Granger Causality Analysis of fMRI Data:
Techniques, Caveats and Applications
Xiaoping Hu
Wallace H. Coulter Department of
Biomedical Engineering
Emory University and Georgia Tech
Atlanta, GA, USA
fMRI Deriving Connectivity
• Functional connectivity
– “Temporal correlations between remote
neurophysiological events”
neurophysiological events”
• Effective connectivity
– “Influence one neuronal system exerts over
another”
Granger Causality Analysis
• Granger causality is based
on the concept of temporal
precedence information
• If including past values of
Y improves the prediction
of future values of X, then
Y is said to have a causal
Time n
T I M
Y is said to have a causal
influence on X
• Originally invented by
Granger for stock market
prediction and awarded
Nobel prize in economics
in 2003
Time series X Time series Y Time n-k
M E
Multivariate Granger Causality and Directed
Transfer Function
• Let
X
(t)=(x
1(t),x
2(t),... x
k(t))
be the data
vector wherein x
kis the time series, a
multivariate autoregressive model with
model parameters
A(
n)
of order p
is given by
• Transforming the equation to the frequency
domain, one has the transfer matrix, H(f)(=A
-1(f)), or the non-normalized directed transfer
∑
= + − = p n t n t n t 1 ) ( ) ( ) ( ) ( AAAA XXXX EEEE X XX X ) ( ) ( ) ( ) ( ) ( f A 1 f E f H f E f X = − =1
(f)), or the non-normalized directed transfer
function (DTF).
• H(f) is multiplied by partial coherence to
emphasize direct connections and summed
over all frequencies to obtain direct DTF,
where the partial coherence is given by
and M
ij(f) is the minor of the cross-spectrum
matrix between the time series.
∑
= f ij ij ij h f f dDTF ( )η ( ) ) ( ) ( ) ( ) ( 2 f M f M f M f jj ii ij ij = ηStatistical Testing using Surrogate Data
Original Data Retain Power Spectrum Randomize Phase Surrogate Data Calculate dDTF on Surrogates 2500 Times Empirical Null DistributionOriginal time series
Phase randomized
0 10 20 30 40 NULL SURROGATE DISTRIBUTION
Significant !
5s
12s
fMRI Impulse-Response Function
τ
2s
A Poor Man’s Application of
Granger Analysis: Investigation
of Slow Causal Influences
of Slow Causal Influences
Motor Fatigue Experiment
• Subjects repeated a hand contraction task guided
by visual feedback (50% maximum force). The
duration of each contraction was 3.5 s, followed
by a 6.5 s rest.
by a 6.5 s rest.
• The fatigue task lasted 20 minutes, with a total of
120 contractions performed by each subject.
• fMRI images acquired at every 2 seconds.
ROIs and Integrated Time Courses
■ ▲ * ♥ ◄ ■ SMA ▲ M1 * S1 ♥ P ◄ PM 0 100 200 300 400 500 600 -15 -10 -5 0 5 10 15 Time (TRs) 0 20 40 60 80 100 120 10 15 20 25 30 35 40 Time Deshpande et al., HBM (2009)Effective Connectivity during Fatigue
SMA M1 PM C P S1 SMA M1 PM C P S1 SMA M1 PM C P S1First Window Middle Window Last Window
Deshpande et al., HBM (2009)
• In window 1, the strong output from S1 indicates fine tuning of motor activity by sensory feedback.
• In window 2, cerebellum’s role increases indicating more motor control; the shift from window 1 to window 2 likely reflects learning.
• In window 3, the connectivity pattern is similar to that of
window 2 but there is a general reduction in connectivity due to fatigue.
Network from raw time series
SMA M1 P S1 SMA M1 P S1 SMA M1 P S1window 1 window 2 window 3
PM C PM C PM C
• Networks derived from the raw data exhibit more
causal paths that are less significant, with no
apparent driving node(s) and little change with
time.
Effect of Slow Sampling and
Hemodynamic Response on Fast
Causal Influences
BOLD Signal and LFP
Simulations
• LFP signal X
sampled at 1ms
interval. Y
obtained by shifting it
by d
ms
• Hemodynamic impulse response modeled by modeled by Gamma functions A = time to peakW = full width at half maximum K = scaling factor
• TR=0.5, 1, 1.5 and 2 seconds
fMRI simulation from LFP
Original LFP time series X(red)→Y(blue): 0.3
Y(blue)→X(red): 0.0024 LFP convolved with HRF X(red)→Y(blue): 1.8 Y(blue)→X(red): 0.7 X(red)→Y(blue): 80 Y(blue)→X(red): 21
HRF Difference (0.5 s) Opposite the Neuronal Delay
X(red)→Y(blue): 44 Y(blue)→X(red): 12 No Noise Noise (SNR=50) added to simulated fMRI X(red)→Y(blue): 29±5 Y(blue)→X(red): 15±4HRF Difference (0.5 s) Opposite the Neuronal Delay
preserving HRF shape
HRFs shifted in time; shape preserving (results below)
HRFs with different rise time; shape altering (results above)
Physiologically feasible model
Aguirre et al, NeuroImage 1998
preserving (results below) shape altering (results above)
No Noise X(red)→Y(blue): 0.85 Y(blue)→X(red): 1.73 SNR=50 X(red)→Y(blue): 0.9 ± 0.3 Y(blue)→X(red): 1.6 ± 0.2 Reverse direction inferred
Desphande, Sathian & Hu. Effect of Hemodynamic Variability on Granger Causality Analysis of fMRI. NeuroImage 52: 884-96, 2010.
