Elly Matul Imah

A comparative study on Daubechies Wavelet Transformation, Kernel PCA and PCA as feature extractors for arrhythmia detection using SVM

The electrocardiogram (ECG) plays an important role in monitoring and preventing heart attacks. In this paper, we propose and compare the use of Daubechies WT (Daubechies Wavelet Transformation), Kernel PCA (Principal Component Analysis), and PCA as feature extraction methods in improving arrhythmia signals classification. The Kernel PCA employs linear, polynomial, and Gaussian kernels. We examine Support Vector Machines (SVM) pattern classifier with various kernels including wavelet, linear, Gaussian and polynomial. The ECG signals are obtained from MIT-BIH arrhythmia database. The task is to classify or distinguish four different arrhythmias from normal ECG. The overall classification system is comprised of three components including data preprocessing, feature extraction and classification. In data preprocessing which depends on how the initial data is prepared, we reduce the baseline noise with cubic spline and cut the signal beat by beat using pivot R peak. Finally, ECG signal is classified by SVM using various kernels, our experimental results show that wavelet gives better results compared to other feature extraction methods. The accuracy of Wavelet Daubechies for feature extraction is 100% and the best kernel function for the SVM classification is Linier kernel and wavelet kernel.

Adaptive Multilayer Generalized Learning Vector Quantization (AMGLVQ) as new algorithm with integrating feature extraction and classification for Arrhythmia heartbeats classification

Electrocardiogram (ECG) plays an important role in monitoring and preventing heart attacks. In this paper, we propose a new method Adaptive Multilayer Generalized Learning Vector Quantization (AMGLVQ) that integrated feature extraction and classification for the automatic classification of heartbeats in an ECG signal. Since this task has specific characteristics such as, inconsistency optimization on feature extraction and classification, unclassifiable beats and a strong class unbalance, so in this study we proposed new algorithm to handle the problems. The algorithm will be evaluated on real ECG signals from the MIT arrhythmia database. The Experiments show that the proposed method can improve the accuracy of classification better than SVM or back-propagation NN and also able to handle some problems of heartbeat classification: unbalance class, inconsistency between feature extraction and classification and detecting unknown beat on testing phase.

Developing smart telehealth system in Indonesia: Progress and challenge

Indonesia is developing country with high population. There are more than 200 million residents living in the country. As a developing country, Indonesia has several health problems. First, Indonesia has a high value of mortality caused by heart and cardio vascular diseases. One of the major cause is the lack of medical checkup especially for heart monitoring. It is caused by limited number of medical instrumentation e.g. ECG in hospital and public health center. The supporting factor is the small number of cardiologist in Indonesia. There are 365 cardiologists across the country, which is a very small number compared to the 200 million of Indonesia population. Furthermore, they are not distributed evenly in all provinces, but only centered in Jakarta and other capital cities. Therefore, it is difficult for residents to get appropriate heart monitoring. Second, the mortality rate of mother and baby during delivery of the baby in Indonesia is also high. One way to solve this problem is to devise a system where the health clinics in rural areas can perform fetal biometry detection before consulting the results to the expert physicians from other areas. The proposed system will be equipped with algorithms for automatic fetal detection and biometry measurement. By the end of this development, we have several results, the first is a classifier to automatic heartbeat disease prediction with accuracy more than 95%, the second is compression method based on wavelet decompositon, and the third is detection and approximation a fetus in an ultrasound image with hit rate more than 93%.

Support vector machine with multiple kernel learning for image retrieval

Content-based image retrieval (CBIR) is any technology that helps to organize digital image archives by their visual content, with the rapid growth of image amount, the presence of image retrieval is very helpful for people now. Basically image retrieval is part of classification problem. Support Vector Machine (SVM) is one of technique that can solve classification problem well. In fact, SVM is a linear classifier, its mean that this algorithm can only be used to classify linear separable data. In order to classify not linear separable data, this algorithm should be modified with kernel learning. Determine the appropriate kernel during the learning process is difficult, therefore, many researchers are trying to develop more flexible kernel learning called Multiple Kernel Learning (MKL). This framework will build a classification model based on SVM algorithm modified with multiple kernel learning and applied to image retrieval. The image retrieval contain five categories of image and the experimental uses k-fold cross-validation. The result show that SVM with multiple kernel learning has good accuracy with 78 % and also has sort computation time, where it needs about 64.35 seconds for training session and 26.15 seconds for retrieve session.