Abstracts

Techniques from nonlinear dynamics are potentially valuable for prediction of epileptic seizures based on preliminary results with reported prediction times of up to several minutes before electrographic onset. These results must be validated using a database containing more subjects and longer time series in order to establish their predictive sensitivity and specificity. Here, we examine the correlation integral (CI), a measure commonly used in seizure prediction algorithms.
We analyzed a dataset consisting of 1146 hours of single channel ECoG from 10 patients who underwent invasive evaluation for epilepsy surgery. The signals were amplified and digitized to 240 Hz with 10 bits of precision. The generic Osorio-Frei seizure detection algorithm (Epilepsia 39(6):615-627) was run continuously on the data for the purpose of automated seizure scoring and comparison with the correlation integral methods. After clustering together events seperated by less than 60 seconds, there were 502 automated detections, of which all but 65 were scored as seizures through expert visual analysis (I. Osorio). The signal was time delay embedded with a delay time of .07 s, and embedding dimensions of 5 to 35. CI was computed continuously for a full range of characteristic length scales in 15 second sliding windows overlapped by 5 s, with a Theiler correction of .125 s, which was chosen based on analysis of interictal autocorrelations. To assess long-term predictive ability, the resulting time series were statistically compared using the Kolmogorov-Smirnov (K-S) statistic in three 15 minute epochs: 0-15 minutes, 30-45 minutes, and 120-135 minutes prior to seizure. Also, short-term prediction (less than 15 minutes prior to seizure) was evaluated through analysis of overlaid pre-ictal correlation integral time series. Seizure detection ability was determined through analysis of false and true positives and compared against the Osorio-Frei maximum ratio (MR).
In most cases, the distributions of CI values in the three epochs were significantly different (p [lt] .05), but the K-S statistic was still too low for a specific correlation integral value to be assigned correctly to its corresponding probability distribution with high confidence. While no prediction was apparent, CI showed improved ability to detect seizure activity over the MR for two patients, due in part to a larger window size. Of the remaining 8 patients, the methods performed approximately equivalently on two, and MR performed best on the remaining 6. Due to a lack of a linear scaling region, the correlation dimension could not be computed reliably from CI either ictally or interictally.
We do not find evidence of prediction ability with CI. However, methods based on this measure may have some value for seizure detection, but this must be weighed against the computational intensity of these methods and the loss of detection speed associated with larger window sizes.
[Supported by: NIH grant #1R43NS43100-01]