•In the absence of HRF variability, even tens of milliseconds of neuronal delay can be inferred from GC analysis of fMRI.
•In the presence of HRF delays which oppose neuronal delays, the minimum detectable neuronal delay may be hundreds of milliseconds.
•In the more realistic scenario of unknown neuronal and hemodynamic delays within their normal physiological range, the accuracy of
detecting the correct multivariate network from fMRI is well above chance and up to 90% with faster sampling.
•Under all conditions, faster sampling and low measurement noise improve the sensitivity of GC analysis of fMRI data.
• Aim
: to investigate the neural
circuitry underlying tactile
spatial acuity at the human
finger pad
• Spatial task
: linear, 3-dot
arrays, applied to the
immobilized right index
finger pad using a
computer-Tactile Spatial Acuity Experiment
finger pad using a
computer-controlled, MRI-compatible,
pneumatic stimulator
• Control task
: Temporal offset
stimulus instead of spatial
offset
• Activity specific for
spatial processing
revealed activity in a
distributed
fronto-parietal cortical
network
• Levels of activity in
right posterior
right posterior
intraparietal sulcus
(pIPS) significantly
predicted individual
acuity thresholds
135
• Multivariate Granger
causality
relationships among
selected ROIs
• Top: Better
2 66• Top: Better
• Bottom: Poorer
What determines acuity ?
• Regression shows that in the better group, the paths predicting acuity converged from the left postcentral sulcus and right frontal eye field onto converged from the left postcentral sulcus and right frontal eye field onto the right pIPS.
• These connections were selective for the spatial task • Their weights predicted the level of right pIPS activity
• Conclusion: The optimal strategy for fine tactile spatial discrimination involves interaction in the pIPS of a top-down control signal, possibly attentional, with somatosensory cortical inputs, reflecting either
visualization of the spatial configurations of tactile stimuli or engagement of modality independent circuits specialized for fine spatial processing
Comparing functional connectivity and
Granger-based effective connectivity
E ff e c ti v e C o n n e c ti v it y M a tr ix
"Better" group "Poor" group
F u n c ti o n a l C o n n e c ti v it y M a tr ix R=-0.08, p=0.45 R=-0.02, p=0.85
Correlation-purged Granger Causality
• Given n time series X(t) = [x1(t) x2(t) … xn(t)], the traditional VAR model of order p is given below
X(t) = A(1)X(t-1) + A(2)X(t-2) + ... + A(p)X(t-p) + E(t)
where A(1) … A(p) are the coefficients of the model and E(t) is the model error
• In order to account for the lag correlation effects, we introduce the zero-• In order to account for the lag correlation effects, we introduce the
zero-lag term
X(t) = A' (0)X(t) + A' (1)X(t-1) + A' (2)X(t-2) + ... + A' (p)X(t-p) + E' (t)
• The inclusion of the zero-lag term affects the value of other coefficients and hence A'(1) … A' (p) ≠ A(1) … A(p)
• GC obtained from A'(1) … A' (p) are linearly independent of zero-lag correlation, which we call correlation-purged GC (CPGC)
Simulation
• CASE 1: Consider two time series x(n) and y(n) modeled as a first
order VAR process such that the causal influence between them is zero but the instantaneous correlation is nonzero
= = 1 0.5 0.5 1 Cov and 0 0 0 0 A(1)
• Assuming x(n) and y(n) represent LFPs sampled at 1ms, they were
convolved with HRF and downsampled 1000/2000 times to simulate convolved with HRF and downsampled 1000/2000 times to simulate fMRI series with TRs of 1 s and 2 s
• CASE 2: Subsequently, in order to demonstrate the efficacy of CPGC
for recovering neuronal causal influences from fMRI, we generated x(n)
and y(n) such that a unidirectional causal influence exists from x(n) to
y(n) with no correlation between them. The corresponding fMRI time
series, x' (n) and y' (n), were derived and zero-lag correlation, GC and CPGC were calculated from them
TR Zero-lag correlation Granger causality Correlation-purged Granger causality
x' (n) ↔y' (n) x' (n) →y' (n) y' (n) → x' (n) x' (n) →y' (n) y' (n) → x' (n)
1 s 0.49 ±0.05 0.47 ±0.01 0.47 ±0.01 0.01 ±0.02 0.01 ±0.02
2 s 0.49 ±0.06 0.40 ±0.07 0.40 ±0.07 0.00 ±0.09 0.00 ±0.09
TR Zero-lag correlation Granger causality Correlation-purged Granger causality
x' (n) ↔y' (n) x' (n) →y' (n) y' (n) → x' (n) x' (n) →y' (n) y' (n) → x' (n)
1 ms 0.5 ±0.09 0.00 ±0.01 0.00 ±0.01 0.00 ±0.01 0.00 ±0.01
