Time-frequency evaluation of segmentation methods for neonatal EEG signals

In order to analyse non-stationary signals, like neonatal EEG, it is sometimes easier to segment signals into pseudo-stationary segments. An evaluation was performed on three previously proposed EEG segmentation methods in order to determine which method is most suited to neonatal EEG analysis. The three methods evaluated are spectral error measurement (SEM), generalised likelihood ratio (GLR) and non-linear energy operator (NLEO). A windowed version of NLEO was also tested in an attempt to minimise the effect of any temporary transients on the segmentation algorithm. The results from the segmentation algorithm were compared with the time-frequency distribution of the original signal to determine the appropriateness of the segments. It was found that GLR was the most appropriate segmentation method, and that the windowed version of the NLEO method performed better than the non-windowed version, both of which are less computationally expensive than the other methods