City Research Online

Application of Higher-Order Statistics and Subspace-Based Techniques to the Analysis and Diagnosis of Electrocardiogram Signals

El-Khafif, S. H. (2002). Application of Higher-Order Statistics and Subspace-Based Techniques to the Analysis and Diagnosis of Electrocardiogram Signals. (Unpublished Doctoral thesis, City, University of London)


The first and main contribution of this research work is the higher-order statistics (HOS)-based non-linear analysis and subsequent diagnosis of abnormal electrocardiogram (ECG) signals, particularly myocardial ischaemia. In the time domain; the second-, third-, and the fourth-order cumulants have been used in the analysis. In the frequency domain; up to the tenth-order polyspectra have been exploited. This HOS-based analysis of normal and ischaemic electrocardiogram signals has led to the identification of certain key discriminant features for the two physiological states of the heart. These features are then fed to different backpropagation-based multiple layer perceptrons for classification. The second contribution is a proposed new methodology to discriminate patients with angina pectoris or with old myocardial infarction (MI) during the first 60 seconds of stress test (or in some cases using rest ECG). It is based on the pseudo-spectral Multiple Signal Classification (MUSIC) and has the potential of being highly sensitive diagnostic signal processing tool. The third contribution is the development of a novel higher-order statistics, high-resolution estimator for quadratically coupled frequencies based on subspace spectral estimation.

Extensive studies of cumulants, bispectra and bicoherence-squared of normal and ischaemic ECG signals collected from MIT and ST-T European databases has enabled us to see key discriminant features in both the third- and fourth-order cumulant domains. In the frequency domain, the polyspectral study has been extended to the lOth-order poly spectra. By calculating one-dimensional polyspectrum slices using an algorithm developed by Zhou and Giannakis (1995) a considerable reduction in the CPU time has been achieved. Furthermore, Zhou’s algorithm has been further extended to estimate the polycoherency slices which are used to characterise non-linearities in normal and ischaemic ECG signals. An important finding in this thesis is the decrease of the order of non-linearity representing the electrocardiogram signals of ischaemic patients.

This thesis also includes the results of a pilot study involving eighteen healthy subjects (MIT database) and confirmed that the ECG signal is non-Gaussian, cyclostationary and quasi periodic. Combined spectral and bispectral analysis of the signal revealed that there are unique harmonic characteristics for the P-wave, QRS complex and T-wave and other frequencies due to harmonic interactions.

In this work three linear and one non-linear adaptive filtering/predictions techniques have been applied to noisy ECG signals and their respective performances appraised. It is shown that the Kalman filter gives the best mean-square error MSE error but its comparatively long execution time and problems arising from ill-conditioning of the state-error covariance matrix render it of limited use in ECG applications. It is also shown that the LMS-based quadratic and cubic Volterra filters are the most superior for the ECG signal prediction.

For ECG classifications; three multi-layer perceptrons employing back-propagation and modified back-propagation algorithms, and using two sets from the higher-order most discriminant features as their inputs, have yielded fairly high classification rates.

Publication Type: Thesis (Doctoral)
Subjects: R Medicine > R Medicine (General)
T Technology > T Technology (General)
Departments: School of Science & Technology > Engineering
School of Science & Technology > School of Science & Technology Doctoral Theses
Doctoral Theses
[thumbnail of El-Khafif thesis 2002 PDF-A.pdf]
Text - Accepted Version
Download (15MB) | Preview


Add to AnyAdd to TwitterAdd to FacebookAdd to LinkedinAdd to PinterestAdd to Email


Downloads per month over past year

View more statistics

Actions (login required)

Admin Login Admin Login