PATIENT-SPECIFIC EPILEPTIC SEIZURE DETECTION USING EEG WAVELET PACKET ANALYSIS
Abstract number :
2.002
Submission category :
3. Clinical Neurophysiology
Year :
2009
Submission ID :
9719
Source :
www.aesnet.org
Presentation date :
12/4/2009 12:00:00 AM
Published date :
Aug 26, 2009, 08:12 AM
Authors :
Ali Shahidi Zandi, G. Dumont and M. Javidan
Rationale: We propose a novel wavelet-based algorithm for real-time detection of epileptic seizures using scalp electroencephalogram (EEG). This automated seizure detection procedure generating alarms shortly after the unequivocal seizure onset (UEO) has high sensitivity and specificity and is suitable for seizure monitoring units. Methods: We applied the seizure detection algorithm on ~35.5 hours of multi-channel scalp EEG from 9 patients with temporal lobe epilepsy (TLE), recorded according to the International 10-20 system in Vancouver General Hospital (VGH) with a sampling rate of 256 Hz. This dataset included 33 epileptic seizures and was visually inspected by an electroencephalographer to determine the UEOs. As a non-stationary signal, EEG can be appropriately analyzed using the wavelet packet transform (WPT) to localize the patterns of interest and discriminate them from the rest of the signal. In this study, the EEG signal from each channel was segmented into 2-second windows (epochs) with 1-second overlap and decomposed by the WPT using Daubechies-12 wavelet, resulting in a set of coefficients for each epoch. Using the coefficients of the last decomposition level obtained from a sample seizure period and a 30-min interictal reference, a measure was developed for each patient to quantify the difference between the seizure and non-seizure phases for a range of frequencies (1-30 Hz). Utilizing this measure, a frequency band representing the maximum difference between the two states was determined. This patient-specific frequency band was then employed to develop a regularity index as a measure of rhythmicity and an index sensitive to the ratio of the current epoch energy to that of a moving reference. In addition to these indices computed for each EEG channel, a multivariate index incorporating all EEG channels was also developed to improve the specificity of the algorithm by considering the coherence of EEG channels. Combining the three abovementioned indices, a normalized index, i.e. ranging from 0 to 1, was derived. The index value remains close to zero in non-seizure periods and increases during the ictal state.To inspect the variations of this index in each channel and generate alarms, a cumulative sum (CUSUM) procedure was employed as a robust statistics minimizing the detection delay for any fixed false alarm rate. Analyzing alarms from all channels, a seizure alarm was generated if at least 3 channel alarms (not in the same channel) occurred within 5s. Results: The proposed algorithm detected 31 seizures (~94% sensitivity) with a false positive rate of 0.36/h and median delay of 7s with respect to the UEO where the detection latency for ~74% of the detected seizures was less than 15s. Conclusions: The novel wavelet-based seizure detection method is able to detect the seizure onset in a very short time after the onset and has a higher sensitivity than the currently commercially available methods.
Neurophysiology