Caracterización dinámica de registros ECG para identificación de arritmias

We present a methodology for feature extraction by means of adaptive filter banks in case of automatic identification of arrhythmias using ECG recording. Proposed filter banks, which are supposed to track in more accurate way any change of parameters of time-varying sequence, is developed for biorthogonal wavelet bases using Teager algorithm. Besides, adaptive lifting schemes, which allow filter order change, are used for filter bank implementation. Lifting schemes are introduced because lower computational complexity and less processing time. As features, both maximum value and variance of different wavelet decomposition levels are selected for brain zone classification. Results are provided using MIT database, in case of 2, 4, and 6 wavelet vanishing moments, for classification of 10 different arrhythmias. As a result, classification performance level of 98.5% value, estimated by means of bayesian classifier with Mahalanobis distance, is reached which is better than in 5% in comparison to those obtained figures for filter banks but having fixed parameters.