ABSTRACTOffline EEG signal processing for seizure detection is an important clinical application. In this paper, we introduce the offline seizure detection algorithms using single-channel and one-second-length of short-time EEG signals based on time domain and novel frequency-sequency features and also novel feature reduction scenarios. After preprocessing stages, the EEG signals are decomposed to five brain rhythms using uniform discrete Fourier transform (DFT)-based filter bank. Then, transform-based features by the DFT and discrete Walsh–Hadamard transform are defined. With computing spectral correlation model (SCM) using Fourier and Walsh patterns of the rhythms, the Walsh power is computed as a modulus maximum feature that is called spectral correlation power. By segmenting the SCM, the histogram-based statistical features namely; kurtosis, skewness and negentropy as non-Gaussianity measures of probability density function for each epoch of the SCM are calculated. These extracted features are fused in multi-level structures with four different arrangements of kernel-principle component analysis, linear discriminant analysis and kernel matrix. These feature reduction methods are called multi-level dimensionality reduction (MLDR) algorithms. Ultimately, the reduced features with a hybrid model of particle swarm optimization and probabilistic neural network are applied to select the optimal features. To classify the EEG signals into the seizure and non-seizure classes, an optimized non-linear support vector machine classifier with the Gaussian radial basis function kernel is employed. The proposed methods are evaluated on selected 104-hour of EEG records from 23 pediatric patients and the sensitivity rates of 86.38%, 86.05%, 84.71% and 84.65%, precision rates of 95.33%, 96.97%, 95.33% and 95.57% and false detection rates of 0.0853, 0.0805, 0.0926 and 0.0921 per hour are obtained for different MLDRs algorithms, respectively. With regard to utilize the short-time EEG single-channel signal processing scenarios for different patients with different seizures foci, our proposed algorithms obtain the acceptable performances.