Abstracts

Rationale: An automatic seizure detection algorithm is important to reduce the workload posed by the large amount of EEG data recorded in an epilepsy monitoring unit. The development of such an algorithm is demanding, due to mainly two factors: the inter-patient variability of seizure characteristics and artifacts contaminating the EEG data. The limitation of the currently available detection systems is their high false detection rate (0.5/h). Methods: The algorithm presented here extracts features from the EEG, which guide the neurologists visually, when reading EEG data. These features include change in frequency distribution, asymmetry between the hemispheres, morphology characterized by repetitive sharp waves, and evolution in amplitude, frequency and topology. Therefore, the main steps of the algorithm are analysis of power spectral density in different frequency bands, brain symmetry index modified for contralateral channel pairs, extraction of wave segments, and retrieval of a series of repetitive waves with gradually evolving amplitude. The algorithm was tested on 680 hours of EEG data containing 26 temporal lobe seizures and 95 extratemporal lobe seizures from patients with refractory partial epilepsy, who underwent a presurgical evaluation. Results: The sensitivity for temporal lobe seizures was 84.4% while for extratemporal seizures 46%. Missed seizures either had short ictal pattern on EEG or were difficult to detect visually as well. The number of false detections was 0.15/h, and were mainly due to rhythmic, low-frequency patterns of sleeping and drowsiness. Conclusions: Our automatic seizure detection algorithm is able to mimic the visual interpretation of the human observer and can work real-time. It is a sensitive method for detection of temporal lobe seizures, but not extratemporal lobe seizures. An advantage of the algorithm is its high specificity, which could be further improved if sleeping and drowsiness patterns could be excluded. Seizures which are difficult to detect visually will also be missed by our algorithm. Methods revealing hidden patterns of EEG might be employed for such seizures, before applying the proposed steps of detection.