Patient-specific seizure onset detection based on CSP-enhanced energy and neural synchronization decision fusion

Abstract

This paper presents a patient-specific seizure onset detector based on the fusion of classification decisions from a common spatial pattern (CSP)-enhanced energy based detector and a neural synchronization based detector. Specifically, one level of the detector evaluates the amount of neural synchrony present within the electroencephalography (EEG) channels by calculating the condition number (CN) of the EEG matrix. On a parallel level, the detector first enhances the EEG via CSP and then evaluates the energy contained in four EEG frequency subbands. The information is then fed into two independent and parallel classification units based on support vector machines to determine the electrographic onset of a seizure event. The decisions from the two classifiers are then coupled according to two fusion techniques to determine a global decision. Experimental results demonstrate a sensitivity of 100%, detection latency of 1.75 seconds, and a false alarm rate of 3.14 per hour for the detector based on the AND fusion technique. The OR fusion technique achieves a sensitivity of 100%, and significantly improves delay latency (0.61 seconds), yet it achieves 14.26 false alarms per hour.