CAN WAVELET TRANSFORM ANALYSIS IMPROVE AUTOMATED SPIKE DETECTION?

Objectives: The purpose of this project is to use wavelet transform analysis to identify and classify two different epileptic EEG spikes, Rolandic and generalized. The performance of the wavelet algorithm was evaluated and compared to time correlation analysis.

Methods: The data are taken from interictal EEG segments of 5 pediatric patients with generalized spikes and 5 pediatric patients with benign Rolandic spikes. The project developed a method using wavelets and correlation analysis to assign a score to each point in an EEG segment, which is a measure of the likelihood that a spike exists. Using the EEG files for patients with known spikes, probability distribution curves were generated for the identification scores for segments with and without spikes. From this information, maximum likelihood detection was used to estimate the performance of this matching technique, and compare it to results using a time domain analysis.

Results: The results show significant improvement with wavelets over conventional time correlation analysis for spike identification, with most maximum likelihood false-alarm rates below 0.15% and detection rates above 96%. Wavelet spike identification proved especially useful in driving the false-alarm rates down, which is important in real-world implementations. It also produced rates of over 96% for correctly classifying the type of known spikes in most cases. Conclusion: These data show wavelet transform analysis to be effective in spike identification. The wavelet algorithm used in this study was more sensitive and specific than time correlation and allowed for correct distinction between Rolandic and generalized spikes.