Search

Abstract We present a new approach to classifying the sleep stage that incorporates a computationally inexpensive method based on permutations for channel selection and takes advantage of deep learning power, specifically the gated recurrent unit (GRU) model, along with other deep learning methods. By systematically permuting the electroencephalographic (EEG) channels, different combinations of EEG channels are evaluated to identify the most informative subset for the classification of the 5-class sleep stage. For analysis, we used an EEG dataset that was collected at the International Institute for Integrative Sleep Medicine (WPI-IIIS) at the University of Tsukuba in Japan. The results of these explorations provide many new insights such as the (1) drastic decrease in performance when channels are fewer than 3, (2) 3-random channels selected by permutation provide the same or better prediction than the 3 channels recommended by the American Academy of Sleep Medicine (AASM), (3) N1 class suffers the most in prediction accuracy as the channels drop from 128 to 3 random or 3 AASM, and (4) no single channel provides acceptable levels of accuracy in the prediction of 5 classes. The results obtained show the GRU’s ability to retain essential temporal information from EEG data, which allows capturing the underlying patterns associated with each sleep stage effectively. Using permutation-based channel selection, we enhance or at least maintain as high model efficiency as when using high-density EEG, incorporating only the most informative EEG channels.

Abstract Effective management of dementia requires the timely detection of mild cognitive impairment (MCI). This paper introduces a multi-objective optimization approach for selecting EEG channels (and features) for the purpose of detecting MCI. Firstly, each EEG signal from each channel is decomposed into subbands using either variational mode decomposition (VMD) or discrete wavelet transform (DWT). A feature is then extracted from each subband using one of the following measures: standard deviation, interquartile range, band power, Teager energy, Katz's and Higuchi's fractal dimensions, Shannon entropy, sure entropy, or threshold entropy. Different machine learning techniques are used to classify the features of MCI cases from those of healthy controls. The classifier's performance is validated using leave-one-subject-out (LOSO) cross-validation (CV). The non-dominated sorting genetic algorithm (NSGA)-II is designed with the aim of minimizing the number of EEG channels (or features) and maximizing classification accuracy. The performance is evaluated using a publicly available online dataset containing EEGs from 19 channels recorded from 24 participants. The results demonstrate a significant improvement in performance when utilizing the NSGA-II algorithm. By selecting only a few appropriate EEG channels, the LOSO CV-based results show a significant improvement compared to using all 19 channels. Additionally, the outcomes indicate that accuracy can be further improved by selecting suitable features from different channels. For instance, by combining VMD and Teager energy, the SVM accuracy obtained using all channels is 74.24%. Interestingly, when only five channels are selected using NSGA-II, the accuracy increases to 91.56%. The accuracy is further improved to 95.28% when using only 8 features selected from 7 channels. This demonstrates that by choosing informative features or channels while excluding noisy or irrelevant information, the impact of noise is reduced, resulting in improved accuracy. These promising findings indicate that, with a limited number of channels and features, accurate diagnosis of MCI is achievable, which opens the door for its application in clinical practice.

Abstract Early detection of Parkinson’s disease (PD) is very important in clinical diagnosis for preventing disease development. In this study, we present efficient discrete wavelet transform (DWT)-based methods for detecting PD from health control (HC) in two cases, namely, off-and on-medication. First, the EEG signals are preprocessed to remove major artifacts before being decomposed into several EEG sub-bands (approximate and details) using DWT. The features are then extracted from the wavelet packet-derived reconstructed signals using different entropy measures, namely, log energy entropy, Shannon entropy, threshold entropy, sure entropy, and norm entropy. Several machine learning techniques are investigated to classify the resulting PD/HC features. The effects of DWT coefficients and brain regions on classification accuracy are being investigated as well. Two public datasets are used to verify the proposed methods: the SanDiego dataset (31 subjects, 93 min) and the UNM dataset (54 subjects, 54 min). The results are promising and show that four entropy measures: log energy entropy, threshold entropy, sure entropy, and modified-Shannon entropy (TShEn) lead to high classification accuracy, indicating they are good biomarkers for PD detection. With the SanDiego dataset, the classification results of off-medication PD versus HC are 99.89, 99.87, and 99.91 for accuracy, sensitivity, and specificity, respectively, using the combination of DWT + TShEn and KNN classifier. Using the same combination, the results of on-medication PD versus HC are 94.21, 93.33, and 95%. With the UNM dataset, the obtained classification accuracy is around 99.5% in both cases of off-and on-medication PD using DWT + TShEn + SVM and DWT + ThEn + KNN, respectively. The results also demonstrate the importance of all DWT coefficients and that selecting a suitable small number of EEG channels from several brain regions could improve the classification accuracy.

