Learning representative waveforms to analyze, summarize, and compare long-term neural recordings
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Neuroscientists and clinicians are often faced with the tedious and time-consuming task of visually inspecting neurological time series that can last hours to days, and finding recurrent patterns that might be only a few milliseconds long. Furthermore, the experts use the morphological features of such patterns, like the sharpness of their peaks, to classify them, and to study physiological phenomena, like the coupling between two brain oscillations. Although standard methods, like the spectrogram and the Morlet wavelet convolution, are well-established and highly useful in the time-frequency analysis of neurological time series, they might not be sufficient or appropriate for such tasks. More recent convolutional dictionary learning algorithms aim to discover a set of recurrent waveforms that represent underlying physiological processes and are linearly combined to approximate the original time series. Those algorithms, however, are currently too slow to be run on long (i.e., with millions of samples) time series. ☐ In this dissertation, we developed a set of methods to either find recurrent patterns from long neurological time series, through an energy-guided search and a density-based clustering, or to learn approximations of such patterns, through a k-means algorithm that is invariant to local temporal shifts. Furthermore, we used the dictionary of waveforms learned to build a bag-of-waves representation of the time series, akin to the bag-of-words model in natural language processing, to perform time series classification. ☐ We achieved high classification performance in two classification tasks, one binary and one multi-class, showing that our methods can learn features from long neurological time series that are not only predictive but also highly interpretable.
Description
Keywords
EEG analysis, Neurological time series, Neural recordings, Data mining, Brain oscillations, Time-frequency analysis