are especially convenient to use in self-care and allow to achieve, as previously reported, high computerized ischemia detection accuracy. Methods We mathematically simulate the effect of the six pairwise reversals of the LA, RA, LL, and C2 electrodes on the three ECG leads I, II, and V2. Our approach then consists in performing serial comparisons of the newly recorded 3D ECG and of the six derived ECGs simulating an electrode reversal with a standard, 12-lead reference ECG by means of the CAVIAR software. We further use a scoring method to compare these analysis results and then apply a decision tree model to extract the most relevant measurements in a learning set of 121 patients recorded in ICU. Results The comparison of the seven sets of serial analysis results from the learning set resulted in the determination of a composite criteria involving four measurements of spatial orientation changes of QRS and T and providing a reversal identification accuracy of 100%. Almost the same results, with 99.99% of sensitivity and 100% of specificity, were obtained in two test sets from 90 patients, composed of 2098 and 2036 representative ECG beats respectively recorded during PTCA balloon inflation, a procedure which mimics ischemia, and before PTCA for control. Conclusion Personalized automatic detection of ECG electrode cable interchanges can reach almost the maximal accuracy of 100% in self-care, and can be performed in almost real time.

Transfer Learning in ECG Classification from Human to Horse Using a Novel Parallel Neural Network Architecture

Automatic or semi-automatic analysis of the equine electrocardiogram (eECG) is currently not possible because human or small animal ECG analysis software is unreliable due to a different ECG morphology in horses resulting from a different cardiac innervation. Both filtering, beat detection to classification for eECGs are currently poorly or not described in the literature. There are also no public databases available for eECGs as is the case for human ECGs. In this paper we propose the use of wavelet transforms for both filtering and QRS detection in eECGs. In addition, we propose a novel robust deep neural network using a parallel convolutional neural network architecture for ECG beat classification. The network was trained and tested using both the MIT-BIH arrhythmia and an own made eECG dataset with 26.440 beats on 4 classes: normal, premature ventricular contraction, premature atrial contraction and noise. The network was optimized using a genetic algorithm and an accuracy of 97.7% and 92.6% was achieved for the MIT-BIH and eECG database respectively. Afterwards, transfer learning from the MIT-BIH dataset to the eECG database was applied after which the average accuracy, recall, positive predictive value and F1 score of the network increased with an accuracy of 97.1%.

Automated ECG beat classification using DWT and Hilbert transform-based PCA-SVM classifier

Automated ECG beat classification using DWT and Hilbert transform-based PCA-SVM classifier

Diagnosis of Left Ventricular Hypertrophy Based on Electrocardiogram Signal and Fuzzy Logic

Sokolow – Lyon index in detection of left ventricular hypertrophy is a hard limited index, so the clinical manifestation of the disease can be ignored when the measured index is near the threshold. Several proposed studies incorporate multiple index to improve diagnostic quality. However, the process of examination and diagnosis will be longer due to the need to collect more data. To solve this problem, the paper proposes a method of classifying left ventricular hypertrophy using fuzzy logic combining with digital signal processing techniques. The proposed method mainly uses the Sokolov-Lyon index (SV1+RV5/V6 ≥ 35 mm) for major changes in ECG signal but with four soft thresholds corresponding to the different clinical manifestations of the disease. In addition, a program is written in C++ language with QT Creator compiler also is developed to implement the algorithm. From there, the doctors can refer and propose to the patient's treatment regimen.
 Keywords
 ECG, left ventricular hypertrophy, signal processing, fuzzy logic.
 References
 [1] Malcolm S. Thaler, The only EKG book, seventh ed. Lippincott Williams & Wilinks, Philadelphia, 2012. [2] Vakili BA, Okin PM, Devereux RB, Prognostic implications of left ventricular hypertrophy, Am Heart J, 141(3) (2001) 334-341.https://doi.org/10.1067/mhj.2001.113218.[3] Tran Do Trinh, Tran Van Dong, How to read EGC signal, Medical Publishing House, 2011 (in Vietnamese).[4] Himanshu Gothwal1, Silky Kedawat1, Rajesh Kumar, Cardiac arrhythmias detection in an ECG beat signal using fast fourier transform and artificial neural network, J. Biomedical Science and Engineering 4 (2011) 289-296.https://doi.org/10.4236/jbise.2011.44039.[5] El-Sayed A. El-Dahshan, Genetic algorithm and wavelet hybrid scheme for ECG signal denoising, Journal of Telecommunications Systems 46(3) (2011) 209-215.https://doi.org/10.1007/s11235-010-9286-2.[6] C. Li, C. Zheng, and C. Tai, Detection of ECG characteristic points using wavelet transforms, IEEE Trans.Biomed. Eng 42(1) (1995) 21-28.https://doi.org/10.1109/10.362922.[7] A.K.M. Fazlul Haque1, Md. Hanif Ali1, M. Adnan Kiber2 and Md. Tanvir Hasan, Detection of small variations of ECG features using Wavelet, ARPN Journal of Engineering and Applied Sciences 4(6) (2009) 27-30.[8] Krimi Samar, Ouni Kas, Noureddine Ellouze, Using Hidden Markov Models for ECG Characterisation, Hidden Markov Models, Theory and Applications, 4 (2011) 151-165.https://doi.org/10.5772/13916.[9] Van Ngoc Tuyet, Bang Ai Vien, Nguyen Van Tri, Medical Journal Ho Chi Minh city, Diagnosis of left ventricular hypertrophy by ECG I 15(1) (2011) 135-140 (in Vietnamese).[10] Buckley, James J., Eslami, Esfandiar, Introduction to Fuzzy Logic and Fuzzy Sets, Physica-Verlag Heidelberg, Berlin, 2002. [11] Phan Xuan Minh, Nguyen Doan Phuoc, Fuzzy Control Theory, Science and Technics Publishing House, Ha Noi, 2006 (in Vietnamese).

