Classification of Arrhythmia on ECG signal with Stacked Auto Encoder

Abstract

Computerized systems that can be employed as a screening tool for the classification of cardiac arrhythmias play an important role not only for the patients but also can help the doctors in the decision making of correct treatment selection. For recognizing cardiac arrhythmia in ECG signals, common signal processing techniques such as pre-processing, segmentation, feature extraction, feature selection and classification are crucial. The purpose of this invention is to provide an appropriate model for classification of arrhythmia into different classes using ECG signal.

Specification

Description:FIELD OF INVENTION
[001] Artificial Intelligent Technology.
BACKGROUND AND PRIOR ART
[002] The invention of efficient methods for predicting CVD received a lot of interest from the scientific community in the last decades. Several studies have demonstrated that arrhythmias are the primary causes of the majority of CVD cases. So, detection of arrhythmias is essential.
[003] Ours is a data-driven, technologically advanced society. The usage of Information Technology (IT) has an enormous impact on society as a whole and the development of the nation's economy. Thanks to advancing technology, since improvements are possible in this domain.
[004] Since ECG signals are typically non-stationary, it is particularly challenging to observe disease symptoms. Hence, the ECG patterns are recorded for longer than a hour to detect the heart irregularities associated with arrhythmia. A lot of data is produced when the ECG signals are recorded. Thus, it becomes difficult and time-consuming for medical specialists to identify irregularities in ECG signals. Also manual heartbeat classification of long-term ECG recording requires a lot of practice for junior practitioners. Therefore, automated computer-based methods are required for the diagnosis of different arrhythmias from the massive volume of data. The development of the methodology utilized in the study of arrhythmias is probably going to help cardiologists in the diagnosis and screening of CVD.
[005] A wide range of currently available techniques were employed in various studies for the categorization of arrhythmias. Many machine learning techniques, including the Ensemble Method, Artificial Neural Network (ANN), Support Vector Machine (SVM), Conditional Random Field (CRF), Self-Organizing Map (SOM), and Linear Discrimination Analysis (LDA), were used for the arrhythmia classification process. Existing research approaches for classification of arrhythmias had limitations of imbalanced data and overfitting problems.
SUMMARY OF THE INVENTION
[006] Arrhythmia classification helps to analyze the condition of the heart and is a vital support to the doctor in the decision making of correct treatment selection for arrhythmia patient. Since ECG signals are typically non-stationary, it is particularly challenging to observe disease symptoms. The development of the methodology used in the study of arrhythmias is probably going to help cardiologists in the diagnosis and screening of CVD.
BRIEF DESCRIPTIONS OF DRAWINGS
[007] A brief description of the several views of the drawings. It consists of 4 stages. At the pre-processing stage, the input ECG signal is subjected to Normalization, Filtering and 'R' peaks detection operations. Figure 1 depicts the overview of the proposed framework for arrhythmia classification.
[008] During the second stage, Wavelet, Burg, Yule, Entropy and Statistical related attributes are extracted using feature extraction techniques. The next stage is feature selection. For feature selection the optimization techniques namely Whale Optimization Algorithm (WOA), Grey-Wolf Optimization Algorithm (GWOA) and Grasshopper Optimization Algorithm (GOA) are used. In this study to choose the most pertinent features from the best solutions of the WOA, GWOA and GOA the ensemble feature selection method based on feature correlation is proposed for arrhythmia and ventricular arrhythmia classification. The advantage of the suggested ensemble feature selection approach is that it may choose the pertinent characteristics from the finest outcomes of the GWOA, GOA and WOA methods. For the classification of arrhythmia, classifiers such as the Deep Stacked Autoencoder method is employed.
DETAILED DESCRIPTION OF THE INVENTION
[009] The performance of the classification depends not only on the classifier but also impacted by the variations in the original data. At the pre-processing stage, the input ECG signal is subjected to Normalization, Filtering and 'R' peaks detection operations. To reduce high-frequency noise and to remove baseline wander, a second-order Butterworth low pass filter with a 30 Hz cutoff frequency is used. The well-known Pan Tompkins Algorithm (PTA) is used to accomplish the R-peak detection.
[010] The pre-processed ECG signals are used as input x for feature extraction. From the input data, the Discrete Wavelet Transform, Burg's, Yule features, Entropy features, and Statistical features are retrieved. To choose the most pertinent features, the extracted features are applied for ensemble feature selection process.
[011] The entropy features are extracted from the pre-processed signals such as Renyi Entropy, Tsallis Entropy, Sample Entropy, Differential Entropy. The statistical features are extracted from the pre-processed signals such as Variance, HOS Cumulative Features, Skewness, kurtosis.
[012] The hybrid methods of WOA, GOA and GWOA are used with the extracted features for feature selection. Poor convergence is the limitation of the feature selection method such as Backtracking Search Optimization (BSO) and Lower exploitation is the main limitation of the feature extraction method such as Fruitful Optimization Algorithm (FOA). The WOA approach avoids trapping in local optima in feature selection. Also the GWO and GOA perform better in exploitation and have lower convergence. Higher exploitation and exploration are possible with the help of the GWOA, GOA, and WOA ensemble approach. The proposed ensemble feature selection method of GWOA, GOA and WOA provide high exploitation and exploration.
[013] The ensemble feature selection method determines the correlation between the chosen features in three optimization methods and chooses the features with the highest correlation for classification. The association between the chosen features offers pertinent data and aids in avoiding local optimum features. Equation describes how to measure the ensemble feature selection output using the correlation coefficient matrix as shown in Equation.
[014] An autoencoder is a type of unsupervised learning structure that has three layers. i.e. the input layer, the hidden layer, the output layer and is used to perform arrhythmia classification operation. The dropout strategy is used in the neural network training process to minimize the over fitting issue. Rectified Linear Unit (ReLU) activation function is used. The classified output is used to evaluate the efficiency of the model and compare it with the existing techniques such as Random Forest (RF), Multi Support Vector Machine (MSVM), K Nearest Neighbors (KNN), and Decision Tree (DT).
[015] The accuracy of the suggested ensemble feature selection approach is 98.23% and the accuracy of the current SVM with the made by hand feature selection approach is 96.3% in ventricular arrhythmia classification.The accuracy of this present approach is 97.7% percent, whereas the suggested ensemble model has 98.11% percent in arrhythmia classification. The accuracy improvement of the suggested method is 1.93% for ventricular arrhythmia classification and 0.41% for arrhythmia classification. , C , Claims:[016] 1. The invention provides a system and method for detecting arrhythmia using AI technology. Classification of arrhythmia in CAD system, generally undergo signal processing stages such as pre-processing, segmentation, extraction of characters, selection of characters, and classification. A machine learning model enhances detection accuracy by recognizing patterns of ECG signal data. The contribution is concerned with development of a model to classify cardiac abnormalities into four and eight different classes using the MIT-BIH ventricular arrhythmia database (VFDB) and the MIT-BIH arrhythmia database (MITDB). In this approach a Deep Neural Network (DNN) in conjunction with the ensemble feature selection method is applied to classify arrhythmias.

Documents