The thesis treats methods for pattern recognition in multichannel electroencephalogram (EEG) signals, for application to diagnostics of epilepsy. Parts I-IV treat methods for feature extraction and clustering of EEG spikes, occurring between epileptic seizures, and part V presents a method for filtering seizure onset EEG signals. In part I, Hermite functions are used for parametric description of multichannel spikes, and the L-means method for clustering. The method is evaluated on real life signal sets. Part II treats the problem of making clustering methods statistically robust, and evaluates the fuzzy L-means method and the k-component graph-theoretic method in this respect. The effect of incorporation of the physical geometry of the electrode positions is also studied. In part III, the problem of alignment of the signals that are to be clustered is treated. Two new estimators are suggested and evaluated on simulated as well as real life signals. Part IV treats denoising of signal matrices under the assumption that the signals have low rank and the noise has full rank. A method is suggested consisting of representation in a basis of unit rank matrices, and algorithms for determination of such bases are given. Applicability to real life signals is demonstrated. Part V describes a technique for approximate Wiener filtering of non-stationary seizure onset EEG signals. The method relies on a time-frequency domain identity which is approximately valid for underspread stochastic processes, and estimates the Weyl symbol of the Wiener filter with aid of coherence functions between channels.
|Award date||1999 Oct 15|
|Publication status||Published - 1999|
Bibliographical noteDefence details
Place: Room E:1406, E-huset, LTH, Lund
Name: Koski, Timo
Subject classification (UKÄ)
- Electrical Engineering, Electronic Engineering, Information Engineering
- EEG signal processing
- statistical robustness
- feature extraction
- signal alignment
- Wiener filtering
- time-frequency analysis.
- Signal processing