Proposing feature engineering method based on deep learning and K-NNs for ECG beat classification and arrhythmia detection.

Arrhythmia is slow, fast or irregular heartbeat. Manual ECG assessment and disease classification is an error-prone task because of vast differences in ECG morphology and difficulty in accurate identifying ECG components. Moreover, proposing a computer-aided diagnosis system for heartbeat classification can be useful when access to medical care centers is difficult or impossible. Therefore, the main aim of this study is classifying ECG beats for arrhythmia detection (four beat classes are considered). Previous studies have proposed different methods based on traditional machine learning and/or deep learning. In this paper, a novel feature engineering method is proposed based on deep learning and K-NNs. The features extracted by our proposed method are classified with different classifiers such as decision trees, SVMs with different kernels and random forests. Our proposed method has reasonably good performance for beat classification and achieves the average Accuracy of 99.77%, AUC of 99.99%, Precision of 99.75% and Recall of 99.30% using fivefold Cross Validation strategy. The main advantage of the proposed method is its low computational time compared to training deep learning models from scratch and its high accuracy compared to the traditional machine learning models. The strength and suitability of the proposed method for feature extraction is shown by the high balance between sensitivity and specificity.

A NARX Model to Predict Myocardial Ischemic Beats from ECG Using Features Extracted by ICA and WPD

This paper presents a methodology for predicting Myocardial Ischemic Beats from ECG signal using a Nonlinear Autoregressive Neural Network with Exogenous inputs (NARX) model. This technique utilizes the features extracted by integrating independent component analysis (ICA) and Wavelet packet decomposition (WPD) on ECG for detecting Myocardial Ischemic beats. At first, the denoised ECG beat segments are projected on the bases to create the independent component (IC) vectors. Further, these IC vectors are disintegrated by WPD. The feature set for distinguishing ischemic beats is extracted by calculating entropy, mean and standard deviation from wavelet coefficients. These features are input to NARX model for uncovering ischemic beats from normal beats. Several architectures of NARX models were tested for predicting myocardial ischemic beats. The efficacy of NARX architectures are assessed by comparing MSE and correlation coefficient. The NARX model with 2 hidden neurons and 2 delay lines provided the best results with a MSE of 0.0002 and correlation coefficient 0.99, which implies that the NARX neural network has huge potential in the prognosis of Myocardial Ischemic Beats.

Phase Space Reconstruction Based CVD Classifier Using Localized Features

This paper proposes a generalized Phase Space Reconstruction (PSR) based Cardiovascular Diseases (CVD) classification methodology by exploiting the localized features of the ECG. The proposed methodology first extracts the ECG localized features including PR interval, QRS complex, and QT interval from the continuous ECG waveform using features extraction logic, then the PSR technique is applied to get the phase portraits of all the localized features. Based on the cleanliness and contour of the phase portraits CVD classification will be done. This is first of its kind approach where the localized features of ECG are being taken into considerations unlike the state-of-art approaches, where the entire ECG beats have been considered. The proposed methodology is generic and can be extended to most of the CVD cases. It is verified on the PTBDB and IAFDB databases by taking the CVD including Atrial Fibrillation, Myocardial Infarction, Bundle Branch Block, Cardiomyopathy, Dysrhythmia, and Hypertrophy. The methodology has been tested on 65 patients’ data for the classification of abnormalities in PR interval, QRS complex, and QT interval. Based on the obtained statistical results, to detect the abnormality in PR interval, QRS complex and QT interval the Coefficient Variation (CV) should be greater than or equal to 0.1012, 0.083, 0.082 respectively with individual accuracy levels of 95.3%, 96.9%, and 98.5% respectively. To justify the clinical significance of the proposed methodology, the Confidence Interval (CI), the p-value using ANOVA have been computed. The p-value obtained is less than 0.05, and greater F-statistic values reveal the robust classification of CVD using localized features.

