Graph Signal Processing Based Epileptic Seizure Event and its Onset Detection from Multi-Channel EEG Signals

This is the authors' abstract. We don't add key points for this paper.

Epileptic seizure detection plays a vital role in the diagnosis of a chronic neurological brain disorder called epilepsy. In this paper, a novel graph signal processing based seizure detection method is proposed by generating spatio-temporal graphs from the image representations of multivariate EEG signals. The proposed scheme consists of two sequential steps: identifying seizure events from the graphs of imaged EEG signals and deriving epileptic seizure onsets from them. Initially, long multi channel EEG signals are represented in the ordered sequence of fixed size images. Each image represents a specific time epoch, with the amplitude at each sample point from its corresponding channel serving as the 2D image’s pixel property with vertically aligned EEG channels and temporal dimensions representing the rows and columns, respectively. The graph corresponding to each image is then obtained by defining non-overlapping sub-images as vertices of the graph. A distance-based weighting function is used to construct a weighted graph from the relationships among the sub-images. The first non-zero eigenvalue of the graph Laplacian, referred to as the Fiedler value or spectral gap λF, is used as a graph-spectral feature for seizure-event detection. Moreover, a post processing method on detected seizure events is developed to mark the onset of the seizure event. The performance of the proposed method is evaluated using benchmark CHB-MIT EEG database using metrics of accuracy, sensitivity, and false-positive rate with several benchmark classifiers. The results show that the proposed method can detect subject-specific epileptic seizure along with minimal onset-offset latencies, which are comparable with the state of the art literature. Further, the efficacy of the proposed approach is also validated for cross subject analysis by using normalised feature across all subjects. Moreover, the effect of adding white Gaussian noise with varied signal-to-noise ratios was tested to evaluate the effectiveness of the suggested technique.