Abstract High-density Electroencephalography (HD-EEG) has proven to be the EEG montage that estimates the neural activity inside the brain with highest accuracy. Multiple studies have reported the effect of electrode number on source localization for specific sources and specific electrode configurations. The electrodes for these configurations are often manually selected to uniformly cover the entire head, going from 32 to 128 electrodes, but electrode configurations are not often selected according to their contribution to estimation accuracy. In this work, an optimization-based study is proposed to determine the minimum number of electrodes that can be used and to identify the optimal combinations of electrodes that can retain the localization accuracy of HD-EEG reconstructions. This optimization approach incorporates scalp landmark positions of widely used EEG montages. In this way, a systematic search for the minimum electrode subset is performed for single- and multiple-source localization problems. The Non-dominated Sorting Genetic Algorithm II (NSGA-II) combined with source reconstruction methods is used to formulate a multi-objective optimization problem that concurrently minimizes (1) the localization error for each source and (2) the number of required EEG electrodes. The method can be used for evaluating the source localization quality of low-density EEG systems (e.g. consumer-grade wearable EEG). We performed an evaluation over synthetic and real EEG datasets with known ground-truth. The experimental results show that optimal subsets with 6 electrodes can attain an equal or better accuracy than HD-EEG (with more than 200 channels) for a single source case. This happened when reconstructing a particular brain activity in more than 88% of the cases in synthetic signals and 63% in real signals, and in more than 88% and 73% of cases when considering optimal combinations with 8 channels. For a multiple-source case of three sources (only with synthetic signals), it was found that optimized combinations of 8, 12 and 16 electrodes attained an equal or better accuracy than HD-EEG with 231 electrodes in at least 58%, 76%, and 82% of cases respectively. Additionally, for such electrode numbers, lower mean errors and standard deviations than with 231 electrodes were obtained.

Abstract In this study we explore how different levels of emotional intensity (Arousal) and pleasantness (Valence) are reflected in electroencephalographic (EEG) signals. We performed the experiments on EEG data of 32 subjects from the DEAP public dataset, where the subjects were stimulated using 60-s videos to elicitate different levels of Arousal/Valence and then self-reported the rating from 1 to 9 using the self-assessment Manikin (SAM). The EEG data was pre-processed and used as input to a convolutional neural network (CNN). First, the 32 EEG channels were used to compute the maximum accuracy level obtainable for each subject as well as for creating a single model using data from all the subjects. The experiment was repeated using one channel at a time, to see if specific channels contain more information to discriminate between low vs high arousal/valence. The results indicate than using one channel the accuracy is lower compared to using all the 32 channels. An optimization process for EEG channel selection is then designed with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) with the objective to obtain optimal channel combinations with high accuracy recognition. The genetic algorithm evaluates all possible combinations using a chromosome representation for all the 32 channels, and the EEG data from each chromosome in the different populations are tested iteratively solving two unconstrained objectives; to maximize classification accuracy and to reduce the number of required EEG channels for the classification process. Best combinations obtained from a Pareto-front suggests that as few as 8–10 channels can fulfill this condition and provide the basis for a lighter design of EEG systems for emotion recognition. In the best case, the results show accuracies of up to 1.00 for low vs high arousal using eight EEG channels, and 1.00 for low vs high valence using only two EEG channels. These results are encouraging for research and healthcare applications that will require automatic emotion recognition with wearable EEG.