Damla Kuntalp

Feature extraction for ECG heartbeats using higher order statistics of WPD coefficients

This paper describes feature extraction methods using higher order statistics (HOS) of wavelet packet decomposition (WPD) coefficients for the purpose of automatic heartbeat recognition. The method consists of three stages. First, the wavelet package coefficients (WPC) are calculated for each different type of ECG beat. Then, higher order statistics of WPC are derived. Finally, the obtained feature set is used as input to a classifier, which is based on k-NN algorithm. The MIT-BIH arrhythmia database is used to obtain the ECG records used in this study. All heartbeats in the arrhythmia database are grouped into five main heartbeat classes. The classification accuracy of the proposed system is measured by average sensitivity of 90%, average selectivity of 92% and average specificity of 98%. The results show that HOS of WPC as features are highly discriminative for the classification of different arrhythmic ECG beats.

A multi-stage automatic arrhythmia recognition and classification system

This paper describes an automatic classification system based on combination of diverse features for the purpose of automatic heartbeat recognition. The method consists of three stages. At the first stage, heartbeats are classified into 5 main groups defined by AAMI using optimal feature sets for each main group. At the second stage, main groups are classified into subgroups using optimal features for each subgroup. Then the third stage is added to the system for classifying beats that are labeled as unclassified beats in the first two classification stages. A diverse set of features including higher order statistics, morphological features, Fourier transform coefficients, and higher order statistics of the wavelet package coefficients are extracted for each different type of ECG beat. At the first stage, optimal features for main groups are determined by using a wrapper type feature selection algorithm. At the second stage, optimal features are similarly selected for discriminating each subgroup of the main groups. Then at the third stage, only raw data is used for classifying beats. In all stages, the classifiers are based on the k-nearest neighbor algorithm. ECG records used in this study are obtained from the MIT-BIH arrhythmia database. The classification accuracy of the proposed system is measured by sensitivity, selectivity, and specificity measures. The system is classified 16 heartbeat types. The measures of proposed system are 85.59%, 95.46%, and 99.56%, for average sensitivity, average selectivity, and average specificity, respectively.

Arrhythmia classification using higher order statistics

In this work, the features are extracted for the arrhythmia classification from the electrocardiograph (ECG) signals by using Higher order statistics. K-nearest neighborhood algorithm is used as classifier. Cumulants are calculated from the raw signals obtained from consecutive sample values of each R peak in ECG signals and used as features. In addition to these features, different features obtained from the relations of cumulants are also used. Simulation results shows that features obtained from the relations among cumulants are more discriminative than the cumulants.

Topographic analysis of morphologic features of ECG beats

In this work, the arrhythmias in the electrocardiograph (ECG) signals are analyzed by using Self Organizing Maps (SOM). Morphologic features obtained from consecutive sample values of each R peak are used for training the SOM networks. The maps are examined using U-matrix representation method. Consequently, the high dimensional data are examined in two dimensions. When the shapes of the distributions obtained by U-matrix representation are considered, it is realized that a simple linear classifier is not able to classify these patterns correctly.

Effect of the Input Window Size in Arrhythmia classification with Multilayer Perceptron Network Structures

In this work, the arrhythmias in the electrocardiograph (ECG) signals are analyzed by using multi-layer perceptron (MLP) network. For training MLP network back-propagation with adaptive learning rate method is used. Feature vectors obtained from consecutive sample values of each peak in different window sizes are normalized and used for training the networks. Performances of different classifiers are examined depending on the average value of sensitivity, specificity, selectivity and accuracy of the classifiers. The results show that for the proposed classifier the optimal feature vector is a 71-point vector with 35 before and 35 after the R peak point of the ECG.

Feature reduction method using self organizing maps

In this work, five main groups of arrhythmias in electrocardiograph (ECG) signals are tried to be classified using the features obtained from the output of a Self Organizing Map (SOM) network. The raw ECG signal consists of 81 sample points (60 point before and 20 point after the R peak point of the ECG). Consecutive sample values of a moving window (20 points of width) are used as the input vector of the SOM network. The output of the SOM network is used as the input vector to a classifier. K-nearest neighbor (k-NN) algorithm is chosen as the classifier. The performance of the classifier is evaluated by the average values of sensitivity, specificity, selectivity and overall accuracy. As a result, 96%, 91%, 99%, and 97% sensitivity, selectivity, specificity, and overall accuracy values are obtained.

Classification of arrhythmias using spectral features with K Nearest Neighbor method

In this work, five types of arrhythmias observed in electrocardiograph (ECG) signals are analyzed by using their spectral features. K-Nearest Neighbor (KNN) method is used as the classifier. The frequency spectrum of the samples are divided into a variable number of distinct bands and average band powers are used as the feature vectors. The performance of the classifier is tested by changing the width of the frequency bands, the number of neighbors and distance metric. The results are examined based on the average sensitivity, specificity, selectivity and accuracy values. The results show that the optimal KNN classifier is the one which uses 1 nearest neighbor, cityblock distance metric and 0.7Hz width frequency band.

Genetic Algorithm Assisted Radial Basis Function Multiuser Detector for DS-CDMA

The optimal receiver for detecting the direct sequence code division multiple access (DS-CDMA) signals suffers from computational complexity which increases exponentially with the increasing number of users. Thus several sub-optimal multiuser detectors (MUD) are proposed. Radial basis function (RBF) MUD is one of these sub-optimal receivers which has a very high performance but which still suffers from computational complexity since the number of center functions increases exponentially with the increasing number of users. In this contribution, we propose a new method to minimize the number of center functions of RBF MUD using genetic algorithm (GA) and least mean squares (LMS) algorithm. Simulation results showed that the proposed method immensely reduced the complexity of the RBF MUD and the MUD was capable of succesfully tracking the bit error rate of the single-user detector.

Musical Genre Classification Using Higher-Order Statistics

In this study, we examine the effects of higher order statistics of timbral features to improve performance of genre classification. It was seen that the first and second order statistics of the features extracted, in this research, is not as discriminative as the third and forth order statistics of the features. For the purpose of designing a classifier, which could be used for real time applications in future studies, randomly taken 3 second-long segments are used for classification. Out of 225 songs from 3 genres, ISO of them are used for training and 45 of them are used for testing. Five different lists that are created using different train and test sets are used to reduce the dependency of the results to the test set while increasing the number of validation data. Average values of validation test results are compared with the results of the similar works, which are based on MIDI format, using the same data set.

New methods of density estimation for vehicle traffic

This study is prepared to produce solutions for estimating vehicle traffic density which is one of the biggest problems of urban life today. Four different algorithms are proposed for density estimation problem with different perspectives. All these proposed algorithms have been tried to estimate next state of the road by looking at history of density data. Algorithms are inspired by the methods used to estimate spectrum holes in cognitive radio channels. A similar approach is used for estimation of traffic density. In this study, Istanbul Metropolitan Municipality Traffic Control Center data received from the busiest roads of Istanbul in 2013, have been used as traffic data. Different simulations have been performed using these algorithms and results are evaluated based on several performance criteria.