Evaluating Shift-Invariant Dictionary Learning Techniques for Automated EEG-Based Epilepsy Diagnosis

• Daniel England, Electrical Engineering, University of Delaware

• Austin Brockmeier, Department of Electrical and Computer Engineering, University of Delaware

The goal of this project is to use machine learning to identify waveforms related to epilepsy in EEG scans, and use the occurrence patterns of the waveforms to assist in computer-aided epilepsy diagnosis. This is important because epilepsy is the most common neurological condition in children, and manual examination of routine scans are often inconclusive, requiring additional follow-ups and delayed diagnosis. By uncovering waveform patterns associated with epilepsy, this project aims to improve accuracy, reduce diagnosis time, and enable more targeted interventions for individuals with epilepsy. Specifically, we build on existing work, implementing, and evaluating the effectiveness of modifications of dictionary learning techniques that learn a set of waveforms that can be used to reconstruct EEG windows by shifting and scaling the waveform. Six different shift-invariant k-means clustering algorithms were tested. These algorithms differ based on their waveform update step: standard averaging, Rank-1 Singular Value Decomposition (SVD) of aligned window segments, and weighted averages using Non-Negative Least Squares (NNLS). Each of these three algorithms were tested with and without spherical normalization of the windows to create the main six algorithms. We initially test this on synthetic EEG signals created with known waveforms (Gabor/Morlet wavelets) and varying levels of noise. The learned waveforms for each algorithm are compared to true waveforms using the cosine similarity: the similarity of the best matched learned waveform for each known waveform is computed. Through analysis of the cosine similarities of all six algorithms at a variety of different signal to noise ratios, we have determined that the best performing algorithm is rank-1 singular value decomposition without spherical normalization. We will now test this on real human (and preclinical mouse models) EEG signals and see if the improvement in waveform learning translates to improved diagnosis using machine learning epilepsy classifiers.