Abstract— Presented paper describes a system of biomedical signal classifiers with preliminary feature extraction stage based on matched wavelets analysis, where two structures of classifier using Neural Networks (NN) and Support Vector Machine (SVM) are applied. As a pilot study the rules extraction algorithm applied for two of mentioned machine learning approaches (NN & SVM) was used. This was made to extract and transform the representation of knowledge gathered in Black Box parameters during classifier learning phase to be better and natural understandable for human user/expert. Proposed system was tested on the set of ECG signals of 20 atrial fibrillation (AF) and 20 control group (CG) patients, divided into learning and verifying subsets, taken from MIT-BiH database. Obtained results showed, that the ability of generalization of created system, expressed as a measure of sensitivity and specificity increased, due to extracting and selectively choosing only the most representative features for analyzed AF detection problem. Classification results achieved by means of constructed matched wavelet, created for given AF detection features were better than indicators obtained for standard wavelet basic functions used in ECG time-frequency decomposition. I. INTRODUCTION T is a challenging task to obtain explicit knowledge from black-box type machine learning systems like e.g. Neural Networks (NN) [12],[13] or Support Vector Machine (SVM) [2] to explain the relationships hidden in classification decisions made by these structures for human users. The difficulty in the interpretation of knowledge gathered in such a systems is one of the main reasons, that although NN or SVM structures are known to be a relatively robust classifiers, their application in medical diagnosis support systems is limited. A classifier learns from training data and stores learned knowledge into the classifier parameters, such as the weights of a neural network classifier. However, it is difficult to interpret the knowledge in an understandable format by the classifier parameters. Hence, it is desirable to extract IF– THEN rules to represent valuable information in data. Common way used to improve the classifier performance, where the original input signal described in N-element space N 1 X x ℜ ⊆ ∈ is mapped to output classifier vector space Manuscript received April 22, 2009. This work was supported in part by the Polish Ministry of the Science and High Education, under Grant No 1311/B/T02/2007/33. P. S. Kostka is with the Silesian University of Technology, Institute of Electronics, 16 Akademicka St. Gliwice, Poland (corresponding author phone: +48 32 2371160; fax: +48 32 2372225; e-mail: pkostka@ polsl.pl). E. J. Tkacz is with the Silesian University of Technology, Institute of Electronics, 16 Akademicka St. Gliwice, Poland (e-mail: etkacz@polsl.pl). with K-class labels } ,... , { K 2 1 y y y Y y = ∈ (N>>K) is the reduction of too high input feature vector size in intermediate feature extraction and selection stage. The basic goal of this preliminary stage is to reveal only the most discriminate features for given task and discard remain, reducing also the classifier complexity. Proposed feature extraction tools almost always must depend on the specificity of classification task to be sensitive to features, which will be able to distinguish between health and pathology cases. The application field of presented multi-domain feature extraction is the trial of detection of atrial fibrillation (AF), which is a supraventricular tachyarrhythmia characterized by uncoordinated atrial activation with consequent deterioration of atrial mechanical function. On the electrocardiogram (ECG), AF is described by the replacement of consistent P waves by rapid oscillations or fibrillatory waves that vary in size, shape, and timing, associated with an irregular, frequently rapid ventricular response when atrio-ventricular conduction is intact. Because of disturbed haemodynamic, atrial fibrillation and atrial flutter are between of the most usual causes of thrombi-embolic events [1]. II. CLASSIFIER STRUCTURE WITH FEATURE EXTRACTION STAGE BASED ON MATCHED WAVELETS DECOMPOSITION A. General structure of whole classifier system.. General structure of described classifier system for screening examination of patients suffered from atrial fibrillation, presented on fig.1 includes feature extraction stage with proposed mixed-domain structure, following by SVM classiffier structure with preliminary original ECG signal preprocessing with important ventricular activation (QRST) cancelation stage. 1) ECG signal preprocessing with cancellation of ventricular activity On the first preprocessing stage of input ECG signal analysis apart from standard ECG filter implementation the elimination of ventricular activity by QRS complex and T- wave cancellation determining the quality of whole procedure was carried out [6][7]. Literature review of papers connected to AF detection problem shows positive influence of ventricular activation cancelation by removing QRST complex from original signal for further analysis [8]. Traditional methods based on simple QRS cycles averaging and subtraction was not enough in case of significant beat- to-beat changes in real ECG signal, which cannot be treated Rules extraction in SVM and Neural Network Classifiers of Atrial Fibrillation Patients with Matched Wavelets as a Feature Generator. Pawel S Kostka, Ewaryst J Tkacz, Member, IEEE I