Automatic detection of epileptic seizure events using the time-frequency features and machine learning. (August 2021)

• Record Type: Journal Article
• Authors: Zeng, Jiale; Tan, Xiao-dan; Zhan, Chang'an A.
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Epileptic seizure detection -- Time-frequency feature extraction -- Empirical wavelet transform -- Discrete wavelet transform -- Machine learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102916
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Prior knowledge based algorithm development for seizure detection. Superior performance consistent across all the tested cases. The number of signal decomposition levels determines the algorithm performance. The ways of using signal features greatly influence the algorithm performance. Abstract: Computer-aided seizure detection from the long-term EEG has shown great potential in improving the epilepsy diagnosis accuracy and efficiency. This study was aimed to utilize prior knowledge about the epileptic EEG signals in choosing the algorithms and parameters in order to improve the performance and robustness for seizure detection. Our choices included the kurtosis-based channel selection, five-level empirical wavelet transform (EWT) for EEG signal decomposition adaptive to the power spectra of individual channels, the direct use of the instantaneous time–frequency features, and post-processing of classification outcomes. Using the publicly available CHB-MIT epileptic EEG database, we tested our algorithm against the high performance ones published in the recent literature and the high potential alternatives. Our algorithm achieved the mean sensitivity, specificity, and accuracy of 99.77%, 99.88%, and 99.88%, respectively, surpassing the best results in the literature in terms of mean performance. More importantly, the performance was highly consistent across all the tested patient cases. Statistical analysis showed that the levels of EWT and the direct use of time–frequency features had the greatest impacts on the final seizure detection performance. In addition, our results showed that k-nearest neighbors (KNN) and support vector machine (SVM) outperformed random forest (RF) classifier, contrary to the finding the RF being the best for seizure detection task in the literature. Our findings demonstrate the value of prior-knowledge-based feature extraction and suggest the equal importance of feature extraction and classifier in the seizure detection algorithm.