Online Epileptic Seizure Detection: Statistical Signal Processing Based on Multichannel Electroencephalogram Sensors

Abstract

This paper presents the framework for an online algorithm for epileptic seizure detection based on component and discriminant analysis combined with a Bayesian decision approach. In our approach, feature vectors representing ictal, interictal, and normal brain activity are extracted from multichannel electrocorticographic (ECoG) sensor data sets based on autoregressive (AR) model parameter estimation. AR model parameters are estimated according to the Yule-Walker equation for offline feature extraction, whereas the weighted least square estimation (WLSE) is adopted for efficient online AR model parameter estimation. We also consider principal component analysis (PCA) and linear discriminant analysis (LDA) for dimensionality reduction and the most discriminative information extraction to aid online epileptic seizure detection. The extracted information is imported to the online detector implanted on the brain surface. We consider PCA combined with cosine similarity and LDA-based similarity score as a decision rule. Simulation result shows that the PCA method combined with cosine similarity has average 96.6% detection accuracy and 1.2 ms detection latency for the test vectors under study. For fair comparison of detector performance, the line length detection algorithm is tested under the same simulation condition and the result shows that the approach based on the PCA method with cosine similarity measure outperforms the LDA-based approach and the line length detection algorithm. Based on the study, we further propose a hybrid approach combining PCA with cosine similarity and an LDA-based similarity score method. The simulation result shows that the hybrid method has a higher detection rate under the same false detection rate compared to the PCA method with cosine similarity measure.