Simulation 1: Only correlation and no causality in LFP data
TR Zero-lag correlation Granger causality Correlation-purged Granger causality
x' (n) ↔y' (n) x' (n) →y' (n) y' (n) → x' (n) x' (n) →y' (n) y' (n) → x' (n)
1 s 0.29 ±0.09 0.47 ±0.1 0.27 ±0.09 0.21 ±0.02 0.01 ±0.02
2 s 0.29 ±0.09 0.40 ±0.09 0.20 ±0.07 0.14 ±0.02 0.00 ±0.02
TR Zero-lag correlation Granger causality Correlation-purged Granger causality
x' (n) ↔y' (n) x' (n) →y' (n) y' (n) → x' (n) x' (n) →y' (n) y' (n) → x' (n)
1 ms 0.0 ±0.09 0.5 ±0.01 0.0 ±0.01 0.5 ±0.01 0.0 ±0.01
Simulation 2: Only causality and no correlation in LFP data
Functional connectivity
• Temporal correlations of low frequency
fluctuations exist in the brain, even at “rest”
– Biswal et al. Magn Reson Med 34:537 (1995)
• Connectivity of functionally related areas
– Examples: Motor, visual, language, “default mode”
networks
Resting State Networks (RSNs)
• Internally directed cognitive processing (specifically, self referential and mental simulation) by Default Mode
Network (DMN)
• Obtained using posterior cingulate (PCC) seed
• Internally directed cognitive processing (specifically,
memory encoding and retrieval ) by Hippocampal Cortical
Memory Network (HCMN)
• Obtained using hippocampus (HC) seed
• Externally directed cognitive processing by Dorsal Attention Network (DAN)
• Obtained using middle temporal (MT) seed
• Executive control of anti-correlated DMN/HCMN and DAN by Fronto-parietal Control Network (FPCN)
• Obtained usinganterior prefrontal (aPFC) seed
DMN HCMN DAN FPCN Frontal Parietal Temporal Cingulate
DMN HCMN DAN FPCN Frontal Parietal Temporal Cingulate
DMN HCMN DAN FPCN Frontal Parietal Temporal Cingulate
PCC and pIPL: The Transit Hubs
• Central location on layout ideal for this role • High resting state metabolism
• PCC seed-based correlation analysis will only reveal DMN and HCMN ROIs • Drives the characterization of different groups of ROIs in different networks • Functional segregation of different networks is rather a soft boundary
Memory Encoding • Predominant inputs to HF from:
• Parietal ROIs provide perceptual content
• DAN ROIs may provide the context, i.e. those perceptual contents which are being attended to
Deshpande et al., NeuroImage (2010)
Integration and Control
• aPFC is at the apex of the control hierarchy
• Implicated in integrating the outcomes of multiple cognitive operations • Integration of internal and external representations from the anti-correlated
DMN/HCMN and DAN systems:
• The input to R aPFC from PCC brings the internal representations from memory
• Inputs from bilateral MT in the DAN bring the information about the external environment
• SVC successfully learned patterns of functional connectivity
capable of predicting MDD from HC
capable of predicting MDD from HC
– Uncovered differences not discovered by t-test analysis
• Feature selection substantially improved the prediction
accuracy of SVC
PCE may affect behavior and functioning by increasing baseline arousal and altering the excitatory/inhibitory balancing mechanisms involved in cognitive resource allocation.
PCE is associated with activation changes in different regions, so can connectivity changes across these regions be used to predict PCE?
Medial PFC
Left DLPFC DLPFCRight
Prediction of PCE status with functional/effective connectivity
Left amygdala Right amygdala
Medial PFC Left Parietal cortex Right Parietal cortex ACC PCC
Prediction of PCE status with functional/effective connectivity
Correction of HRF latency with breath holding
Breath Holding
3.8 sec 3.8 sec 3.8 sec 3.0 sec 11.4 sec
inhale inhale inhale inhale get ready
hold release
… 16 repetitions
Face Perception
90 faces randomly presented
ISI = 3.92sec (mean) ± 2.1sec (SD)
2 fMRI scan runs, ~6min TR = 1sec or 2sec
Data Processing
Mean signal as reference
1. ROI definition by regular GLM
V1 and FFA
2. BH modulated cortical voxels
3. Voxel-wise signal latency
ROI signal extraction 4. Voxel-wise latency correction Reference Each voxel
Shift to find the best correlation
FFA original V1 original FFA corrected V1 corrected Shift Latency applied
5. Compare GCA results of uncorrected & corrected data
Effect of temporal resolution on GCA
Block (30sec) design, visual (flashing checker board) motor (finger tapping) task
Parallel imaging with TR=2.4sec, 0.6sec, and 0.3sec, 4 subjects Six ROIs defined according to fMRI activation
Left LGN Right LGN Visual cortex Left primary motor Right primary motor SMA