Connectivity Modeling Anthony Randal McIntosh Department of Psychology University of Toronto Overview • Theoretical issues and basis for analytic approach • Structural equation modeling – Integrating anatomy and function • Partial Least Squares – Identification of distributed systems • Applications – Sensory learning – Working memory Theoretical focus • All behavioural and cognitive operations in the brain come about through the the action of distributed networks • In order to assess this, need methods that can measure the whole brain and analyses that look at more than one region at a time • Ideally, we would like to analyze spatial and temporal patterns of brain function at the same time Causal patterns in brain research Design/Task Brain Behaviour Causal patterns in brain research Response Stimulus Causal patterns in brain research Response Stimulus Theory to Analysis • Examine the influences between brain areas – Interregional correlation (Horwitz, et al, 1984) – Structural equation modeling (McIntosh & Gonzalez-Lima, 1991, Buchel & Friston, 1997) – Multiple regression and extensions (e.g., Kalman filters, Buchel & Friston, 1998) – Bayes networks (Dynamic Causal Modeling, Friston, Penny, et al, 2003) • Identification of interacting regions – Partial Least Squares (McIntosh, Bookstein, et al, 1996) – Canonical Variates Analysis (Strother et al, 1995) – Independent Components Analysis - 32 flavours (McKeown et al, 1998, Calhoun et al, 2001, Beckmann, Smith, et al., 2002) Functional and Effective Connectivity Structural Equation Modeling • Multivariate multiple regression • Combines interregional covariances with anatomical framework • Provides means to assess whether effective connections are modified by task-demands or differ between groups • Is not meant to be a model test in the coventional sense – Goodness of fit not as relevant Structural Equation Modeling w = 0.611 A x = 0.011 y = 0.614 C A B C D B z = -0.553 D A 1.00 B 0.48 1.00 C 0.62 0.16 1.00 D 0.24 -.41 0.06 Structural Equations A = xB + yC + B = wA + zD + B 1.00 Dorsal vs. Ventral Cortical Visual Streams Spatial Location Object Identification 7 19d 46/47 46/47 17 18 17 18 19v 37 19v 37 21 Path Coefficients Positive Negative 0.7 to 1.0 0.4 to 0.6 0.1 to 0.3 0 McIntosh et al, J. Neurosci, 1994 21 What inferences does Structual Equation Modeling allow? Space Object 21 21 21 7 37 19d 19v 7 46 46 46 46 21 7 37 37 19d 19d 19v 19v Negative Positive 0.65 - 1.0 0.35 - 0.65 0.1 - 0.35 0 7 19d 19v 37 What inferences does Structual Equation Modeling allow? Space Object 21 21 21 7 37 19d 19v 7 46 46 46 46 21 7 37 37 19d 19d 19v 19v Negative Positive 0.65 - 1.0 0.35 - 0.65 0.1 - 0.35 0 7 19d 19v 37 Partial Least Squares • “Least-squares” decomposition of “part” of a covariance matrix • PLS is optimized to explain the relation between two or more blocks of data – what pattern in one block most strongly covaries with a pattern in another block? • Ignores the relation among items within data blocks • Statistical assessment through resampling algorithms – Permutation test and bootstrap estimation of standard error McIntosh, Bookstein, et al, Neuroimage, 1996 Task PLS Compute matrix Mdev which is n*k by m n is the number of subjects or repetitions, k is the number of scans, m is the number of voxels. Each voxel is centered relative to the grand mean. • Compute matrix X, a matrix of scan means expressed as deviations from the grand mean. Alternative: project (correlate) set of orthonormal contrasts (design matrix) on to the data matrix Matrix X now the covariance of image activity with the experimental design. Task PLS • Perform a singular value decomposition (SVD) on X to define the latent variables (LV): SVD(X) = [U,S,V] where: X = U*S*VT U is the k by m orthonormal matrix containing voxel weights (singular or eigen image). S is a diagonal matrix of k singular (“eigen”) values. (The kth singular value is zero because we