A NOVEL FRAMEWORK FOR CARDIAC ARRHYTHMIA CLASSIFICATION BASED ON SUBSPACE PROJECTION AND DECISION TREE

International audience

Due to the capacity of processing signal with low energy consumption, compressive sensing (CS) has been widely used in wearable health monitoring system for arrhythmia classification of electrocardiogram (ECG) signals. However, most existing works focus on compressive sensing reconstruction, in other words, the ECG signals must be reconstructed before use. Hence, these methods have high computational complexity. In this paper, the authors propose a cardiac arrhythmia classification scheme that performs classification task directly in the compressed domain, skipping the reconstruction stage. The proposed scheme first employs the Pan–Tompkins algorithm to preprocess the ECG signals, including denoising and QRS detection, and then compresses the ECG signals by CS to obtain the compressive measurements. The features are extracted directly from these measurements based on principal component analysis (PCA), and are used to classify the ECG signals into different types by the proposed semi-supervised learning algorithm based on support vector machine (SVM). Extensive simulations have been performed to validate the effectiveness of the proposed scheme. Experimental results have shown that the proposed scheme achieves an average accuracy of [Formula: see text] at a sensing rate of 0.7, compared to an accuracy of [Formula: see text] for noncompressive ECG data.

Electrocardiogram (ECG) signals consist of data containing measurements of electrical activity in the heartbeats. These signals include relevant information used to detect abnormalities such as arrhythmia. In this study, a recognition system is proposed for detection and classification of heartbeats in ECG signals. Heartbeats in the ECG data were detected by using the wavelet transform (WT) method and these beats are segmented with determined periods. For obtaining distinctive features from the beats, multi-resolution WT is applied to these segmented signals, and wavelet coefficients are obtained from different frequency levels. Feature vectors are generated on these coefficients by using various statistical methods. The proposed recognition system is trained on feature vectors by using the Online Sequential Extreme Learning Machine (OSELM) classifier during the learning phase to automatically recognize the signals. Five different beat types were obtained from the MIT-BIH arrhythmia dataset. The multi-class dataset that includes five classes and the binary-class dataset that includes two classes were created among these beat types. Performance tests of the proposed wavelet-based-OSELM (W-OSELM) method were realized with these two datasets. The proposed recognition system provided 97.29% correct beat detection rate from raw ECG signals. The classification accuracy is 99.44% for the binary-class dataset and 98.51% for the multi-class dataset. Furthermore, the proposed classifier has shown very fast recognition performance on ECG signals.

A heart disease called cardiovascular diseases (CVD) is another leading cause for the death. There are several reasons that lead the CVD in human beings. The early detect of CVD helps to take necessary medical attentions to prevent the harms. The conventional techniques for CVD detection were manual and expensive which often delivers the inaccurate diagnosis. Since from the last decade the other inexpensive Computer Aided Diagnosis (CAD) based methods gained significant medical attentions. The CAD based techniques mainly based on raw Electro Cardiogram (ECG) signals of patient for the accurate and economical detection of CVD at early stage. In recent past, there are several CAD systems designed for CVD diagnosis utilizing raw ECG signals, however accuracy of CVD detection utilizing ECG bothered through several issues of research like QRS beats extraction, artefacts, efficient features extraction. This research paper present CVD novel framework, utilizing raw ECG signals and designed hybrid pre-processing algorithm for extracting artefacts and noise through raw ECG signal. Further designed simple and efficient dynamic thresholding based technique to extract the beats such as Q, R, S, and ST segment through pre-processing ECG signal. Third step perform the fusion of extracted beats and apply the feature extraction method called Normalized Higher Order Statistic (NHOS). The normalized HOS techniques asses the complexity among all the QRS based beats and delivers the more unique features for the accuracy enhancement. The final step is the classification by using five different classifiers for the CVD detection. The simulation results presented in this paper demonstrate that proposed framework achieved the significant accuracy improvement.

