Cyclic alternating pattern estimation based on a probabilistic model over an EEG signal

Abstract

A probabilistic model for sleep analysis is proposed in this work, modeling the temporal relation between the sleep structure and the presence of the electroencephalogram (EEG) Cyclic Alternating Pattern (CAP) with a Hidden Markov Model (HMM). Sleep scoring is frequently performed by assigning a state to each thirty second epoch. However, this approach does not provide enough time resolution to efficiently detect the CAP since, by definition, the CAP cycles are assessed by applying the scoring rules to one second epochs of the EEG signal. Thus, a clustering analysis was employed, with a one second epoch, over the EEG signal to create clusters that were then encoded using symbolic dynamics to produce words. Two algorithms for clustering were analyzed, specifically the self-organizing map and the Gaussian Mixture Model (GMM). The words were then fed to the HMM to determine the presence of the CAP. Both single-channel and multi-channel (based on sensor fusion) approaches were tested. The best results were attained using the GMM with three Gaussians, achieving an average accuracy, sensitivity, specificity and area under the receiver operating characteristic curve of, respectively, 72%, 66%, 75% and 0.71 for single-channel and 76%, 61%, 85% and 0.73 for multi-channel. This results are in the specialist agreement range with visual analysis. Therefore, the proposed model is capable of providing a new view over the CAP cycles by simplifying the complex EEG signal to a simple sequence of symbols. Such analysis can be significantly challenging to perform in more abstract models.