Statistical distance of quadtree partitioned attractors: detecting changes in the behavior of ECG signals

Abstract

Describes a method to detect different types of behaviors in an ECG signal using quadtree-partitioned attractors. The method is based on the calculation of the statistical distance between two attractors in a phase space of a dynamical system where a quadtree decomposition was applied. A short portion of an ECG signal with normal behavior (of a patient) is used to reconstruct (using a time-delay technique) the trajectory of an attractor in a 2D phase space. The phase space is then partitioned using a quadtree algorithm. The distributions in the quadtree mesh are interpreted as statistical distributions corresponding to the normal dynamical behavior of the ECG recording in the phase space. An algorithm is then used to compute the correlation distance between this distribution and all other distributions that are built using reconstructed attractors from a sliding temporal window over the signal. It was noticed, in normal cases, that the distance was almost constant and below a threshold. For pathological cases (with abnormal transients), on the abnormal portion of ECG, the distance increased consistently with morphological changes.