Detection of neonatal EEG burst-suppression using a time-frequency approach

In newborn EEG, the presence of burst suppression carries with it a high probability of poor neurodevelopmental outcome. This paper presents a novel method to detect neonatal bust suppression from multichannel EEG using a time-frequency (T-F) based approach. In this approach, features are extracted from T-F representations of EEG signals obtained using quadratic time-frequency distributions (QTFDs). Such features take into account the non-stationarity of EEG signals and are shown to be able to discriminate between burst and suppression patterns. The features are based on the energy concentration of the signals in the T-F domain, instantaneous frequency of the signals, and Renyi entropy and singular-value decomposition (SVD) of the TFDs of EEG. For each feature, the receiver operating characteristic (ROC) is found and the area under the ROC curve (AUC) is calculated as the performance criterion. Experimental results using EEG signals with burst suppression acquired from 3 term neonates show that the features extracted from the singular values of TFDs and energy concentration outperform others. Amongst different QTFDs, features extracted from the optimized extended modified B distribution exhibit the best performance. Also, a classifier which uses these features achieves a total accuracy of 99.6%.