Article

Comparing a supervised and an unsupervised classification method for burst detection in neonatal EEG.

School of Engineering, University College of Boraş, Sweden.
Conference proceedings: ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference 02/2008; 2008:3836-9. DOI: 10.1109/IEMBS.2008.4650046
Source: PubMed

ABSTRACT Hidden Markov Models (HMM) and Support Vector Machines (SVM) using unsupervised and supervised training, respectively, were compared with respect to their ability to correctly classify burst and suppression in neonatal EEG. Each classifier was fed five feature signals extracted from EEG signals from six full term infants who had suffered from perinatal asphyxia. Visual inspection of the EEG by an experienced electroencephalographer was used as the gold standard when training the SVM, and for evaluating the performance of both methods. The results are presented as receiver operating characteristic (ROC) curves and quantified by the area under the curve (AUC). Our study show that the SVM and the HMM exhibit similar performance, despite their fundamental differences.

0 Bookmarks
 · 
106 Views
  • [Show abstract] [Hide abstract]
    ABSTRACT: To investigate whether the periodic EEG patterns seen in healthy and sick full term neonates (trace alternant and burst suppression, respectively) have different frequency characteristics. Burst episodes were selected from the EEGs of 9 healthy and 9 post-asphyctic full-term neonates and subjected to power spectrum analysis. Powers in two bands were estimated; 0-4 and 4-30 Hz, designated low- and high-frequency activity, respectively (LFA, HFA). The spectral edge frequency (SEF) was also assessed. In bursts, the LFA power was lower in periods of burst suppression as compared to those of trace alternant. The parameter that best discriminated between the groups was the relative amount of low- and high-frequency activity. The SEF parameter had a low sensitivity to the group differences. In healthy neonates, the LFA power was higher over the posterior right as compared to the posterior left region. Spectral power of low frequencies differs significantly between the burst episodes of healthy and sick neonates. These results can be used when monitoring cerebral function in neonates.
    Clinical Neurophysiology 12/2004; 115(11):2461-6. · 3.14 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Fisher's linear discriminant, a feed-forward neural network (NN) and a support vector machine (SVM) are compared with respect to their ability to distinguish bursts from suppression in burst-suppression electroencephalogram (EEG) signals using five features inherent in the EEG as input. The study is based on EEG signals from six full term infants who have suffered from perinatal asphyxia, and the methods have been trained with reference data classified by an experienced electroencephalographer. The results are summarized as area under the curve (AUC) values derived from receiver operating characteristic (ROC) curves for the three methods, and show that the SVM is slightly better than the others, at the cost of a higher computational complexity.
    Conference proceedings: ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Conference 02/2007; 2007:5136-9.
  • [Show abstract] [Hide abstract]
    ABSTRACT: An automatic EEG pattern detection unit was developed and tested for the recognition of burst-suppression periods and for the separation of burst from suppression patterns. The median, standard deviation and the 95% edge frequency were computed from single channels of the EEG within a moving window and completed by the continuous computation of frequency band power via an adapted Hilbert resonance filter. These parameters were given to the inputs of two hierarchically arranged artificial neural networks (NNs). The output signals of NNs indicate the suppression and burst phases. The burst recognition was focused on the precise recognition of the burst onset. In subsequent processing steps the time course of percentages of burst patterns within their corresponding burst-suppression-phases was calculated and the time locations of burst onsets can be used to trigger an averaging for a burst-related analysis. The data for our investigations were derived from the routine EEG derivations of 12 patients with various neurosurgical diseases. A group-related training of the NNs was realized. For the group-related trained NNs EEG data for 6 patients were used for training and the data of 6 other patients for testing the classification performance of the pattern recognition units. Additionally, the reliability of the detection algorithm was tested with data of two patients with convulsive state, resistant to treatment, and burst-suppression like pattern EEG.
    Journal of Clinical Monitoring and Computing 07/1999; 15(6):357-367. · 0.71 Impact Factor