Automatic detection for epileptic seizure using graph-regularized nonnegative matrix factorization and Bayesian linear discriminate analysis

Abstract Epilepsy is a neurological disorder characterized by excessive neuronal discharge which results in many problems in terms of behavior, state of mind, consciousness, and can threaten the lives of patients. An automatic epileptic seizure detection method with graph-regularized non-negative matrix factorization (GNMF) and Bayesian linear discriminate analysis (BLDA) is presented in this paper. First, discrete wavelet decomposition is applied to analyze raw electroencephalogram (EEG) signals, and the normalization based on differential operator is used to guarantee the nonnegative constraint and reinforce the distinction between seizure and non-seizure signals. Then, GNMF is employed to dimensionality reduction and feature extraction for EEG data, which could capture a parts-based representation of samples and obtain more discriminative features. The EEG features are calculated and entered into the BLDA classifier for categorized results. The public Freiburg EEG database is used to evaluate the performance of the proposed seizure detection method. The results showed event-based sensitivity of 95.24%, epoch-based sensitivity of 93.20%, and a false-alarm rate of 0.5/h. These results demonstrate the potential clinical value of this method for automatic seizure detection.