Abstracts

Rationale: To determine the efficacy of different agents in the prevention or mitigation of epilepsy, it is necessary to identify the presence of seizures in prolonged (greater than 1 month) video-EEG recordings. Currently, most investigators employ expert personnel to determine the presence and frequency of seizures, resulting in both large time commitments as well as great expense. In this presentation, we describe an automated method that can be used to accurately and efficiently identify both latency period and seizure frequency. Methods: Rats used in this study were implanted with bilateral subdural electrodes above temporolimbic cortex (1) prior to and (2) 4 weeks after pilocarpine treatment to determine the accuracy of the algorithm for detecting the initial spontaneous seizures and late spontaneous seizures >1 month after pilocarpine-induced status epilepticus. Both sets of rats were monitored using video-EEG for periods of approximately 1 month. The EEG recordings were then reviewed by an expert technician with Racine seizure stage being confirmed on video recording. The autonomous algorithm was implemented using subroutines written in Visual Basic, implemented on a Windows platform. The software exploits characteristics present in EEG recordings of seizures: (1) a change in frequency or amplitude, (2) an increase in coherence between the two recorded channels, (3) a decrease in dimension of the signal during a seizure, (4) characteristic evolution of the seizure, and (5) the presence of EEG changes following the seizure. The algorithm developed uses a novel, efficient technique for identifying changes in frequency and amplitude in which a local mean line is constructed. The areas above and below the line and line crossing distances are determined and averaged over a fixed time interval. A metric is then determined based on these parameters. The calculation of the metric is independent of the individual rat. Results: A total of over 1000 hours of video-EEG were analyzed by the expert and using the computer algorithm. There were a total of 1260 seizures identified by the expert. The algorithm identified 100% of the seizures. The algorithm had a false positive rate of 0.063 per hour. False positives were more likely in the record early after status epilepticus, or at times when the amplitude and duration of seizures decreases. Signal artifact substantially increased the number of false positives. Conclusions: These results demonstrate that we have developed an accurate algorithm for identifying seizures in prolonged video-EEG recordings employing several novel techniques. This algorithm will significantly decrease the time and cost of identification of seizures in prolonged EEG recordings. Further testing will be performed to demonstrate efficacy in other animal models and with signals containing significant artifact. Additional work will also be performed to create an adaptable network which will learn the pattern for each individual rat, thereby reducing the number of false positives. Acknowledgements: Work was performed using support from grants 5 K08 NS053610-05 (AMW) and NIH R01 NS051710 (ABK).