Ewaryst Tkacz

An application of wavelet neural network for classification of patients with coronary artery disease based on HRV analysis

Presents some recently obtained results concerning the possibility of application of wavelet neural networks (WNN) for classification purposes in the case of patients with coronary artery disease of different levels. Patients with respectively one, two and three coronary arteries blocked have been taken into consideration. The Heart Rate Variability signal has been registered for 5 minutes for each of such patients. All the patients have been previously preliminary classified by an experienced cardiologist with regard to the estimation of the number of coronary arteries blocked. Then half of each HRV record has been applied for teaching the neural network after features selection from raw HRV through the application of a wavelet transform being the first layer of the WNN system. The second half of data has been used for classification. Due to the fact that four classification groups were expected the output layer of the neural network has only two output neurons.

Feature extraction for improving the support vector machine biomedical data classifier performance

A support vector machine (SVM) is a relatively novel classifier based on the statistical learning theory. To increase the performance of classification, presented study focuses on the mixed domain (time&frequency) feature extraction preliminary to SVM application. Time and frequency domain selected features and discrete fast wavelet transform coefficients parameters including energy and entropy measures were the component of new feature vector. SVM classifier structure were adjusted by the selection of optimal for analysed application its kernel functions:both polynomial and radial basis functions. System was positively verified on the set of clinically classified ECG signals for control and atrial fibrillation (AF) disease patients taken from MITBIH data base. The measures of specificity and sensitivity computed for the set of 20 AF and 20 patients from control group divided into learning and verifying subsets were used to evaluate presented pattern recognition structure. Different types of wavelet basic function for feature extraction stage were tested to find the best system structure. Obtained results showed, that the ability of generalization for enriched feature extraction (FE)-SVM based system increased, due to selectively choosing only the most representative features for analyzed AF detection problem.

Supervised and Unsupervised Learning Systems as a Part of Hybrid Structures Applied in EGG Signals Classifiers

This paper aims at investigating an unsupervised learnt neural networks in classifier applications and comparing them to supervised perceptron type nets. The proposed solutions focus on combing the time-frequency preliminary analysis by means of wavelet transform with application of self organizing maps. Using wavelet transform as a feature extraction tool allowed to reveal important parameters included both in time and frequency domain of non-stationary electrogastrographic signals, which were classified in elaborated systems. Proposed structures were tested using the set of clinically characterized EGG signals of 62 patients, as cases with different level rhythm disturbances from bradygastria up to tachygastria together with some artifacts of non-stationary character such as muscle thrill etc. Additionally similar control group of healthy patients was analyzed. The results of the proposed methodology are illustrated in the measure of sensitivity and specificity, where the best classifier based on Kohonen maps with preliminary wavelet processing reached the performance above 90%

A new parallel concept for QRS complex detector

The paper presents a new approach to the QRS complex detection problem in noisy ECG signal. Complexes were distinguished into two classes: supraventricular and ventricular. For each class of QRS we introduced special filtering using matching filters facilities. The processing of the real time ECG incoming signal we performed in parallel for filters and decision concerning presence of particular complex was based on the output of these filters.

Feature extraction based on time-frequency and Independent Component Analysis for improvement of separation ability in Atrial Fibrillation detector

Due to redundancy of over-dimensioned information, observed often in originally recorded biomedical signals, feature extraction and selection has become focus of much researches connected with biomedical signal processing and classification. Mixed new feature vector combined from time-frequency signal representation (obtained after wavelet transform) and Independent Component Analysis (ICA) applied for non-stationary signals is proposed as a preliminary stage in ECG waveform classification for patients with Atrial Fibrillation (AF). Discrete fast wavelet transform coefficients parameters including energy and entropy measures and components extracted as a result of FastICA algorithm implementation after optimization gave the best classifier performance of whole AF ECG classifier system. System was positively verified on the set of clinically classified ECG signals for control and atrial fibrillation (AF) disease patients taken from MITBIH data base. The measures of specificity and sensitivity computed for the set of 20 AF and 20 patients from control group divided into learning and verifying subsets were used to evaluate presented pattern recognition structure. Different types of wavelet basic functions for feature extraction stage and kernels for SVM classifier structure calculation were tested to find the best system architecture. Obtained results showed, that the ability of generalization and separation for enriched feature extraction based system increased, due to selectively choosing only the most representative features for analyzed AF detection problem.

