Electroencephalographic (EEG) activity recorded during the entire sleep cycle reflects various complex processes associated with brain and exhibits a high degree of irregularity through various stages of sleep. The identification of transition from wakefulness to stage1 sleep is a challenging area of research for the biomedical community. In this paper, spectral entropy (SE) is used as a complexity measure to quantify irregularities in awake and stage1 sleep of 8-channel sleep EEG data from the polysomnographic recordings of ten healthy subjects. The SE measures of awake and stage1 sleep EEG data are estimated for each second and applied to a multilayer perceptron feed forward neural network (MLP-FF). The network is trained using back propagation algorithm for recognizing these two patterns. Initially, the MLP network is trained and tested for randomly chosen subject-wise combined datasets I and II and then for the combined large dataset III. In all cases, 60 % of the entire dataset is used for training while 20 % is used for testing and 20 % for validation. Results indicate that the MLP neural network learns with maximum testing accuracy of 95.9 % for dataset II. In the case of combined large dataset, the network performs with a maximum accuracy of 99.2 % with 100 hidden neurons. Results show that in channels O1, O2, F3 and F4 (A1, A2 as reference), the mean of the spectral entropy value is higher in awake state than in stage1 sleep indicating that the EEG becomes more regular and rhythmic as the subject attains stage1 sleep from wakefulness. However, in C3 and C4 the mean values of SE values are not very much discriminative of both groups. This may prove to be a very effective indicator for scoring the first two stages of sleep EEG and may be used to detect the transition from wakefulness to stage1 sleep.

No abstract

This paper presents a novel ranking method to select spectral entropy (SE) features that discriminate alcoholic and control visual event-related potentials (ERP'S) in gamma sub-band (30-55 Hz) derived from a 64-channel electroencephalogram (EEG) recording. The ranking is based on a t test statistic that rejects the null hypothesis that the group means of SE values in alcoholics and controls are identical. The SE features with high ranks are indicative of maximal separation between their group means. Various sizes of top ranked feature subsets are evaluated by applying principal component analysis (PCA) and k-nearest neighbor (k-NN) classification. Even though ranking does not influence the performance of classifier significantly with the selection of all 61 active channels, the classification efficiency is directly proportional to the number of principal components (pc). The effect of ranking and PCA on classification is predominantly observed with reduced feature subsets of (N = 25, 15) top ranked features. Results indicate that for N = 25, proposed ranking method improves the k-NN classification accuracy from 91 to 93.87% as the number of pcs increases from 5 to 25. With same number of pcs, the k-NN classifier responds with accuracies of 84.42-91.54% with non-ranked features.Similarly for N = 15 and number of pcs varying from 5 to 15, ranking enhances k-NN detection accuracies from 88.9 to 93.08% as compared to 86.75-91.96% without ranking. This shows that the detection accuracy is increased by 6.5 and 2.8%, respectively, for N = 25, whereas it enhances by 2.2 and 1%, respectively, for N = 15 in comparison with non-ranked features. In the proposed t test ranking method for feature selection, the pcs of only top ranked feature candidates take part in classification process and hence provide better generalization.