Independent Component Analysis (ICA) has received a lot of attention in statistical as well as in biomedical signal processing. It is widely used in blind source separation (BSS) problems, as it is a convenient method to separate signals from different sources, without any prior information about them or the mixing process. In the first part of the dissertation we report a theoretical background of ICA, analyzing what are the preprocessing steps that are needed and how ICA works, and then giving more details on the two algorithms that are compared, fastICA and ProDenICA. In the second part we present experimental results in a simulation environment to see what ICA achieves and what are the merits and drawbacks of the two ICA algorithms while in the third part we consider a real surface Electromyography (sEMG) dataset. sEMG is one type of bioelectrical signals produced by the human body and contain significant information about muscle activity. ICA is applied to sEMG signals in order to recover the original signals originating from each muscle. Besides, a post-ICA method that overcomes the independent component ordering ambiguity is proposed, allowing them to be related to the suitable corresponding muscles. ICA and the post-ICA steps that are described, manage to reach more than 79% accuracy on three hand gesture classification problems. The experimental results in both simulation and sEMG dataset indicate that ICA is an appropriate method for signal recovering and identification of hand gestures using sEMG signals.
