Abstracts

Rationale: In epilepsy patients and animal models of epilepsy, seizure monitoring is essential for accurate diagnosis and evaluation of antiepileptic treatments. Presently, there are few validated software for automatic analysis of seizure recurrence in large data sets. Quantification of seizure frequency and severity is integral outcome indices of epileptogenesis. Here we describe a predictive data analytics protocol for automated quantification of epileptic seizures and the outcomes were validated by an independent observer. Methods: The protocol consists of a two-step algorithm to find potential epileptic seizures. First, the algorithm extracts local energies in different frequency bands of the EEG recordings, and employs an ensemble machine learning method called random forests to identify the presence or absence of an epileptic episode over 1 second intervals. Second, it uses a high-level rule sets to parse the outputs of the random forest method to reduce the number of artifacts misclassified as epileptic seizures. This two-step protocol was implemented in Matlab and R to automate the epileptic seizures detection process. Results: The results suggest that our new two-step protocol to be extremely accurate and very efficient for the detection of epileptic seizures. The algorithm was tested on two hour long EEG recordings from 20 different rat specimens. 28 of 31 epileptic seizures, which were manually detected, were confirmed. The protocol also identified new seizures (true positive) which were overlooked by the current protocols. We found the most sensitive features for detecting seizures to be variance in the EEG signals extracted over 1 sec and 0.25 sec time windows, and energies over the 0-4 Hz frequency band extracted over 1 sec time window. Conclusions: Overall, the new protocol achieved an accuracy rate of 91%, with a sensitivity of 97% and a specificity of 87%. This software protocol allows for sensitive, high-throughput counting of epileptic seizures recorded in any file formats, and thus provides an automated quantitative tool for monitoring of epilepsy in animals and patients. Funding: N/A