Independent Component Analysis In Biomedical Signal Processing Applications

The extraction of fetal electrocardiogram (FECG) is of vital importance fromclinical point of view, because it can provide information about the health and thepossible diseases of fetus. Before delivery, noninvasive techniques to acquire FECGare preferred, i.e., extraction FECG from the maternal ECG measured from mother'sbody. However, the desired fetal heartbeat signals appearing at the electrode outputalways are buried in an additive mixture of noises and interferences. The strongestone is the maternal electrocardiogram (MECG) contributions, with extremely higheramplitude than the fetal heartbeat signals. Besides, mother's respiration, thermal noisedue to the electronic equipment, and other interferences, also corrupt the fetalheartbeat signals.Many methods have been proposed to address this problem, such as coherentaveraging, matched filtering, auto-and cross-correlation based methods, adaptivefiltering, singular value decomposition (SVD), multireference adaptive noisecancellation. But all these methods have some drawbacks; for example, the signalacquisition is highly sensitive to electrode placement, human interaction, position offetus, etc. The dissatisfying results call for a complete reformulation of the problem,whereby attention would to paid to the fundamental aspects behind the biologicalproblem.Independent component analysis (ICA) is a recently developed signal processingtechnique. Its powerful potentials have been showed in many fields, such aselectroencephalogram analysis, noise reducing, wireless communication, imageprocessing, audio signal processing and so on. Since the problem of FECG extractioncan be modeled as the ICA model, it is natural to apply ICA technique to extractingFECG.In this paper, the basic knowledge of ICA has been first introduced, including thestatistics theory and information theory that are necessary to understanding andmastering ICA technique. Then the basic principle of ICA has been introduced.Next, two famous ICA algorithms, i.e., extended Infomax algorithm and fastfixed-point algorithm, have been introduced. The former is an online ICA algorithm,and the latter is offline algorithm. Both algorithms have been applied to the problemof FECG extraction. The results have been compared to that of LMS algorithm, one ofclassic adaptive filter algorithms. Conclusion can be safely drawn that ICA algorithmscan separate much clearer FECG than adaptive filter algorithms, and are non-sensitiveto the electrodes position.Another focus of this paper is the introduction of several algorithms proposed byus. One is mExtICA algorithm. The new algorithm is obtained by introducing askewness-adjusting parameter to the Pearson mixture density model, which enablesthe model cover a wider range of sub-Gaussian distribution including asymmetricaland multi-modal ones, resulting in more precisely approximating source's density.When dealing with skewed mixed sources, the new algorithm switches among severalfixed non-linearities according to the sources' skewness and kurtosis; while dealing...