Abstracts

Rationale: The analysis of long-term EEG recordings required for pre-surgical workups of drug-resistant epilepsies is extremely time consuming and expensive. An automatic online seizure detection system could significantly reduce data review effort, and it could alert medical staff to a beginning seizure in order to set appropriate medical actions improving patient safety and to perform further systematic neurological testing during the seizure. Here, an online seizure detection system is presented, which is currently being evaluated within a clinical study at the Neurological Department Rosenh gel at General Hospital Hietzing, Vienna.Methods: The seizure detection algorithm is based on a periodic waveform analysis detecting ictal rhythmic EEG patterns. An adaptation module automatically adjusts the algorithm to a patient-specific, frequency-dependent EEG baseline. Thus, suspicious patterns lead to a seizure alert only if they stand out against a patient s normal EEG. The algorithm was evaluated off-line using 4.300 hours of unselected EEG recordings from 48 patients with temporal lobe epilepsy. All data were taken as recorded , i.e., without any data pre-selection. A total number of 186 electrographically visible seizures were included in the recordings from 38 patients. Results: The average sensitivity was 88% and the average false alarm rate was of 0.26 false alarms per hour (FA/h), which is approximately one false alarm within four hours. For 68% of the patients, 100% of the seizures were detected by the seizure detector, for 27% of the patients, sensitivities between 50 and 99% were achieved, and only for the remaining 5% the sensitivity was below 50%. The false-alarm rates were below 0.25 FA/h for 62% of the patients, between 0.25 and 0.5 FA/h for 27% of the patients, and between 0.5 and 1 FA/h for the remaining 11% of the patients. Conclusions: In long-term monitoring of epilepsy patients, automatic seizure detection systems must prove robustness against all kinds of signal artifacts and corrupted EEGs. In this study, the complete data of each patient were evaluated. Thus the results can be expected to carry over to clinical practice. For 89% of the patients, false alarm rates below 0.5 FA/h were achieved. This is less than one false alarm within two hours, which should be an absolutely acceptable value. For 1/3 of the patients we missed seizures, but only in 5% of the patients less than 50% of the seizures were detected. An automatic system cannot yet relieve personal from continuously monitoring patients during long-term recordings. However, for 2/3 of the patients, all seizures could be detected automatically. For these patients, the data review effort could significantly be reduced, and seizures might be noticed much earlier due to an automatic seizure alert, resulting in improved patient safety and diagnostic potentials.