The existing sleep stages classification methods are mainly based on time or frequency features. This paper classifies the sleep stages based on graph domain features from a single-channel electroencephalogram (EEG) signal. First, each epoch (30 s) EEG signal is mapped into a visibility graph (VG) and a horizontal VG (HVG). Second, a difference VG (DVG) is obtained by subtracting the edges set of the HVG from the edges set of the VG to extract essential degree sequences and to detect the gait-related movement artifact recordings. The mean degrees (MDs) and degree distributions (DDs) P (k) on HVGs and DVGs are analyzed epoch-by-epoch from 14,963 segments of EEG signals. Then, the MDs of each DVG and HVG and seven distinguishable DD values of P (k) from each DVG are extracted. Finally, nine extracted features are forwarded to a support vector machine to classify the sleep stages into two, three, four, five, and six states. The accuracy and kappa coefficients of six-state classification are 87.5% and 0.81, respectively. It was found that the MDs of the VGs on the deep sleep stage are higher than those on the awake and light sleep stages, and the MDs of the HVGs are just the reverse.
1 July 2014 in Computer Methods and Programs in Biomedicine
Epileptogenic localization is a critical factor for successful epilepsy surgery. Determining the epileptogenic hippocampus with single channel intracranial electroencephalography (iEEG) recording is beneficial to decrease the risk of infection compared with that based on multi-channel iEEGs. Delay permutation entropy (DPE) methodology is presented in this study to measure iEEG with different delay lag based on focal epileptogenic zone. A total of 1600 20-s epileptic iEEG are evaluated and are used as features to classify epileptogenic and non-epileptogenic zone. Experimental results show that the DPE index of epileptogenic iEEG is significant lower than that of non-epileptogenic hemisphere when delay lag ranges from 5 to 30 (p=0.01). In addition, the accuracy of identifying epileptogenic region with the DPE index is increased when the delay lag between 5 and 25, compared to the performance of the PE index.
This paper proposes a novel horizontal visibility graph entropy (HVGE) approach to evaluate EEG signals from alcoholic subjects and controlled drinkers and compare with a sample entropy (SaE) method. Firstly, HVGEs and SaEs are extracted from 1,200 recordings of biomedical signals, respectively. A statistical analysis method is employed to choose the optimal channels to identify the abnormalities in alcoholics. Five group channels are selected and forwarded to a K-Nearest Neighbour (K-NN) and a support vector machine (SVM) to conduct classification, respectively. The experimental results show that the HVGEs associated with left hemisphere, \(C\)1, \(C\)3 and FC5 electrodes, of alcoholics are significantly abnormal. The accuracy of classification with 10-fold cross-validation is 87.5 \(\%\) with about three HVGE features. By using just optimal 13-dimension HVGE features, the accuracy is 95.8 \(\%\). In contrast, SaE features associated cannot identify the left hemisphere disorder for alcoholism and the maximum classification ratio based on SaE is just 95.2 \(\%\) even using all channel signals. These results demonstrate that the HVGE method is a promising approach for alcoholism identification by EEG signals.
Identifying stroke subtypes from electromagnetic imaging systems is usually based on frequency domain using radar or tomography algorithms which is computationally expensive. This paper presents a novel graph degree mutual information (GDMI) approach to distinguish Intracranial Haemorrhage (ICH) from Ischemic Stroke (IS). A total of 50 ICH and 50 IS signals simulated using a 16-antenna electromagnetic head imaging system are analysed to evaluate GDMI. The data collected from each model consists of 256 reflected and received signal. Subsequently, noise is injected into the collected signals to generate three groups of signals with different signal-to-noise ratios (40 dB, 25 dB and 10 dB SNR), to emulate measurement noise and to test the algorithm's robustness. Each signal is converted into a graph to avoid the variable signal amplitudes. Then, the relationship between each pair of graph degrees is calculated by mutual information and forwarded to a support vector machine to identify stroke type. The results indicate that signals from ICH subjects exhibit a significantly higher GDMI compared to the IS group ($p<; 0.01$). An accuracy of 88% is achieved in identifying ICH from IS without the need to use time- and resource-expensive brain image reconstruction algorithms even under 25 dB signal-to-noise levels. The execution time for graph feature extraction and classification is less than one minute on a PC. Such a short time is suitable for stroke emergency requirements.