Rhythm classification using reconstructed phase space of signal frequency sub-bands

A preliminary study in the use of frequency sub-bands with reconstructed phase spaces (RPS) to distinguish between normal and abnormal atrial activity in an attempt to separate the atrial activation components from the ventricular activation components is presented. Two-second ECG Holter recordings of sinus rhythm (SR), atrial flutter (AFL), atrial fibrillation (AF), supraventricular tachycardia (SVT) and ventricular tachycardia (VT) were filtered into four sub-bands (0.5-5, 5-10, 10-20, and 20-32 Hz) and embedded into a 3-dimensional RPS. Gaussian mixture models of the sub-banded RPS were learned. The models learned over the 5-10 and 10-20 Hz bands had the best overall classification accuracy. SR was best classified in the 5-10 Hz band with no false positives in the 10-20 Hz band. AFL's highest classification was in the 10-20 and 20-32 Hz bands, AF in the 0.5-5 Hz band, SVT in the 5-10 Hz band, and VT in 10-20 Hz band. When the atrial arrhythmias were folded together into one class, the highest overall classification accuracy increased from 79% in the 10-20 Hz band to 92% in the 5-10 Hz band. These results are promising for the use of sub-banded RPS in the classification of atrial arrhythmias from surface ECGs.