Automated Rapid Seizure Detection in the Human ECoG

Automated seizure detection with high specificity and sensitivity is a highly desirable but elusive goal. The failure to develop a reliable system despite decades of effort is due in part to the non-stationary and noise in the EEG/ECoG signals, as well as to the rudimentary mathematical treatment it has received. Another important limitation of present methods is their inability to perform seizure detection in real time. The most popular methods currently in use for on-line analysis are generally unable to make a detection within ten seconds of the onset of a seizure. We have developed a method of automated seizure detection based on a combination of linear and nonlinear filtering techniques, including the discrete wavelet transform. To minimize noise, this methods was first developed for intracranial signals, then later adapted to scalp recordings. Preliminary results indicate that this new method may be the fastest and most reliable to date. The generic algorithm (without any patient-specific tuning) has been tested on 5 patients and a total of 20 seizure segments and 7 interictal segments recorded form intracranial electrodes. No seizures were missed, no false detections occurred, and the seizure detections were, on average, 16 seconds BEFORE the clinical onset of the seizure. We have also compared the method to expert visual analysis. performed independently and retrospectively by an epileptologist/electroencephalographer through review of polygraph tracings of the signals, and have found our method to be both fast and highly accurate. we are generally able to detect the electrographic seizure within a second of the time marked by the EEGer, and detected the onset sooner in 8 of the 20 cases. The method is also highly adaptable it automatically accounts for signal changes over time, and has a number of parameters which may be tuned to further improve accuracy for a given individual patient or for a particular signal or group of signals being monitored. The algorithm has been implemented and now allows real t ime monitoring and detection on a 486/DX 33 Mhz PC. Based on our preliminary results, we believe this method to be a sign%&& improvement both in accuracy and speed over other currently available methods for automate seizure detection. [Full paper not received in time for publication] Proceedings of the Eighth Annual IEEE Symposium on Computer-Based Medical Systems (CBMS '95) 1063-7125/95 $10.00 © 1995 IEEE