Extracting ECG signal characteristics based on non-linear transformations and wavelets

Different extraction methods were compared regarding the characteristics of normal ECG signals and those emitted in the presence of events related to ischemic cardiopathy based on diagnosis measurements, wavelet transformation and nonlinear analysis of main components. Methods were developed for automatic recognition between normal and ischemic ECG signals. Two effective feature selection techniques were proposed; one used multivariate statistical methods and the second univariate ones. Linear discriminatory evaluation and vector support machines were used for evaluating the proposed feature extraction techniques, comparing error when classifying different states of cardiac functionality. Nonlinear PCA offered slightly better performance compared to wavelet representation but was much better compared to diagnosis measurement. There was up to 0.22% error compared to 6.78% in the case of wavelets and 24.22% in the case of diagnostic measurements. Support vector machines increased the performance for all analysed feature extraction methods; more discriminating characteristics were obtained when using wavelets applied to heartbeat having up to 0.1% classification precision compared to 0.12% in the case of nonlinear analysis of main components and 5.11% in the case of diagnostic measurements