Separation and Identification of Biomedical Signals based on Frequency Domain Independent Component Analysis

Biomedical signals are characterized by the presence of several components belonging to different physiological and non physiological sources. A method is proposed to separate and identify sources of interest from multichannel data acquisitions of vital signs, taking advantage of general hypotheses of statistical independence among the underlying sources. The method is based on a convolutive mixtures model and the strong hypothesis of an instantaneous mixing of the sources component is relaxed. The independent components are extracted by a frequency domain analysis, where the convolutive model can be solved by instantaneous mixing model approach. Two different methods are proposed to identify components separated by the convolutive mixtures model: the first one is based on the areas subtended by the autocorrelation function of the components magnitude spectrograms while the other and more general one requires a classification procedure of the features extracted from the spectrograms by means of singular value decomposition (SVD). Simulations are carried out to test the validity of the proposed algorithm in convolutive mixtures of electrocardiographic (ECG) and electromyographic (EMG) signals. The method is also tested on real ECG and EMG acquisitions that present reciprocal contamination.