Comparing a Supervised and an Unsupervised Classification Method for Burst Detection in Neonatal EEG

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.
  • [Show abstract] [Hide abstract] ABSTRACT: To aid with prognosis and stratification of clinical treatment for preterm infants, a method for automated detection of bursts, interburst-intervals (IBIs) and continuous patterns in the electroencephalogram (EEG) is developed. Results are evaluated for preterm infants with normal neurological follow-up at 2 years. The detection algorithm (MATLAB®) for burst, IBI and continuous pattern is based on selection by amplitude, time span, number of channels and numbers of active electrodes. Annotations of two neurophysiologists were used to determine threshold values. The training set consisted of EEG recordings of four preterm infants with postmenstrual age (PMA, gestational age + postnatal age) of 29-34 weeks. Optimal threshold values were based on overall highest sensitivity. For evaluation, both observers verified detections in an independent dataset of four EEG recordings with comparable PMA. Algorithm performance was assessed by calculation of sensitivity and positive predictive value. The results of algorithm evaluation are as follows: sensitivity values of 90% ± 6%, 80% ± 9% and 97% ± 5% for burst, IBI and continuous patterns, respectively. Corresponding positive predictive values were 88% ± 8%, 96% ± 3% and 85% ± 15%, respectively. In conclusion, the algorithm showed high sensitivity and positive predictive values for bursts, IBIs and continuous patterns in preterm EEG. Computer-assisted analysis of EEG may allow objective and reproducible analysis for clinical treatment.
    Full-text · Article · Sep 2011
  • [Show abstract] [Hide abstract] ABSTRACT: The paper presents an application of Simplified Social Impact Theory based Optimization on feature subset selection for automated neonatal sleep EEG recognition. The target classifier is 3-Nearest Neighbor classifier. We also propose a novel initialization of iterative population based optimization heuristics, which is suitable for feature subset selection, because it reduces the computational complexity of whole feature selection process and can help to prevent overfitting problems. Our methods leads to a significant reduction of the original dimensionality while simultaneously reduce the classification error.
    Article · Jan 2013 · Physiological Measurement
  • [Show abstract] [Hide abstract] ABSTRACT: General anesthesia is a drug-induced, reversible condition comprised of hypnosis, amnesia, analgesia, akinesia, and autonomic stability. During the deepest levels of anesthesia, burst suppression is observed in the EEG, which consists of alternating periods of bursting and isoelectric activity. By accurately tracking anesthesia-induced burst suppression, it may be possible to provide a higher level of care for patients receiving general anesthesia. We develop a probabilistic framework for detecting burst suppression events. The algorithm uses multinomial regression to estimate the probability of burst, suppression, and artifact states at each time given EEG frequency-domain data. We test the efficacy of this method on clinical EEG acquired during operating room surgery with GA under propofol.
    Conference Paper · Nov 2013