Abstracts

RATIONALE: We describe the implementation and evaluation of a method (Schindler et al., Clin Neurophysiol 2001;112:1006-1017) for detection of spiking patterns in non-invasive scalp EEG registrations that are associated with epileptic seizures.
METHODS: Our detection method consists of a number of consecutive signal processing stages.
At the initial stage, each EEG signal is low pass filtered. The mean of all EEG channels is then computed and subtracted from each of the channels.
The next signal processing stage marks parts of the EEG signals where the slopes changes from low to high values with a sequence of unit pulses. For that purpose, the absolute values of derivatives (slopes) are calculated for each DC-corrected EEG channel. Next, each of the derivatives is normalized using a continuously updated standard deviation of the slope computed from a prior interval of the same channel. In case the normalized slope exceeds a threshold, a unit pulse is created.
In the third stage, the trains of unit pulses act as input for a leaky integrate and fire unit (LIFU). The LIFU is a classical simple cell model transforming input pulse trains ([ssquote]action potentials[ssquote]) from the previous stage in a sequence of slower excitatory postsynaptic potentials (EPSP) that sum up if they are spaced closely enough. Each time the summated EPSPs exceed a threshold, the LIFU emits a spike and resets the signals across all the internal circuits to zero.
In the final stage, the output spikes of the LIFU are used to calculate a spiking rate (SR). If SR exceeds the given SR-threshold, a seizure onset is detected.
We evaluated the method on 47 recordings obtained from 15 randomly selected patients suffering from different types of drug-resistant epilepsy, who were evaluated for possible epilepsy surgery at the Inselspital in Bern. Recordings consisted of more than 11 hours of EEG data and contained 47 seizures. All the patients signed an agreement that the EEG data might be used for research purposes. A standard 10-20 system of scalp electrodes and two FoV electrodes with four contacts each were used for EEG recording. The evaluation of the method was done for scalp electrodes only.
Detection parameters were optimized for a training set of data and then kept fixed. The actual evaluation was done in an independent test set of recordings.
RESULTS: 41 seizures (87%) were detected, 6 seizures (13%) were missed. There were 13 false detections during the evaluation. In terms of sensitivity and positive predictive accuracy it is 87% and 75% respectively.
The method is implemented and integrated in a clinical EEG software package (EEMAGINE Medical Imaging Solutions GmbH) within an environment providing a user friendly interface with predefined analysis protocols, viewing facilities adapted to EEG signals and import facilities for a wide range of EEG data formats.
CONCLUSIONS: We conclude that the performance of the method and its implementation are acceptable for off-line automatic detection of epileptic seizures in scalp EEG. The method is fast and therefore is applicable for on-line implementation.
Our aims for the future are further improvement of the performance and on-line implementation of the method.
Furthermore, recent results indicate that a slightly modified version of the method may be used to detect pre-seizure EEG changes.
(Disclosure: Salary - M. D mpelmann is employee at ANT Software and eemagine Medical Imaging Solutions. He receives salary from ANT Software. He was not involved in the evaluation of the method. eemagine Medical Imaging intends to exploit the presented method in comme)