Abstracts

Rationale: Scalp EEG-video monitoring is the standard of care for the pre-surgical evaluation of patients suffering from medically intractable epileptic seizures. The efficiency of this time and labor intensive task depends on how quickly and accurately the recorded seizures can be identified in multi-day recordings. Therefore, several computer applications offering automated seizure detection (ASD) have been used in epilepsy monitoring units (EMUs). However, due to their high false detection rates (> 10 per day), this ASD software has had limited clinical utility. We previously reported initial results of a scalp EEG based ASD algorithm in a small training dataset. The algorithm detects seizures by extracting multi-dimensional spatio-temporal patterns of signal complexity, frequency, and amplitude variation. The objective of this study is to complete the initial study by expanding the training dataset and to validate the algorithm performance in a separate test dataset. Both training and test results are compared with commercially available ASD software. Methods: Multi-channel scalp EEG recordings obtained from 47 patients (total length ~ 3653 hours with a total of 141 seizures) with a history of intractable epileptic seizures were used for algorithm training. A separate set of 32 scalp EEG recordings, containing a total of 75 epileptic seizures in 2,388 hours of recordings, were used for algorithm performance validation. For each subject, the entire long-term EEG recording (no clipping) was included in the analysis for performance evaluation. This method created the most representative clinical data by including as many patients’ normal physiological conditions and recording artifacts as possible. We evaluated the performance of seizure detection by estimating the detection sensitivity and the false detection rate per 24 hours. After training process, all the detection criteria and parameters were fixed in the validation study except the parameter that was used for generating detection receiver operating characteristic (ROC) curves. The detection performance was compared with one commercially available ASD application (Persyst Reveal®) that is widely used in EMUs. Results: ROC curves show the detection performance from the least sensitive to the most sensitive by changing the pre-determined ROC parameter. In the training study, the test ASD software performed significantly better than Reveal at any sensitivity level (paired-T test of false detection rate with respect to different level of sensitivity; p < 0.001). Using the same ROC parameters in the validation study, again, the test software performed significantly better than Reveal at any sensitivity level (p < 0.001). According to the ROC curves, the Reveal can reached 80% sensitivity with about 5.7 false detections per 24 hours of recording, while the test algorithm gave 83% sensitivity with about 1.6 false detections per 24 hours of recording. Conclusions: The results from this study suggest that the proposed ASD algorithm can be clinically useful for detection seizures in long-term scalp video-EEG recordings.