multivariate time series analysis bijan pesaran center for neural science new york university
DESCRIPTION
Multivariate data, Imaging data types MEG/EEG/LFP fMRI Optical imaging Need to find spatial projection to reduce dimensionality Combine spectral and multivariate toolsTRANSCRIPT
Multivariate time series analysisBijan PesaranCenter for Neural ScienceNew York University
Overview
Singular value decomposition
Application to space-time data
Application to space-frequency data
Periodic stacking method
Multivariate data, Imaging data types
MEG/EEG/LFP fMRI Optical imaging
Need to find spatial projection to reduce dimensionality
Combine spectral and multivariate tools
Singular value decomposition Eigenvalue decomposition Calculates directions/modes in data space that
contain maximum variance
Singular valuespectrum
Spatialmode
Temporalmode
Application to space-time data
Spatial and temporal correlations modes of spatial correlation matrix
modes of temporal correlation matrix
Truncation defines subspace
Noise tailfor a pxq
matrix
fMRI data set with 1877x500data points, sampled at 5 Hz for 10 s.
2
Spectrum of temporal modes
Reveals physiological features across multiple modes
Application to space-frequency data
A geometric interpretation Project time series into a subspace Use an orthogonal basis set Local-in-frequency projection operator
Ttktk WfPXWfX ;;~,
Advantages of local-in-frequencybasis Combine information across this basis
Ensemble averaging
Choose properties of this basis Select time and frequency
Project onto multiple different subspaces centered on different frequencies
Space-frequency decomposition Local-in-frequency projection
Dimensionality reduction
Multivariate Coherence, Assess degree of low-dimensionality
fMRI data set
Complex-valued spatial modes
Spatial segregation of physiological modes.
1st order modes
fMRI example
Presence or absence of visual stimulus Digitization rate: 5 Hz Duration: 110 s Visual stimulation with red LED patterns (8
Hz).
Visual response in fMRI signal
No stimulus - dashed
Visual stimulus - solid
Visual response in fMRI signal
No stimulus
Visual stimulus
Spatially restricted visual response Coronal slice at the occipital pole
Optical imaging example Isolated procerebral lobe of Limax Presence of voltage sensitive dye Digitization rate: 75 Hz Duration: 23 s 600 um by 200 um
Optical imaging response
Limax procerebral lobe during olfactory stimulation
Principal spatial modes
Quantifying traveling waves Express leading spatial mode
2.5 Hz 1.25 and 2.5 Hz
x
y
LFP example Rubino et al. 2007 Electrode arrays in M1 and PMd of awake
monkey Digitization rate: 1 kHz Duraction Visual instructional cue response
Phase gradients in M1 PMd
M1 M1 PMd
Activity between 10 - 45 Hz
Waves reflect anatomical connections
Periodic stacking
If you have repeated measurements of the response to multiple stimuli, you can order your data to take advantage of the multitaper harmonic analysis methods that we have been shown.
1 2 1 2 1 2
Extraction of the average and differential dynamical response in stimulus-locked experimental data. J Neurosci Methods. 2005 Feb 15;141(2):223-9.
Odd harmonics - differences between responsesEven harmonics - average dynamics
Generalization to N stimuli: N’th harmonics are average dynamics, the rest are differences amongst the stimuli