Search

This article proposes a novel seizure detection algorithm based on multidimensional covariance matrix and hybrid mutation of binary Harris hawks optimizer. First, raw electroencephalogram (EEG) signals are preprocessed using empirical mode decomposition (EMD) and discrete wavelet transform (DWT) to obtain several subband signals. Second, single-channel single-subband (SCSS) matrices, single-channel multisubband (SCMS) matrices, and multichannel single-subband (MCSS) matrices are constructed and transformed into covariance matrices. After computing the eigenvalues of each covariance matrix, statistical parameters, which include mean, variance, and skewness, are extracted to form a feature set. Then, a binary Harris hawks optimization algorithm with Cauchy–Gaussian mutation (CGBHHO) is proposed to enhance global search capability and local search capability of binary Harris hawks optimization (BHHO) algorithm to accomplish iterative feature selection. Finally, random forest (RF) classifier is employed for automatic seizure detection. The proposed method achieves favorable experimental results with a tenfold cross-validation evaluation on the CHB-MIT database. The results show that the accuracy (ACC), specificity (SPE), sensitivity (SEN), and F1 score (F1) of the method can reach 99.32%, 99.22%, 99.45%, and 99.36%, respectively.

Research on the electroencephalogram-electromyo- gram (EEG-EMG) functional network is of great significance for exploring the correlation and diagnosing neurological diseases between EEG and EMG. To detect interictal and ictal periods of West syndrome with reliability and accuracy, this article proposes seizure detection method based on a node-optimized EEG-EMG fusion network. In contrast to conventional brain networks, the EEG-EMG fusion network constructs connections between EEG and EMG functional networks, effectively using complementary data from multiple modalities, and combines enhanced artificial rabbit optimization (ARO) algorithm to choose the ideal network node combination. First, the EEG and EMG data are preprocessed. Then, three functional networks are constructed using the preprocessed EEG signals, EMG signals, and EEG-EMG fusion signals. The EEG signals are divided into five frequency bands, and mutual information (MI) is used as an indicator for the connection between network nodes. Finally, the nodes for fusion network are chosen using enhanced ARO algorithm. The EEG and EMG data of 17 West syndrome patients are records by the Children’s Hospital of Zhejiang University School of Medicine and which are evaluated by fivefold cross-validation method. The results of experiment indicate that average accuracy, precision, sensitivity, specificity, and $F1$ scores can reach 97.72%, 97.73%, 97.86%, 97.38%, and 97.54%, respectively.

Long-term electroencephalography (EEG) rec- ordings contain multiple types of artifacts (eye blink artifacts, chewing artifacts, electromyography (EMG) artifacts, etc.), that severely interfere with the EEG reading by clinicians and researchers. Existing research often focuses on specific artifact detection, which cannot handle the complex multitype artifact detection. Therefore, a multiple artifact detection method based on adaptive scalp region selection and classifier fusion is proposed in this article. The main contributions of the article are as follows: 1) according to the characteristics of different scalp area generated by artifacts, adaptive brain region screening is realized. 2) ReliefF and mRMR are fused to realize two-stage feature evaluation and reduce the feature redundancy. 3) Multimodel cascade fusion can bring stronger robustness and better generalization. The data used in the study are collected from real patients suffering from benign epilepsy with centro-temporal spikes (BECT) at the Children’s Hospital of Zhejiang University School of Medicine (CHZU). Compared with the state-of-the-art methods, the proposed method shows the best performance in the detection of multitype artifacts, and achieves the F1-score of 87%, precision of 87%, recall of 88%, and accuracy of 93% on the CHZU dataset.