It is shown that the discrete cosine transform (DCT) of a given set of electrocardiographic (ECG) cycles can be approximated to a set of narrow band signals. These signals are estimated by performing singular value decomposition (SVD) on a prediction matrix made up of the DCT coefficients. The QRS pulses are then separated out in the DCT domain from a reduced rank approximation of the prediction matrix. The size of this reduced rank matrix is obtained by means of an optimally established orthogonality relationship between two complementary subspaces derived from the signal space of DCT coefficients.