Abstracts

Rationale: Visual review of long-term continuous EEG is a labor-intensive process. Many algorithms for automated seizure detection exist, but none approach the specificity and sensitivity of an experienced electroencephalographer. We have developed a reliable automated tool for seizure detection from EEG in epileptic rats, using a modified Kantz algorithm. Methods: Adult male Sprague Dawley rats were made epileptic using the lithium/pilocarpine model. After the rats developed chronic epilepsy, an electrode array was implanted in 10 cranial locations (Fig. 1). EEGs were continuously recorded with a sampling rate of 512 Hz. A computer code written in MATLAB was used to analyze the EEG and produce automated seizure detection results, which were compared with seizures identified by visual inspection.Our seizure detection algorithm utilizes the computation of maximum Lyapunov exponents (Lmax) through the use of a modified Kantz algorithm. Iasemidis et al. showed that Lmax can be used to predict seizures since the brain goes through state changes between chaotic and ordered as it progresses through a seizure. However, in the Iasemidis study, seizure detection was not optimal, likely due to the use of Wolf’s algorithm, which yields poor estimates of the Lyapunov exponents, especially in the presence of noise. In our work, we use Kantz’s algorithm tuned for seizure detection. This method requires an exponentially greater number of trajectories to be measured in the phase-space so we optimized it by taking advantage of advances in parallel computations. Results: Our method shows that during a seizure the brain transitions through different stages of chaoticity, as measured by Lmax. Fig. 2, shows this transition going from a baseline chaotic inter-ictal stage, to a more ordered pre-ictal stage, to a highly chaotic ictal stage, and finally settling at the baseline chaotic post-ictal stage. Using our method, we were able to achieve 100% sensitivity and 100% specificity in most data sets collected over ~30 weeks combined from three animals. In the remaining sets, we achieved 100% sensitivity for all data; the lowest specificity we obtained with data artificially corrupted with noise and stimulation artifacts was 94.47%. In every case, the Wolf algorithm yielded significantly poorer results than our proposed method. Conclusions: With our parallelizable method for computing the maximum Lyapunov exponent, we were able to achieve extremely high sensitivity and specificity of automated seizure detection. The method uses exponentially more information than the Wolf algorithm and is capable of delivering more reliable detection results. The cost of time is mitigated through the use of multi-core systems to make the detection algorithm operate in real-time. This will be a useful tool in laboratory investigation but still needs to be validated in the study of human epilepsy. Funding: This work was supported by NSF Award ECCS-1102390 and the Barrow Neurological Institute.