EEG waveform analysis by means of dynamic time warping
The feasibility of using dynamic time warping (DTW), a technique widely used in the area of automatic recognition of spoken words, to cluster EEG waveforms was studied. DTW utilizes dynamic programming techniques to compress and extend the time axes of two digitized time series to reduce the effects of minor differences in shape due to noise and normal, random shape fluctuations. The distance that remains after DTW can be used as a similarity index in a clustering procedure. Several modifications were implemented to make the DTW algorithm suitable, and efficient for this new application. A number of experiments were performed with simulated data (half-cycle sine waves and triangular waves) to gain insight in the specificity and sensitivity of DTW as a function of four wave shape parameters- frequency, amplitude, peak location, and phase. It was found that DTW based clustering could distinguish between waves only slightly different in frequency, amplitude, or peak location. However, DTW proved to be much more sensitive to phase changes. The DTW approach was also applied to real EEG data (spikes and sharp waves) and compared with two other methods. The first method, referred to as the heuristic approach, comprised the extraction of four features and using the Euclidean distance measure between these features as a measure of similarity. The second method, termed the peak-alignment approach, proceeded by first aligning the peaks of the two signals to be compared. If needed, zeros were padded to the shorter signal to make both signals of equal length. The sum of the absolute values of the amplitude differences was used as a similarity index. The results showed that the DTW approach yielded the most homogeneous clusters in terms of shape-similarity. The preliminary results presented in this thesis indicate that dynamic time warping based clustering is a viable approach to EEG waveform clustering.