Abstract
Low-frequency signals comprise different types such as Electroencephalogram (EEG), gyroscope and seismic signals. Processing of EEG signals is performed for tasks such as seizure prediction and detection. On the other hand, processing of seismic and gyroscope signals is performed for tasks such as activity classification. This paper presents two efficient models for anticipation of anomalies from low-frequency signals. A detailed study of EEG seizure prediction is introduced in this paper based on wavelet-domain processing and compression techniques as an example. The first model uses different families of wavelet transform, while the second one concentrates on lossy compression techniques and their effect on further processing for seizure prediction in a realistic signal acquisition and compression scenario. The prediction approach adopts statistical processing with training and testing phases. The training phase comprises estimation of six signal attributes: amplitude, derivative, local mean, local variance, local median and entropy. On the other hand, the testing phase is performed with a thresholding strategy on the selected probability bins. A majority voting strategy with a moving average smoothing filter is used for decision making. The suggested models are executed on long-term EEG recordings from the available Physio-Net EEG dataset. Simulation results in the first model show that the Daubechies wavelets demonstrate the best prediction results as the filter lengths in these wavelets are longer than those in the Haar wavelet. The obtained results in the second model prove the feasibility of lossy compression, especially Discrete Cosine Transform (DCT) compression for seizure prediction.