GABC based neuro-fuzzy classifier with hybrid features for ECG Beat classification

The main objective of the proposed methodology is to classify an ECG signal as normal or abnormal using the optimal neuro-fuzzy classifier. The proposed work consists of two phases namely, feature extraction and neuro-fuzzy classifier based classification. The beat signals are initially taken from the physio-bank ATM. Then, three types of features are extracted from each signal namely, Morphological-based features, Haar wavelet-based features, and Tri-spectrum based features. After feature extraction, the optimal neuro-fuzzy classifier is classifying the beat signal as normal or abnormal. Here, Artificial Bee Colony (ABC) algorithm is combined with Genetic Algorithm (GA) for training the neuro-fuzzy classifier. For experimental evaluation, the MIT-BIH Arrhythmia Database is utilized and the performances are analyzed in terms of accuracy, sensitivity, and specificity. The experimental results clearly demonstrated that the proposed technique outperformed by having better accuracy of 93% when compared existing technique achieved 84% only.

An Automated ECG Beat Classification System Using Deep Neural Networks with an Unsupervised Feature Extraction Technique

An automated classification system based on a Deep Learning (DL) technique for Cardiac Disease (CD) monitoring and detection is proposed in this paper. The proposed DL architecture is divided into Deep Auto-Encoders (DAEs) as an unsupervised form of feature learning and Deep Neural Networks (DNNs) as a classifier. The objective of this study is to improve on the previous machine learning technique that consists of several data processing steps such as feature extraction and feature selection or feature reduction. It is also noticed that the previously used machine learning technique required human interference and expertise in determining robust features, yet was time-consuming in the labeling and data processing steps. In contrast, DL enables an embedded feature extraction and feature selection in DAEs pre-training and DNNs fine-tuning process directly from raw data. Hence, DAEs is able to extract high-level of features not only from the training data but also from unseen data. The proposed model uses 10 classes of imbalanced data from ECG signals. Since it is related to the cardiac region, abnormality is usually considered for an early diagnosis of CD. In order to validate the result, the proposed model is compared with the shallow models and DL approaches. Results found that the proposed method achieved a promising performance with 99.73% accuracy, 91.20% sensitivity, 93.60% precision, 99.80% specificity, and a 91.80% F1-Score. Moreover, both the Receiver Operating Characteristic (ROC) curve and the Precision-Recall (PR) curve from the confusion matrix showed that the developed model is a good classifier. The developed model based on unsupervised feature extraction and deep neural network is ready to be used on a large population before its installation for clinical usage.

A cascaded classifier for multi-lead ECG based on feature fusion

Background and ObjectiveElectrocardiogram (ECG) is an important diagnostic tool for the diagnosis of heart disorders. Useful features and well-designed classification method are crucial for automatic diagnosis. However, most of the contributions were in single lead or two-lead ECG signal and only features from single lead were used to classify the ECG beats. In this paper, a cascaded classification system is proposed to extract features and classify heartbeats in order to improve the performance of ECG beat classification via multi-lead ECG. MethodsIn contrast with most of the literatures, ten signal features were chosen and run on each of the 12 leads separately. Based on these features, we developed a novel feature fusion method combining information from all available leads, and then implemented a cascaded classifier utilizing random forest (RF) and multilayer perceptron (MLP). Besides, in order to reduce the feature space dimension, principal component analysis (PCA) was applied in the method. MaterialsFour open source databases including MIT-BIH Arrythmia Database, QT Database, MIT-BIH Supraventricular Arrhythmia Database and St. Petersburg Institute of Cardiological Technics 12-lead Arrhythmia Database (INCART Database) were used in this work. These four databases are different in classes of beats, volume of dataset, number of individual volunteers. Except INCART database, in which the recording format is 12−lead, the other three databases consist of 2−lead recordings. Above all, they all have annotations for every single beat including the type of each beat. ConclusionsExtensive experimental results shown that the average accuracy achieved 99.3%, 99.8%, 97.6% and 99.6% on four databases respectively. Compared with most state-of-the-art methods, our work has better performance and strong generalization capability.