This study aims to present an efficient model for autodetection of cardiac arrhythmia by the diagnosis of self-affinity and identification of governing processes of a number of Electrocardiogram (ECG) signals taken from MIT-BIH database. In this work, the proposed model includes statistical methods to find the diagnosis pattern for detecting cardiac abnormalities which is useful for the computer aided system for arrhythmia detection. First, the Rescale Range (R/S) analysis has been employed for ECG signals to understand the scaling property of ECG signals. The value of Hurst exponent identifies the presence of abnormality in ECG signals taken for consideration with 92.58% accuracy. In this study, Higuchi method which deals with unifractality or monofractality of signals has been applied and it is found that unifractality is sufficient to detect arrhythmia with 91.61% accuracy. The Multifractal Detrended Fluctuation Analysis (MFDFA) has been used over the present signals to identify and confirm the multifractality. The nature of multifractality is different for arrhythmia patients and normal heart condition. The multifractal analysis is useful to detect abnormalities with 93.75% accuracy. Finally, the autocorrelation analysis has been used to identify the prevalent governing process in the present arrhythmic ECG signals and study confirms that all the signals are governed by stationary autoregressive methods of certain orders. In order to increase the overall efficiency, this present model deals with analyzing all the statistical features extracted from different statistical techniques for a large number of ECG signals of normal and abnormal heart condition. Finally, the result of present analysis altogether possibly indicates that the proposed model is efficient to detect cardiac arrhythmia with 99.3% accuracy.

Multi-stream Bi-GRU network to extract a comprehensive feature set for ECG signal classification

This paper presents an algorithm formulated to identify the atrial fibrillation complications through electrocardiogram (ECG) signals. The ECG data for the study was retrieved from Physio Net which consists of normal, atrial fibrillation and other rhythms. The Discrete Wavelet Transform (DWT) was used to remove baseline wanders. Pan Tompkins algorithm was utilized to detect the P, Q, R, S and T peak and thus the ECG signals were segmented based on each cycle. The morphological features were extracted directly from the time-series while statistical features were extracted after Stockwell transform (S-transform) was applied to the data. Genetic Algorithm (GA) and reliefF algorithm have been applied separately to select the optimum features for classification purpose. Bagged Tree ensemble algorithm, Decision Tree and k-Nearest Neighbour (KNN) algorithm were used as classifiers to identify atrial fibrillation through ECG signals. The classification results with and without feature selection techniques are presented. Prior to the feature selection, Bagged Tree is the classifier best performing classifier with 86.50% of accuracy, 84.38% of sensitivity and 91.94% of specificity. After feature selection, all the three classifiers have almost the same performance which is nearly 100% of accuracy, sensitivity and specificity. This shows that the proposed combinations of algorithms are reliable and able to improve the identification rate of the normal, atrial fibrillation and other rhythms using lesser number of features.

The article is devoted to the application of machine learning methods for computerized detection of sleep apnea episodes based on the analysis of single-channel signals of the electrocardiogram (ECG) and electroencephalogram (EEG). To study the possibilities of machine learning to detect apnea based on ECG and EEG analysis, we used Apnea-ECG database and MIT-BIH polysomnographic database from PhysioNet, which contain annotations to each minute of records indicating the presence or absence of apnea/hypopnea at the current time. In order to apply machine learning methods to the problem of automated detection of sleep apnea/hypopnea episodes in ECG and EEG signals, long-term polysomnograms available in MIT-BIH polysomnographic database were segmented according to annotations into shorter sections lasting 30 seconds each. The study used 267 segments lasting 30 seconds for the class "norm", 258 segments for the class "apnea" and 273 segments for the class "hypopnea", a total of 798 simultaneous ECG and EEG recordings. The aim of this work is to identify and compare informative signs of sleep apnea episodes in terms of heart rate variability (HRV) and brain electrical activity, as well as the choice of classification methods that provide the highest accuracy for this task. Features of cardiorhythmograms in time and frequency domains, spectral-temporal and wavelet characteristics, as well as parameters of EEG signals based on energy ratio of EEG rhythms, Hearst index, Higuchi fractal dimension and sample entropy for EEG signals are considered. Using different sets of features, the accuracy of classifiers based on decision trees, discriminant analysis, support vector machines, k-nearest neighbor method, and ensemble training was determined. Based on this, combination of features and classifiers is proposed, which provides the highest accuracy of recognition of sleep apnea episodes according to single-channel ECG and EEG signals, taken separately and in the case of a combination of their features. The best results of classification of signals "norm", "apnea" and "hypopnea" were obtained for the model trained using weighted method k nearest neighbors with 25 features of HRV: the total percentage of correctly identified cases for three classes was 99.9% (797 correctly identified cases of 798). By reducing the number of HRV parameters to 9, the best machine learning result was achieved using the bagging ensemble algorithm with 30 decision trees: the total percentage of correctly identified cases for all three classes was 99.4% (793 correctly identified cases from 798: for "norm" - 265 cases from 267, for "apnea" - 257 cases from 258, for "hypopnea" - 271 cases from 273). The use of EEG parameters as features for apnea/hypopnea recognition showed worse results compared to HRV parameters. In this case, the best result of machine learning was achieved using support vector machines with quadratic kernel function: the total percentage of correctly identified cases for three classes was 91.9% and the signals corresponding to norm were most badly recognized (27 cases were classified as hypopnea, and in 9 cases - as sleep apnea). The combination of HRV and EEG parameters gave the best accuracy of 99.1%, but the results are comparable to using only HRV parameters. The obtained results indicate that HRV parameters allow recognizing sleep apnea and hypopnea with higher accuracy than EEG parameters, but EEG signal undoubtedly reflects signs of sleep apnea/hypopnea and also can be used for apnea recognition.

