On clustering based nonlinear projective filtering of biomedical signals

We propose to modify the method of nonlinear state-space projections (NSSP) by application of the technique of k-means clustering. NSSP performs reconstruction of the state-space representation of the processed signals using the Taken's method of delays. Then it projects each state-space point on the appropriately constructed signal subspace and recovers the one-dimensional signal by averaging the results of all projections. The k-means clustering is applied to form so-called neighborhoods on the basis of which the signal subspaces are created. Within these neighborhoods, local density around each state-space point is estimated, to make construction of the signal subspaces more immune to high energy electromyographic noise. The developed method is applied to process different types of ECG signals. For reference, the original NSSP method and its previously developed modifications are used. In different types of noise environment, the proposed method appears more effective than the original one and in most cases than the other reference methods. Moreover, visual results of fetal phonocardiogram and electronystagmogram processing show the wide range of its possible applications.