A new approach to the P-wave detection and classification based upon application of wavelet neural network

Presents a new approach to the P wave classification problem which is based upon the application of a new recently developed and widely described tool such as the wavelet neural network. The novel idea of classification is based on the creation of our own non-standard wavelet exactly as a P wave morphology template and then calculation of the wavelet transform as a first layer of a classical multi-layer perceptron. This first layer works as a feature selector and extractor.

Feature Selection of Protein Structural Classification Using SVM Classifier

Recursive feature elimination method (RFE), cross validation coefficient (CV) and accuracy of classification of test data are applied as a criterion of feature selection in order to find relevant features and to analyze their influence on classifier accuracy. Feature selection method was compared to principal component analysis (PCA) to understand the effectiveness of feature reduction. Support vector machine classifier with radial basis function (RBF) kernel is applied to find the best set of features using grid model selection and to select and assess relevant features. The best selected feature set is then analyzed and interpreted as the source of knowledge about the protein structure and biochemical properties of amino acids included in the protein domain sequence.

Statistical Analysis of the Impact of Molecular Descriptors on Cytotoxicity of Thiourea Derivatives Incorporating 2-Aminothiazole Scaffold

Chemical reactivity descriptors and lipophilicyty (log P) were evaluated via semi-empirical method for the quantum calculation of molecular electronic structure (PM3) in order to clarify the structure-cytotoxic activity relationships of disubstutited thioureas. Analysed compounds were obtained by the linkage of 2-aminothiazole ring, thiourea and substituted phenyl ring. The detailed examination was carried out to establish correlation between descriptors and cytotoxic activity against the MT-4 cells for 11 compounds. For the most active compounds (6 compounds) cytotoxic activity against three cancer cell lines (CCRF-CEM, WIL-2NS, CCRF-SB) and normal human cell (HaCaT) was determined. 3-(4,5-Dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) reduction and lactate dehydrogenase (LDH) release were assessed. Regression analysis revealed that electrophilicity index and chemical potential significantly contributed to expain the thioureas cytotoxic potential.

A new method for attenuation of respiration artifacts in electrogastrographic (EGG) signals

Electrogastrography (EGG) is a test method designed for noninvasive assessment of gastric slow waves propagation. The EGG signal is obtained from the electrodes respectively arranged on the surface of the patients abdomen. A significant problem during recording of the EGG signal is the elimination of disturbances occurring during registration and unwanted components of other signals such as: components of electrocardiographic (ECG), baseline drift or respiratory disturbances. These components are generally present in the signals registered from the surface of the abdomen of the patient. Since EGG frequency components partly overlap with the frequency components of respiratory artifacts, conventional band-pass digital or analog filtering may cause distortion in electrogastrographic signal. In the paper a method for removing respiratory interference occurring during registration of EGG signal and the effect of filtration on selected parameters of EGG signal analysis is presented. Respiratory artifacts are removed through the use of adaptive filter working in the DCT domain. The applied adaptive filtering method involves the use of the signal including respiratory disturbances. This signal is recorded synchronously with the EGG signal using a thermistor placed near the nose of the patient.

Feature extraction in time-frequency signal analysis by means of matched wavelets as a feature generator

The goal of presented work was to compare the usage of standard basic wave let function like e.g. bio-orthogonal or dbn with the optimized wavelet created to the best match analyzing ECG signals in the context of P-wave and atrial fibrillation detection. A library of clinical expert evaluated typical atrial fibrillation evolutions was created as a database for optimal matched wavelet construction. Whole data set consisting of 40 cases with long term ECG recording s were divided into learning and verifying set for the multilayer perceptron neural network used as a classifier structure. Compared with other wavelet filters, the matched wavelet was able to improve classifier performance for a given ECG signals in terms of the Sensitivity and Specificity measures.