Globally, cardiovascular disease kills more than 500000 people every year, thus becoming the primary reason for death. Nevertheless, cardiovascular health monitoring is essential for accurate analysis and therapy of heart disease. In this work, a novel deep learning-based StrIppeD NAS-Network (SID-NASNet) for arrhythmia categorization into octa-classes with electrocardiogram (ECG) signals is presented. First, the ECG signals are recorded in real time using 12-lead electrodes. Then, the Discrete Wavelet Transform (DWT) is used to denoise the signals to reduce repetition and increase resilience. The noise-free ECG signals are fed into a K-means clustering algorithm to group ECG signal segments into a set number of clusters to identify patterns that may indicate heart abnormalities. Subsequently, the deep learning-based NASNet with Stripped convolutional layers is used to detect ECG irregularities of arrhythmia. Each sample point is examined for its local fractal dimension before extracting the heartbeat waveforms within a predetermined window length. A bio-inspired Dingo Optimization (DO) algorithm is used in the SID-NASNet to normalize the parameters to improve the efficiency of the network with low network complexity. The efficiency of the proposed SID-NASNet is assessed with specificity, accuracy, precision, F1 score and recall based on the MIT-BIH arrhythmia dataset. From the test results, the proposed SID-NASNet achieves an accuracy of 98.22% for effective categorization of ECG signals. The proposed SID-NASNet improves the overall accuracy of 1.24%, 3.76%, 1.87%, and 0.22% better than ECG-NET, Deep Learning (DL)-based GAN, 1D-CNN, and GAN-Long-Short Term Memory (LSTM), respectively.

A novel method for reducing arrhythmia classification from 12-lead ECG signals to single-lead ECG with minimal loss of accuracy through teacher-student knowledge distillation

Electrocardiogram (ECG) signal plays a key role in the diagnosis of arrhythmia, which will pose a great threat to human health. As an effective feature extraction method, deep learning has shown excellent results in processing ECG signals. However, most of these methods neglect the cooperation between the multi-lead ECG series correlation and intra-series temporal patterns. In this work, a multi-domain collaborative analysis and decision approach is proposed, which makes the classification and diagnosis of arrhythmia more accurate. With this decision, we can realize the transition from the spatial domain to the spectral domain, and from the time domain to the frequency domain, and make it possible that ECG signals can be more clearly detected by convolution and sequential learning modules. Moreover, instead of the prior method, the self-attention mechanism is used to learn the relation matrix between the sequences automatically in this paper. We conduct extensive experiments on eight advanced models in the same field to demonstrate the effectiveness of our method.

