Abstracts

Rationale: Epilepsy is a highly prevalent neurological disorder that affects over 65 million people worldwide.  Approximately 1 in 26 Americans will be diagnosed with epilepsy in their lifetime. Even though there are currently numerous antiseizure drugs (ASDs) for the symptomatic treatment of epilepsy, an estimated 38% of patients do not have full seizure control on their current ASD regimen. In the search for more effective therapies for the treatment of pharmacoresistant epilepsy, chronic animal studies often use EEG to capture and quantify seizures. Detecting seizures in EEG data typically relies on human reviewers, which is not only subject to error and bias, but is also labor intensive. The use of computer automated seizure detection algorithms (SDAs) allows for increased efficiency and throughput in experiments where seizures must be detected within EEG data. In order to implement SDAs, they must be validated for their sensitivity and specificity. There is a wealth of previous literature that describes the sensitivity and specificity of individual SDAs. However, a major question that remains unanswered is whether or not a combination of multiple SDAs yields better seizure detection than each individual algorithm. Methods: To test the efficacy of a combined approach, four SDAs were evaluated; i.e., an autocorrelation method, FFT power-spectrum analysis, coastline-burst (line length) index, and a spike frequency analysis. The combined approach was evaluated using a database of 60 days of human-reviewed and annotated EEG data containing 279 seizures from 37 rats with temporal lobe epilepsy as a result of kainic acid-induced status epilepticus. The results of the human review were used to evaluate the sensitivity and specificity of the novel combined approach. Results: Using this method, 98% of reviewer-marked seizures were captured with a false positive rate of 53.52 false positives per real seizure. No individual algorithm was able to detect 98% of reviewer-marked seizures with a false positive rate of less than 413 false positives per real seizure. Subsequent implementation of the consolidated algorithms increased the overall efficiency of the manual EEG review process; e.g., only 17 minutes was required to review an average 24-hour file of EEG data from 12 rats. Manual review of EEG generally took a trained observer approximately two hours. Thus, implementation of the final seizure detection algorithm reduced manual review time by 86%. Conclusions: This increased throughput of chronic animal studies as well as minimization of error and bias due to human review would greatly enhance a laboratory’s ability to evaluate the EEG of rats with epilepsy and assess the impact of treatment on overall seizure burden. Funding: NINDS: HHSN271201600048C