Application of time-scale techniques to detection of epileptiform activity in the EEG (2000)
Type of ContentTheses / Dissertations
Thesis DisciplineElectrical Engineering
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury. Electrical and Electronic Engineering
AuthorsGölz, Hansjergshow all
Epilepsy is a neurological disorder for which the electroencephalogram (EEG) is the most important diagnostic tool. Detection of interictal (between-seizure) epileptiform activity in the EEG is, however, complicated by various artifacts and spike-like transient waveforms in the normal background EEG. Approaches to the automatic detection of epileptiform discharges (EDs) require sophisticated signal analysis methods. We have developed a new system for the detection of EDs spikes, sharp waves, and spike-and-wave complexes in the EEG. It is based on the continuous wavelet transform (CWT) and incorporates statistical pattern recognition and 3D spatial source analysis. Wavelet-based approaches to signal analysis include the discrete wavelet transform (DWT), the CWT, and matching pursuit. Both DWT and CWT are based on a prototype filter, defined by a wavelet function applied at various scales, that can lead to accurate localization in both the time and frequency domains. Wavelet-based methods are more appropriate for the analysis of non-stationary signals - such as the EEG than the more conventional standard and short-time Fourier transforms. Both the DWT and CWT have been applied previously to the spike detection problem but the CWT with a complex-valued wavelet is considered superior due to translation-invariance and independence from the phase of the transient. Statistical pattern recognition covers a wide range of classification techniques including artificial neural networks and both linear and quadratic discriminant functions. Linear and quadratic discriminant functions perform well on the separation of two non Gaussian distributed samples from multidimensional feature spaces. In many medical diagnostic decision problems there is a large number of healthy controls but only a small number of definitively diagnosed patients. The resulting unequal sample sizes pose a problem for classification but can be counter-balanced by a bias term in the discriminant functions. An alternative interpretation of the output of discriminant functions can be gained by using a Bayesian approach to provide an additional measure of confidence in the decision process. A quasi-regular geodesic sampling grid of hypothetical current sources (dipoles) in the source space (brain volume) leads to a new paradigm for the estimation of source location and source orientation. The orientation estimate forms the basis of a new characteristic spatial feature of epileptogenic spikes. The wavelet-based system can be separated into two distinct stages. Stage 1 detects candidate epileptiform discharges (CEDs). The magnitude of wavelet coefficients of a single scale of the CWT are evaluated on a logarithmic scale to adaptively estimate the background activity. Sudden deviations from the background trigger CED detections. In Stage 2 features of CEDs are extracted from their multichannel multiscale context. Important features are derived from the multiscale edges and ridges of the CWT. A linear discriminant function with a Bayesian approach is applied to obtain the probability that a single-channel transient is epileptiform. The source orientation of each CED is estimated with the dipole model and another linear discriminant function is used to estimate the spike probability of the multichannel CED. A training set of 53 EEGs (13 epileptiform and 40 normal) was used in the design and training of the system. The system was evaluated using a test set of 53 EEGs (14 epileptiform and 39 normal). This indicated a mean sensitivity of 57.9%, a mean selectivity of 62.5%, and a mean false detection rate of 6.8 false detections per hour. Detection results compare favourably with those obtained with two other published systems (Hybrid I and Hybrid II) on the same data set. The system is sensitive to weak EDs but fails to reject artifacts in several recordings. It is considered that the system would benefit from the addition of a further stage able to make use of wide-sense temporal context of detections. In summary, it has been shown that the CWT can function as a very sensitive time-adaptive multichannel detector for CEDs. The multichannel multiscale context of EDs provides important features for discrimination of EDs from artifacts via a linear classifier. When used in conjunction with a source-orientation estimate from a dipole model, the resultant system shows considerable potential as an accurate spike detector for applications in clinical neurophysiology.