Electrocardiogram (ECG) signal quality indices (SQIs) are essential for improving diagnostic accuracy and reliability of ECG analysis systems. In various practical applications, the ECG signals are corrupted by different types of noise. These corrupted ECG signals often provide insufficient and incorrect information regarding a patient’s health. To solve this problem, signal quality measurements should be made before an ECG signal is used for decision-making. This paper investigates the robustness of existing popular statistical signal quality indices (SSQIs): relative power of QRS complex (SQIp), skewness (SQIskew), signal-to-noise ratio (SQIsnr), higher order statistics SQI (SQIhos) and peakedness of kurtosis (SQIkur). We analysed the robustness of these SSQIs against different window sizes across diverse datasets. Results showed that the performance of SSQIs considerably fluctuates against varying datasets, whereas the impact of varying window sizes was minimal. This fluctuation occurred due to the use of a static threshold value for classifying noise-free ECG signals from the raw ECG signals. Another drawback of these SSQIs is the bias towards noise-free ECG signals, that limits their usefulness in clinical settings. In summary, the fixed threshold-based SSQIs cannot be used as a robust noise detection system. In order to solve this fixed threshold problem, other techniques can be developed using adaptive thresholds and machine-learning mechanisms.

Detecting epileptic seizure onsets is the main goal of numerous studies, since it has many profits for patients and clinicians. Methods based on electroencephalogram (EEG), electrocardiogram (ECG), and other electrophysiological signals had been used for automatic detection in the literature. For the first time, absence seizures have been detected based on ECG signals in this study. Animal models of absence epilepsy, WAG\Rij rats, with repetitive seizures (duration about few seconds'), have been investigated. After detecting QRS complexes from ECG signal and extracting 38 different linear, nonlinear and frequency domain features from heart rate variability, feature vectors were constructed. In order to obtain high efficiency detection algorithm, feature selection have been implemented based on wrapper approach. Results related to support vector machine (SVM), linear discriminate analysis (LDA), and k-nearest neighbor (kNN), three important classifiers for seizure detection have been compared in this work. The test results for patient- independent detection with 5 selected features in leave-one-out (LOO) train approach had accuracy of 74%, 72% and 71% for SVM, LDA and kNN, respectively. All the algorithms and methods have been optimized to be useful in embedded implementations.

Aim: The aim of the work is to evaluate the accuracy and precision in predicting heart disease using Support Vector Machine (SVM), Random forest (RF), Logistic Regression (LR), Decision Tree (DT) Classification algorithms. Materials and Methods: Classification algorithm is appealed on a heart dataset which consists of 180 records. A framework for heart disease prediction in the medical sector comparing Random forest, Logistic Regression, Decision Tree and SVM classifiers has been proposed and developed. The sample size was calculated as 55 in each group using G power 80%. Sample size was calculated using clincalc analysis, with alpha and beta values 0.05 and 0.5, 95% confidence, pretest power 80% and enrolment ratio 1. Results: The Novel Random Forest Algorithm (92.13%), Support Vector Machine (62.51%), Logistic Regression (84.89%), Decision Tree (86.25%) classifiers produce respectively. SVM, RF exists a statistically significant difference between the two groups (p=0.001,p=.004;p<0.05).LR, RF exists a statistically insignificant difference between the two groups (p=.103, P=.080;p>0.05) both with confidence interval 95%. Hence Random forest is better than SVM, RF, DT classifiers. Conclusion: The results show that the performance of RF is better when compared with SVM, LR and DT in terms of both precision and accuracy.

The paper is devoted to the application of machine learning methods for epileptic seizures prediction based on features of electroencephalogram signal (EEG) and heart rate variability (HRV) parameters derived from simultaneously registered electrocardiogram signal (ECG). To study the possibilities of machine learning to predict epileptic seizures based on EEG and HRV analysis, open databases were used which contain annotations indicating the presence or absence of epileptic seizures at each minute of time. The goals of this work are prediction of epileptic seizures in terms of ictal, pre-ictal and interictal states recognition using features of brain electrical activity and HRV, and the choice of classification methods that provide the highest accuracy for this task. HRV parameters in time and frequency domains, as well as parameters of EEG signals based on energy of EEG rhythms, Hearst index, Higuchi fractal dimension and different estimations of entropy for EEG signals were used to achieve these goals. Using different sets of features, the accuracy of classifiers based on k-nearest neighbors method, decision trees, random forest classifier, multi-layer perceptron, support vector machines, and extreme gradient boosting classifier was determined.