use deviation values in X, i.e., we eliminate the grand mean to create X). VT is the transpose of matrix V, a k by k orthonormal matrix of scan weights. • Project singular image on to original data to obtain “brain scores” Index of how well each subject shows the effect Task PLS Matrix M Brain Images Mean Grand Matrix X Average deviation within scan, of Matrix X Singular Value Decomposition V1 Matrix X Average deviation within scan, Matrix M Brain Images (Singular Image) U1 (Singular Image) U1 S1 Behavior PLS • Compute matrix M which is n*k by m n is the number of subjects/repetitions, k is the number of scans, m is the number of voxels. • Create vector B, which is an n* k by 1 vector of performance measures for each scan. • Create matrix Y, which contains the scan-specific correlations of voxel activity (Yk) and behavior (Bk). Behavior PLS • Perform a singular value decomposition (SVD) on Y to define the latent variables (LV): SVD(X) = [U,S,V] where: X = U*S*VT U is the k by m orthonormal matrix containing voxel weights (singular or eigen image). S is a diagonal matrix of k singular (“eigen”) values. VT is the transpose of matrix V, a k by k orthonormal matrix of scan weights. Behavior PLS Cross-correlation of B and M, Matrix Y Brain Images Matrix M Behavior Measures Matrix B S1 Cross-correlation of B and M, Matrix Y V1 U1 (Singular Image) Singular Value Decomposition of Matrix M 4 3 2 0 Brain Scores -2 -0.6 4 -0.4 -0.2 0.0 0.2 0.4 0.6 2 0 -2 -4 -1.0 -0.5 0.0 0.5 1.0 1.5 8 6 U1 4 (Singular Image) 2 0 Brain Images Matrix M -2 -4 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 Behavior Measure Correlation of Brain Scores & Behavior 1 -1 Statistical Assessment • Assessment of omnibus/latent variable structure through permutation tests – Is the latent variable significantly different from “noise”? • Assessment of the precision of estimates derived from PLS through bootstrap estimation of standard errors – How reliable is the answer? • Procrustes rotation to original solution space used to correct for axis rotation and reflection during resampling Milan & Whittaker, 1995, Royal Stat Society J Why use bootstrap? 0.02 Permutation 0.015 • Estimation of standard errors is a direct assessment of the stability of your data 0.01 Est Standard Error 0.005 0 0.02 Bootstrap – A signal can be significantly different from noise (e.g., P<0.01), but not be reliable 0.015 0.01 0.005 0 0.02 ANOVA 0.015 0.01 0.005 0 -0.04 0 Singular Vector Weight 0.04 Causal patterns in brain research Response Stimulus Multiblock PLS Brain Images Matrix M Behavior Measures Matrix B Average deviation within scan, Matrix X Normalize rows to unit-length Cross-correlation of B and M, Matrix Y Normalize rows to unit-length Grand Mean Brain Images Matrix M Multiblock PLS Average deviation within scan, Matrix X S1 Cross-correlation of B and M, Matrix Y V1 U1 (Singular Image) Matrix Z U1 (Singular Image) Brain Images Matrix M Within-task correlation of brain scores and behavior Average Brain Score within-task Singular Value Decomposition of Matrix Z How do we use this? Target Tone = 1 kHz FM ~ 65 dB - 500 ms duration 1 2 Scans 3 Distractor Tone lowP P(Visual/Tone2)=0.2 Tone highP Reaction Time (msec) Distractor Unpaired trials Paired trials 560 540 520 500 480 460 440 420 4 5 Tone highP Tone lowP Tone highP Tone highP Tone highP 2 3 4 5 Scan Tone highP P(Visual/Tone1)=0.7 6 Distractor McIntosh, Cabeza & Lobaugh, J Neurophys 1998 Distractor 6 Sensory Associative Learning 1 Identify system(s) that respond to change in significance of the tone 2 Identify system(s) that relate to (effect) a change in behavior as a result of learning 3 Identify the overlap between 1 and 2 Task PLS -28 -4 +20 16 Brain scores 14 12 10 8 6 4 2 VD1 TLP THP1 THP2 Scan THP3 VD2 Behaviour PLS -28 -4 Brain Scores +20 TLP THP2 VD2 THP1 THP3 Behavior Multiblock PLS TLP 2 Brain Scores Brain Scores 3 1 0 -1 THP2 VD2 THP1 THP3 -2 -3 TLP