Detection of epileptic seizures from compressively sensed EEG signals for wireless body area networks. (15th June 2021)

• Record Type: Journal Article
• Authors: Aghababaei, Mohammad H.; Azemi, Ghasem; O'Toole, John M.
• Is Part Of: Expert systems with applications. Volume 172(2021)
• Publication Date: 2021-06-15
• Subjects: EEG -- Epileptic seizure -- Compressive sensing -- Wireless body area networks -- Partial energy difference feature; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2021.114630
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Novel features for seizure detection from compressive EEG measurements are proposed. The performance of the proposed features are evaluated using two well-known databases. The features accurately detect seizures at the compression ratios as small as 0.05. The features can be used for real-time tele-monitoring of epilepsy patients. Abstract: Wireless body area networks (WBANs) are gaining popularity for tele-monitoring of biomedical signals such as the electroencephalogram (EEG), with diagnosing and monitoring of epileptic seizures being one of the most important applications. Most seizure-detection algorithms cannot be applied directly to the compressed data in WBANs as they require full reconstruction of original EEG signals. In this study, we propose a novel feature for real-time automatic single-channel seizure detection which does not require complete reconstruction of original EEGs. The feature is based on iteratively applying the orthogonal matching pursuit (OMP) algorithm on the compressed EEG data and computing the rate by which the energies of partially reconstructed signals are increased. The feature, i.e. partial energy difference (PED), is then used for classifying seizure and non-seizure states. We also extend this method to the case for multichannel EEG. In multichannel case, the simultaneous OMP (SOMP) with a low number of iterations (1 and 15 iteration) is applied to the compressed data and the difference between the Frobenius norms of partially reconstructed signals is used as a multivariate feature for the classification of seizure and non-seizure states. The proposed features are used to detect seizure intervals in the EEG database provided by the University of Bonn and the CHB-MIT database. The results show that the proposed features can classify seizure epochs from non-seizures even for compression ratios as small as 0.05. The results also show that the proposed single-channel method achieves improvement of up to 4% for the area under the curve (AUC) with significantly less execution time compared to the benchmark matrix determinant (MD) classification approach. The results of applying the proposed multivariate feature to seizure and non-seizure segments of length 5 s from the CHB-MIT database show that it achieves mean AUC values of 0.941, 0.941, and 0.939 with mean execution times of 9.5, 10.7, and 13.1 s for compression ratios of 0.05, 0.1, and 0.2, respectively. Applying a threshold to the PED in a leave-one-out cross-validation (LOO-CV) scenario generates a sensitivity of 0.873, specificity of 0.710, and accuracy of 0.791 at the compression ratio of 0.05. The proposed single- and multichannel features have the potential for deployment in WBANs for the real-time tele-monitoring of epilepsy patients in healthcare applications.