A comparison of wavelet transforms through an HMM based ECG segmentation and classification system

This work is intended to fill a gap in the ECG analysis literature using wavelets. The wavelet transform is employed as the parameter extraction stage necessary to build the observation sequence used by the hidden Markov models (HMM). We selected a group of five wavelet functions commonly used in ECG analysis and tested them in an original HMM based ECG segmentation system in order to evaluate the strengths of each wavelet function in a real world application. All experiments were realized on the QT database, which is composed of a representative number of ambulatory recordings of several individuals and is supplied with manual labels made by a physician. Our main contribution relies on the consistent set of experiments performed. Moreover, the results obtained in terms of beat detection and segmentation, and premature ventricular beat (PVC) detection compare favorably to others works reported in the literature, independently of the type of wavelets. Finally, we originally combined the strengths of more than one wavelet function, achieving our best performances.