Abstracts

SEIZURE PREDICTION USING MACHINE LEARNING ON BIVARIATE FEATURES FROM EEG

Abstract number : 1.048
Submission category : 3. Clinical Neurophysiology
Year : 2008
Submission ID : 8311
Source : www.aesnet.org
Presentation date : 12/5/2008 12:00:00 AM
Published date : Dec 4, 2008, 06:00 AM

Authors :
Piotr Mirowski, D. Madhavan, Y. LeCun and R. Kuzniecky

Rationale: Recent research suggests that electrophysiological changes develop minutes to hours before the actual clinical onset in focal epileptic seizures. State-of-the-art seizure prediction research has investigated statistical analysis of features derived from intracranial EEG. However, no reliable seizure prediction method is ready for clinical applications. In this study, modern machine learning techniques were used to successfully predict seizures from several bivariate EEG features proposed in the literature. Methods: Continuous preictal and interictal EEG recordings of 21 patients suffering from intractable focal epilepsy were obtained from the Freiburg EEG database. Intracranial EEG was acquired with a combination of 3 focal and 3 extrafocal grid-, strip-, and depth-electrodes, sampled at a 256 Hz, digitized at 16 bits, and further filtered to remove power line noise and dc component. First, a large number of bivariate features were computed on the EEG, for all pairs of channels, and at all times (and all frequencies). Bivariate features measure synchronization of EEG channels, and include cross-correlation, nonlinear interdependence, difference of Lyapunov exponents and phase-locking synchrony. Features were aggregated into spatially and time-varying patterns of features. Second, nonlinear classifiers were trained to discriminate between interictal and preictal patterns of features. Machine learning methods included convolutional networks and support vector machines. Each classifier was trained for a specific patient and a specific type of features. In-sample optimization of the parameters and out-of-sample testing ensured the validity of the predictions. In-sample data consisted of 2/3 of interictal patterns and the earlier 2/3 preictal epochs, whereas out-of-sample data consisted of 1/3 interictal patterns and the later 1/3 preictal epochs. The preictal period was defined as the 2 hours preceding a seizure. Results: Classification performance of 100% sensitivity (at least one feature pattern classified as preictal before each seizure) and no false alarms were obtained on all the training dataset. On the testing dataset, and for each of the 21 patients, at least one method predicted 100% of the seizures on average 60 minutes before the onset, with no false alarm. Convolutional networks predicted all seizures without false alarm on 20 patients out of 21, and support vector machines on 11 patients (on 17 patients with less than 1 false alarm every 4 hours). Cross-correlation features enabled good seizure predictions with less than 0.3 false positives per hour on 13 patients out of 21; nonlinear interdependence on 19, wavelet phase-locking synchrony and phase difference entropy on 14 patients, wavelet coherence on 15. The difference of short-term Lyapunov exponents was the weakest, as it worked only on 3 patients. Conclusions: These extremely encouraging results give hope for clinical applications of seizure prediction, such as implantable devices capable of warning the patient of an upcoming seizure or implanted drug-delivery devices. Further evaluation is being conducted on additional patients.
Neurophysiology