Abstracts

RATIONALE: We previously described the energy as a stand-alone feature, and the accumulated energy as a stand-alone, derived feature for predicting epileptic seizures (Neuron, Vol. 30, 1-20, April 2001). In an effort to trace the earliest precursors back to their site(s) of origin, we expand our work to multiple features and multiple channels by introducing a genetic approach to obtaining the optimal feature for predicting seizures. A genetic approach to this problem automates and systematizes optimal feature selection.
METHODS: Thirty-six 15-minute Intracranial EEG records comprising 16 seizure and 20 baseline records were evaluated on a patient specific basis to predict seizures. The 16 seizure records consisted of 10 minutes of pre-ictal EEG and 5 minutes post-ictal. Each of six first level features were combined with 14 statistical functions using 22 IEEG channels resulting in over 4500 derived channel-feature combinations in the genetic algorithm[scquote]s search space for each first level feature evaluated. Baseline and pre-seizure records used Fischer[scquote]s Discriminant Ratio as the objective function for determining survival of a feature combination. 70% of the data was used for training, with 30% reserved for testing. The channels and derived features were configured into a digitally represented chromosome prior to performing a genetic search to identify the optimal feature combination to achieve prediction.
RESULTS: Our results demonstrate the utility of a genetic algorithm for selecting the optimal feature or set of features to be used for classification. One dimensional (1D) scatter plots and class conditional distribution functions comparing the pre-seizure and baseline records of the test data clearly distinguish the two classes in 100% of the records. Two dimensional scatter plots show that combinations of optimally derived features improve prediction over using one derived feature and warrant further investigation.
CONCLUSIONS: An optimal dervied set of features extracted via subchromosome analysis improved prediction performance and can be used to identify electrode sites critical for seizure prediction. A genetic approach to optimal feature selection may be useful in identifying appropriate features and electrode sites for an implantable therapeutic epilepsy device.
Support: American Epilepsy Society; IntelliMedix; Epilepsy Foundation; Whittaker Foundation.
Disclosure: Equity - Drs. Echauz and Esteller Litt and Vachtsevanos our cofounders of IntelliMedix.