Wave segmentation using nonstationary properties of ECG

This paper examines the nonstationary statistical properties of ECG and uses them in signal segmentation. The ECG signal is grouped in three different groups that are defined by QRS complex, P and T waves. These groups are defined by performing nonstationary spectral analysis. A standard autoregressive model is used to extract coefficients of local stationary ECG beat segment. Such a filter is useful for tracking changes in local statistical properties. Finally a Hidden Markov model is used for automatic segmentation. Due to the locality of signal statistical properties, it is important the Hidden Markov model to be initialised carefully. Thus, the initialisation of Hidden Markov chain is accomplished using autoregressive (AR) filter coefficients. The ECG signals are finally segmented into stationary regimes characterized by different autoregressive models in a beat by beat basis. This method has been tested in different ECG signals from the MIT-BIH database. The results showed that there is a good percentage in identification of ECG basic morphologies while there are a number of problems in starting and ending points of segments. The overall results are not affected by low signal to noise ratio. Another advantage of this algorithm is that it is independent of temporal signal changes because of its adaptive nature. In conclusion this algorithm could be used in automated ECG constituent wave detection, something that is useful in telemedicine based environment