THP1 THP2 THP3 VD2 Behavior Scores Explaining regional activation Behaviour Adjusted rCBF Adjusted rCBF Task Visual Tone Scan 1 Tone Scan 2 Tone Scan 3 Scan Tone Scan 4 Tone - scan 1 Tone - scan 2 Tone- scan 3 Visual Tone scan 4 Visual RT - Difference (unpaired - paired) Seed voxel PLS -4 +20 Adjusted rCBF TLP r = -.57 THP2 r = -0.22 Brain Scores THP1 r = -.15 THP3 r = 0.72 Structural Equation Model Positive 10 0.1 - 0.35 0.35 - 0.65 6 0.65 - 1.0 Negative 42 0.1 - 0.35 0.35 - 0.65 0.65 - 1.0 18 10 18 18 TLP 0 10 10 6 6 6 42 42 42 18 THP1 18 18 THP2 18 18 THP3 Occasions, Trials, Subjects, Groups ERP/MEG/fMRI Data Sets SPACE Voxels, detectors, electrodes Occasions,Trials, Subjects, Groups ERP/MEG/fMRI Data Flatten the matrix Occasions,Trials, Subjects, Groups Space: Voxels/Channels Time and Space Task PLS Matrix M Brain Images Mean Grand Matrix X Average deviation within scan, of Matrix X Singular Value Decomposition V1 Matrix X Average deviation within scan, Matrix M Brain Images (Singular Image) U1 (Singular Image) U1 S1 Spatiotemporal PLS- fMRI Voxel Saliences Brain Images Matrix Y Temporal Brain Scores How strongly does the brain differentiate tasks at each timepoint??? McIntosh, Protzner & Chau, Neuroimage, in press N-back Task Variant Motivation • Working memory may be conceived as the interplay of sustained attention and memory • For a given WM task - is there a dissociation that would reflect the relative contribution of an attentional component v.s. recruitment of memory retrieval? Lenartowicz & McIntosh, submitted N-Back Task Variant 2-back (standard - Std) 0-back (detection) 1-back Time 2-back (Cued) Task PLS 0 8 2 6 6 0 8 4 Brain Score Time (sec) 4 Bootstrap Ratio 4 2 0 -2 -4 -4 10 -6 -8 12 -8 14 +16 +24 +32 +40 +48 +56 +64 0 2 4 6 8 10 Time (sec) 0 -b a c k 1 -b a c k C u e d2 -b a c k S td 2 -b a c k 12 14 Task PLS 0 10 2 8 6 0 8 6 Brain Score Time (sec) 5 Bootstrap Ratio 4 4 2 0 -2- 10 -5 0 2 4 6 8 10 Time (sec) 12 -10 14 -12 0 +12 +24 +36 +48 +60 0 -b a c k 1 -b a c k C u e d2 -b a c k S td 2 -b a c k 12 14 Task PLS 0.3 0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 0 2 4 6 8 10 12 14 0 -b a c k 1 -b a c k C u e d2 -b a c k S td 2 -b a c k Task PLS 0.25 0.2 0.15 0.1 0.05 0 -0.05 -0.1 -0.15 0 2 4 6 8 10 12 14 0 -b a c k 1 -b a c k C u e d2 -b a c k S td 2 -b a c k Task PLS • Dominant effect on anterior cingulate activity • What is the functional connectivity of the anterior cingulate? – Does it change between the 2-back tasks? • Is the pattern of functional connectivity related to behavior? Behaviour/Seed PLS 15 10 1 5 0 Correlation 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 Std 2-back Cued 2-back -5 -10 -16 -4 +8 +20 +32 +44 +56 +68 AC RT Hits Behaviour/Seed PLS 10 1 5 0 Correlation 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -5 -10 -16 -4 +8 +20 +32 +44 +56 +68 -1 Std 2-back Cued 2-back AC RT Hits Summary & Implications • Anterior cingulate activity differentiates tasks based on attentional demands • Functional connectivity varies with attentional demand • Relation of functional connectivity patterns to performance also varies with attentional demand • Behavioural relevance of a region to a cognitive operation depends on its pattern of functional connectivity – Neural Context - McIntosh, Neural Networks 2001 Evaluating the analytic tools: How do we know when the math is right? • Neurobiological interpretation – Identification of new principles • Psychological interpretation – What is the level of nervous system operation that best relates to the cognitive operation? • What is the question? – Level of the answer – Does the experimental design require a certain analytic tool? – Is causality necessary or will correlation suffice? Thank You Acknowledgements: Collaborators: NJ Lobaugh, MN Rajah, CL Grady Funding: CIHR, NSERC, JS McDonnell Fnd http://www.rotman-baycrest.on.ca What inferences does Structual Equation Modeling allow?