Abstracts

RATIONALE: An implantable warning or treatment device for epilepsy requires the ability to detect seizures from intracranial EEG with maximum accuracy, computational efficiency, minimal delay, and demonstrated performance on preictal and baseline data far removed from seizures. We present a seizure detection algorithm, based upon three quantitative features fused by a probabilistic wavelet neural network, which together fulfill these criteria. We present the performance of this system on preictal and baseline data from six patients with temporal lobe epilepsy implanted with intracranial depth and strip electrodes as part of their evaluation for epilepsy surgery.
METHODS: Data from six patients were used in this analysis, including 50 one-hour seizure records and 69 one-hour baseline records, more than 4 hours from any seizure onset or termination, yielding a total of 119 records analyzed. This system is designed based upon a pattern recognition approach that encompasses preprocessing, processing, classification, and validation. Feature extraction and selection were performed off-line within a designed methodological enviroment that encompassed the analysis of 27 features from seven different domains. Different from previous work, an [dsquote]optimal[dsquote] feature vector was selected out of the 27 features analyzed, using a modified add-on algorithm. The feature vector determined included the fractal dimension by Katz, the weighted nonlinear energy, and the absolute value of the fourth wavelet coefficient. The length of the sliding window used to generate each feature was optimized for each patient and each feature using the K-factor, an objective function aimed at maximizing the separability between the seizure onset class and the non-seizure class. In the classification stage, a probabilistic neural network was used as the decision block. A crossvalidation scheme was employed to validate results.
RESULTS: The average seizure electrographic onset detection delay was 1.76 seconds with zero false negatives, with an average of 1.02 false positives per hour, and an average clinical onset prediction time of 11.26 seconds, over the entire data set.
CONCLUSIONS: We present a systematic methodology for implementing patient-specific seizure onset detection from prolonged intracranial EEG recordings. These results indicate that the performance of such a system is approaching the clinical requirements outlined for a successful implantable device. Further research is underway to enhance detection features and improve the false positive detection rate of this system. Our results demonstrate that sliding window length optimization and the complementarity of detection features enhances detector performance.
Support: This work was funded by the American Epilepsy Society, the Epilepsy Foundation, and the Whitaker Foundation. Subsequent to this work, Drs. Esteller, Echauz, Vachtsevanos and Litt co-founded IntelliMedix, a small company which develops intelligent algorithms for medical applications.
Disclosure: Salary - Drs. Esteller and Echauz are employees of IntelliMedix, Inc. Equity - Drs. Esteller, Echauz, Vachtsevanos, and Litt are co-founders of IntelliMedix, Inc. Ownership - IntelliMedix, Inc. Materials - See above.