Informatics Publishing Limited

Cuneyt Yücelbas ¹, Sule Yucelbas ², Seral Ozsen ¹, Gulay Tezel ², Serkan Kuccukturk ³, Sebnem Yosunkaya ³

Affiliations
1 Department of Electrical and Electronics Engineering, Selcuk University, Konya, TR
2 Department of Computer Engineering, Selcuk University, Konya, TR
3 Sleep Laboratory, Faculty of Medicine, Necmettin Erbakan University, Konya, TR

Abstract

Background/Objectives: In this study, Fast Fourier Transform (FFT), Welch, Autoregressive (AR) and MUSIC methods were implemented to detect sleep spindles (SSs) in Electroencephalogram (EEG) signals by extracting features in frequency space. Methods/Statistical Analysis: A database from these signals of five subjects which were recorded at sleep laboratory of Necmettin Erbakan University in Turkey was ready for use. The database consisted of 600 EEG epochs in total. The number of epochs was 300 for both with and without SSs in this database. Comparison of the performances of these methods on SS determination process was performed by using Artificial Neural Networks (ANN) classifier. Findings: According to the test classification results, notable difference was obtained between the applied PSD methods. By using the extracted all features, maximum test classification accuracies were achieved as 84.83%, 80.67%, 80.83% and 80.33% with use of FFT, Welch, AR and MUSIC, respectively. To determine the SSs, Principal Component Analysis (PCA) also was utilized in this study. When PCA was applied, the results were 89.50%, 82.00%, 93.00% and 94.83% by use of the same PSD methods, respectively. Application/Improvements: As a result, the performance of PCA and MUSIC is better than the others. Hence, these methods can be used safely for automatic detection of SSs.

Keywords

AR, EEG, FFT, MUSIC, Sleep Spindle, Welch.

Full Text

   

Sule Yucelbas ¹, Seral Ozsen ², Cuneyt Yucelbas ², Gulay Tezel ¹, Serkan Kuccukturk ³, Sebnem Yosunkaya ³

Affiliations
1 Department of Computer Engineering, Selcuk University, TR
2 Department of Electrical and Electronics Engineering, Selcuk University, Konya, TR
3 Sleep Laboratory, Faculty of Medicine, Necmettin Erbakan University, Konya, TR

Abstract

Background/Objectives: Studies on the field of automatic sleep stage classification have been taking more attention of researchers day by day. Noise in the recordings, nonlinear dynamic feature of EEG signals and some other reasons affect the performance of proposed systems in negative manner. Methods/Statistical Analysis: Sleep can be divided five main stages as Wake, Non-REM1, Non-REM2, Non-REM3 and REM. Almost every proposed method can successfully classify some evident stages like Non-REM2 and REM. But when it comes to the transitions between stages, the systems are not very good in their performances. Thus a different classification strategy was proposed in this study. Five different classifiers were designed especially for transitions between stages using time domain features of EEG, EOG and EMG signals and evaluated these features for each classifier. Sequential backward feature selection process was applied in each classifier to find out which features are dominant in each classification procedure. Artificial Neural Networks was used in designed classifiers. Findings: The highest classification accuracy was obtained as 91.03% for Classifier-3 which predicts stages coming after Non-REM II. The lowest accuracy was recorded as 75.42% for Classifier-2 in which stages are determined after the Non-REM I epochs. Comparatively good results were reached especially if it is taken into account that only used time-domain features of signals. Application/Improvements: The obtain results show that the designed classifiers can be used in automatic sleep staging system, confidently.

Keywords

ANN, Automatic Sleep Stage Classification, EEG, EMG, EOG, Feature Selection.

Full Text

   

Ahmet Coskun ¹, Seral Ozsen ¹, Sule Yucelbas ², Cuneyt Yucelbas ³, Gulay Tezel ², Serkan Kuccukturk ⁴, Sebnem Yosunkaya

Affiliations
1 Department of Electrical and Electronics Engineering, Selcuk University, Konya, TR
2 Department of Computer Engineering, Selcuk University, Konya, TR
3 Department of Electrical and Electronics Engineering, Selcuk University, Konya
4 Sleep Laboratory, Faculty of Medicine, Necmettin Erbakan University, Konya, TR

Abstract

Background/Objectives: Sleep staging is very important phase for diagnosing respiration and sleep diseases. Nowadays, Electroencephalogram (EEG), Electromyogram (EMG), Electrooculogram (EOG) signals are particularly used together in studies on sleep staging. Methods/Statistical Analysis: Associating only EOG signals to sleep staging was distinctly purposed. So, this paper deals with extraction features and classifying for determining REM-NREM states from the EOG signals. In this study, left eye (LEOG) and the right eye (REOG) signals were used. After EOG signals were obtained, 21 different features were extracted from LEOG and REOG in time and frequency domain according to rules of American Academy of Sleep Medicine (AASM). Findings: Artificial Neural Networks (ANN) was adopted on features as method of classification with 3-fold cross validation technique and reached conclusion with the maximum test classification accuracy as 88.05%. To obtain higher classification accuracies, Sequential Backward Selection (SBS) method was used. According to results of SBS, number of the best features combination was determined as 13 and the maximum classification accuracy was obtained as 89.62%. The optimum value of hidden layer node number of ANN was determined as 15 for the best features. Application/Improvements: When looking from the viewpoint of percentage of classification accuracy of this study, a result can be seen that is non-negligible value for literature.

Keywords

Artificial Neural Networks, EOG, Feature Selection, SBS, Sleep Stage.

Full Text