Shrinkage Classification for Overlapping Time Series: An interpretable method for mapping stimulus-differentiated evoked response

Abstract

Multivariate time series from neural electrophysiological recordings are a rich source of information about neural processing systems and require appropriate methods for proper analysis. Current methods for mapping brain function in these data using neural decoding aggregate information across space and time in limited ways, rarely incorporating spatial dependence across recording locations. We propose Shrinkage Classification for Overlapping Time Series (SCOTS), a neural decoding method that maps brain function, while accounting for spatio-temporal dependence, through interpretable dimensionality reduction and classification of multivariate neural time series. SCOTS has two components: first, overlapping clustering from sparse semi-nonnegative matrix factorization gives a data-driven aggregation of neural information across space; second, wavelet-transformed nearest shrunken centroids with sparse group lasso performs multi-class classification with selection of informative clusters and time intervals. We demonstrate use of SCOTS by applying it to human intracranial electrophysiological and MEG data collected while participants viewed visual stimuli from a range of categories. The method reveals the dynamic activation of brain regions with sensitivity to different object categories, giving insight into spatio-temporal contributions of these neural processing systems.