An Algorithm for the Analysis of Temporally Structured Multidimensional Measurements

Analysis of multichannel recordings acquired with contemporary imaging or electrophysiological methods in neuroscience is often difficult due to the high dimensionality of the data and the low signal-to-noise ratio. We developed a method that addresses both problems by utilizing prior information about the temporal structure of the signal and the noise. This information is expressed mathematically in terms of sets of correlation matrices, a versatile approach that allows the treatment of a large class of signal and noise sources, including non-stationary sources or correlated signal and noise sources. We present a mathematical analysis of the algorithm, as well as application to an artificial dataset, and show that the algorithm is tolerant to inaccurate assumptions about the temporal structure of the data. We suggest that the algorithm, which we name temporally structured component analysis, can be highly beneficial to various multichannel measurement techniques, such as fMRI or optical imaging.