SPARSE ANOMALY REPRESENTATIONS IN VERY HIGH-DIMENSIONAL BRAIN SIGNALS
Catherine Stamoulis · 2018
Across scales of organization, brain activity is inherently sparse. This is also the case for transiently occurring signal abnormalities associated with neurological disorders, such as epilepsy. Consequently, for the purpose of characterizing these abnormalities, very high-dimensional brain signals may be represented as sparse combinations of the elements of a comprehensive (overcomplete) dictionary. Such a dictionary may be estimated (learned) from the dataset(s) of interest. However, given the statistical, spectral and signature heterogeneity of brain signals recorded over long periods of times, the size of the dictionary may be suboptimal, particularly in terms of its size. In this paper, signal-specific, dataset-specific and individual-specific anomaly dictionaries, estimated via the K-SVD algorithm from noninvasive high-frequency brain signals collected continuously over several days are explored. It is shown that signal-specific dictionaries may yield substantially more accurate representations than those estimated by combining training signals from multiple electrodes.