A HYBRID MULTI-FEATURE AND MULTI-CHANNEL ANALYSIS OF CONTINUOUS, PROLONGED INTRACRANIAL EEG DATA FOR SEIZURE PREDICTION
Abstract number :
1.123
Submission category :
Year :
2002
Submission ID :
1643
Source :
www.aesnet.org
Presentation date :
12/7/2002 12:00:00 AM
Published date :
Dec 1, 2002, 06:00 AM
Authors :
Maryann D[ssquote]Alessandro, George Vachtsevanos, Ho-seon Lee, Greg Worrell, Steve Cranston, Denise Sewell, Rosana Esteller, Javier Echauz, Gordon Baltuch, Brian Litt. Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA; Dep
RATIONALE: Based upon retrospective studies, there is consensus that a preictal period of at least 20 minutes exists in mesial temporal lobe epilepsy. Given the heterogeneity of epilepsy, it is unlikely that a single quantitative measure will be useful in identifying this period in all patients. As a step towards practical implementation of seizure prediction in humans, we present an individualized method for selecting EEG features and electrode locations for seizure prediction focused on precursors that occur within ten minutes of electrographic seizure onset.
METHODS: Using a systematic approach to feature selection, classification, and validation, we analyzed a 94-hour hospital stay, containing 17 seizures with this method. Four other patient analyses are in progress. The method was trained on 8-minute data epochs leading to 4 seizures, 4 hours of baseline EEG data, and then a set of features and electrodes were selected using a genetic algorithm and probabilistic neural network to optimize the method for individual patients. The trained system was then run on the rest of the patient stay and its performance was assessed.
RESULTS: The complete methodology has been evaluated for one patient. In this patient, the best channel selected by the genetic algorithm was contralateral to the focus channel. The best feature using this channel was the [dsquote]mean of the mean of the curve length,[dsquote] a derived measure related to signal complexity. Validation of this result demonstrated a sensitivity of 100%, with 0.71 false positives per hour (FPh) over the entire 94 hour record. Average prediction time, was 73.57 seconds with a standard deviation of 25.53 seconds. The results from all five patients, including algorithm outputs and receiver operating characteristic (ROC) curves, will be presented.
CONCLUSIONS: The output of the probabilistic neural network (PNN) classifier for the first patient demonstrates that a system based upon multiple features and electrode sites tailored to individual patients produces promising results for seizure prediction. That the electrode site selected as best for short-term prediction was contralateral to the focus channel may indicate the importance of brain outside of the ictal onset zone in generating clinical seizures. This method requires further refinement and validation, but may provide one way of dealing with the heterogeneity of seizure types and individual patterns in seizure prediction technology. It may also provide insight into brain mechanisms that underly seizure generation.
[Supported by: This work is supported by funding from the Whitaker foundation, Epilepsy Foundation, American Epilepsy Society, University of Pennsylvania Research Foundation, and the National Institutes of Health grants #R01NS041811-01 and #MH-62298RO1.]; (Disclosure: Stock - Drs. Echauz, Esteller, Litt, and Vachtsevanos have been awarded a small number of stock options (less than 0 25% of the company s total value) in NeuroPace Inc, resulting from licensure of patents to the company. These patents are all owned, singl)