Partial Least Squares Analysis of Functional Neuroimaging Data Cheryl L. Grady Senior Scientist Rotman Research Institute at Baycrest Toronto, Ontario
Statistical Map superimposed on anatomical MRI image Time Condition 1 Condition 2... ~ 5 min fMRI Signal (% change) Condition Typical fMRI Analysis- “Pure Insertion” & Subtraction Region of interest (ROI) Source: Jody Culham’s fMRI for Dummies web sitefMRI for Dummies
However, Brain Areas are Connected! Although brain regions may activate selectively for different types of stimulus features or processing, this does not mean that activity in these individual areas is solely responsible for cognition Brain areas are heavily interconnected anatomically, so it follows that they also are interconnected functionally Our analytic approaches should be sensitive to covarying brain activity and functional connections between brain regions Multivariate analysis - when brain region “A” is active during a task, what is the rest of the brain doing?
Why use PLS? Computationally efficient – simultaneously dentifies similarities among conditions/groups as well as differences Is quite flexible and can handle “multi-table” questions easily Can be data-driven or hypothesis-driven (task contrasts can be pre-specified) Because the whole brain is assessed in a single computational step, there is no need to correct for multiple comparisons Can be applied to either functional or structural MRI data, or to ERPs For more details see Krishnan et al, Neuroimage, 2011, 56, For a step-by-step manual go to Support + Services Labs PLS Software Support + ServicesLabsPLS Software
Task-related changes Brain/Behaviour Correlations Functional Connectivity Brain TaskBrain Behaviour PLS Options for fMRI Analysis Task PLS Behaviour PLS Seed PLS
PLS Task Analysis Experimental Design Contrasts Brain Images C o n t r a s t C o n t r a s t Covariance between Design and Brain SVD Set of Latent Variables Extracted in order of amount of covariance explained
Composition of the LV Singular Image: each voxel has a weight (salience) that indicates how it relates to task contrast (positively or negatively) Singular Value (S): amount of covariance explained by each LV (1)(2)(3) Contrast across experimental conditions Sum of all voxel saliences
How do we determine the significance of our result? Permutation test to determine probability of each LV Use resampling to determine robustness of each voxel’s contribution to the spatial pattern (bootstrap)
Permutation Test X 1 is task 1 X 2 is task 2 X 3 is task 3 Subject (i 1 =1 of 3) Voxels/VOIs (1,…,j,…J=12) Run some large number of analyses with permuted data and calculate the number of times that the obtained S value for each LV is larger than the original – that is the probability Shuffle the condition labels
How Robust are the voxel contributions? We need to estimate the variance of the weight for each voxel We do this by creating a set of “new” samples by resampling the original set of participants o Sample repeatedly from finite sample o Make sure that probabilities do not change by sampling with replacement
Sampling with Replacement A B C D EF
A Get the first observation
A A Put it in the box
A A Put the original back in the hat
A B Take out another
A B B Put the second one in the box
A B B B Put it back in the hat
A BB EEF And so on, until you get the full Sample Some observations are repeated Some are not repeated
Bootstrap Use a large number of resampled datasets to calculate an estimate of the variance of the salience for each voxel Divide the salience of each voxel by it’s SD This is the bootstrap ratio (BSR), which is analogous to a Z score and can be thresholded to get a robust spatial pattern
Task PLS
Two of the Major Brain Networks Fox et al, PNAS, 2005 Default Network Task Network
19 younger (20-30 yrs, m = 25 yrs); 28 older (56-84 yrs, m = 66 yrs) Healthy, community-dwelling, cognitively normal fMRI at 3T, block design 4 tasks, with stimulus parameters set for 80% accuracy Experiment details
Tasks Detection Perceptual Matching Attentional Cueing Delayed Match to Sample Alternated with Fixation Baseline Grady et al., Cerebral Cortex, 2010
a. LV1b. LV2 Brain Score FixDETPMTATTDMS Young Old FixDETPMTATTDMS
DMN, Young only DMN, Both TPN, Young only TPN, Old only TPN, Both DMN, Old only The Default Network “Shrinks” and the Task Network “Expands” with Age
Functional Connectivity with Seed PLS
Functional Connectivity - Seed PLS Extracted Seed Values Brain Images Within-task Correlation of Seed Activity and Brain Images Singular Image Correlations of Total Brain Activity SVD and Seed Activity Across 3 Tasks
Age Reduction in Correlation between Two DN Hubs What about the rest of the network? Andrews-Hanna et al, Neuron 2007
Default Network Young Both Old Grady et al., Cerebral Cortex, 2010 Default Seed
Reduced Functional Connectivity of the DN in Older Adults Correlation YoungOld Grady et al., Cerebral Cortex, 2010 FC during fixation baseline – correlation of vmPFC activity with whole brain pattern p < Older Adults Young Adults
Task Network Young Both Old Task Seed
Maintained FC of the TN in older adults Correlation Older Adults Young Adults Grady et al., Cerebral Cortex, 2010 Correlation of right IPL activity with whole brain pattern p < 0.002
Summary Both young and old adults show DN activity during baseline and TN activity during the tasks in expected areas Young have more extensive DN activity during fixation and old have more extensive TN activity during all tasks Functional connectivity in the DN is more vulnerable to age than in the TN
Behaviour PLS
Relating Behaviour to Brain Activity Behavioural Measure (s) Brain Images Within-task Correlation of Behaviour and Brain Images Singular Image Correlations of Total Brain Activity SVD and Behaviour Across 3 Tasks
Behaviour Analysis – Across 3 Tasks PMT, ATT, DMS Brain Factor ISD S1 S2 S3 Mean RT Age
Correlations with Behaviour p <.001 Garrett et al, J Neurosci, 2011
But.... Brain’s natural state is a variable one
Younger Older Within-subject RT variability Reflects neural inefficiency Random lapses in attention or executive control Behavioural variability increases with age Macdonald et al., (2006;2009); Dixon et al. (2007)
Correlations with SD Measures p <.001 Garrett et al, J Neurosci, 2011
Correlation Strength Mean Brain ActivitySD of Brain Activity
Comparison of SD and mean Spatial Patterns Regions with both effects SD+ Beh +Mean + Beh -
Summary PLS is a useful tool for examining the many aspects of functional neuroimaging data within the same general framework Using PLS we have shown that there are age differences in DN and TN, with the former showing reduced functional connectivity and the latter showing maintained connectivity BOLD variability measures show promise for understanding aging of the brain and cognition