Abstracts

Rationale: We propose a novel signal feature that allows early and reliable online epileptic seizure detection in combination with standard ECoG recording systems. Seizure detection with a latency of a few seconds between unequivocal seizure onset (UEO) and automatic seizure alert would be of great benefit in clinical environments where medical staff has to react to upcoming seizures. Furthermore, online detection of epileptic seizures is a key technology for intervention systems that will be able to interrupt seizures, e.g. by electrical stimulation or local medication.Methods: In contrast to other approaches our signal feature measures synchronization and energy of ECoG channels simultaneously. This is achieved by forming the inner product of dominant projections to a subspace of periodic signals derived from individual ECoG channels. Due to the nature of the inner product we thus capture energy variations as well as the relative phasing of dominant projections where the latter corresponds to channel synchronization. This method turned out to be especially suitable to detect synchronization of periodic phenomena often seen in the early phase of epileptic seizures. We applied the new feature in our well proven online seizure detector which utilizes an adaptive threshold to obtain decisions. In particular, it performs adaptive estimation of signal parameters and decision feedback. The adaptive parameter estimation allows to substitute intensive offline training of the system with a short initialization time of a few minutes at the beginning of signal processing. Moreover, our system does not rely on any manual channel or data preselection.Results: In our experiments, we used ECoG data from three patients with temporal lobe epilepsy (see Table 1). Data were recorded from 28 up to 32 grid electrodes during 21h and includes 10 seizures with a length between 20 and 180 seconds. Focus channels and UEOs were predetermined via visual inspection by clinical experts. For all three patients we found false detection rates between 0 and 0.18 alarms per hour and mean latencies between 11 and 40 seconds with a sensitivity of always 100% (see Tab. 1 for details). Alerts occurred mainly in the focus channels. Averaging results for the three patients gave a false detection rate of 0.14 per hour, a sensitivity of 100% and a mean latency of 21 seconds with respect to the UEO.Conclusions: Our findings show that reliable online detection of epileptic seizures from ECoG data can be accomplished with a feature which is sensitive to synchronization and energy of ECoG channels. The evaluation of our feature on 21h ECoG data including 10 seizures suggests the application in clinical environments. We obtained good system performance without any manual channel or data preselection. To trigger intervention systems that will be able to interrupt seizures, the detection latencies have to be decreased. This could be done easily at the expense of false alerts per hours. Since this is unacceptable in clinical applications, further research will be necessary to reduce latencies without sacrificing detection rates.