An Overview, Challenges, and Future Directions - SAMSI
An Overview, Challenges, and Future Directions - SAMSI
An Overview, Challenges, and Future Directions - SAMSI
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling of Neuroimaging Data: <strong>An</strong><br />
<strong>Overview</strong>, <strong>Challenges</strong>, <strong>and</strong> <strong>Future</strong> <strong>Directions</strong><br />
F. DuBois Bowman<br />
Department of Biostatistics <strong>and</strong> Bioinformatics<br />
Center for Biomedical Imaging Statistics<br />
Emory University, Atlanta, GA, 30322<br />
Neuroimaging Data <strong>An</strong>alysis<br />
<strong>SAMSI</strong><br />
Research Triangle Park, NC<br />
June 6, 2013<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 1 / 77
Experiments Models Connectivity Prediction Summary<br />
Outline<br />
1 fMRI Experiments<br />
2 Modeling of fMRI Data: Neuroactivation Studies<br />
3 Connectivity<br />
4 Prediction<br />
5 Summary<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 2 / 77
Experiments Models Connectivity Prediction Summary<br />
Hemodynamic Response Function<br />
HRF<br />
Typical BOLD response to an impulse stimulation<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 3 / 77
Experiments Models Connectivity Prediction Summary<br />
fMRI Study Designs<br />
Block Designs<br />
Dominant in the early years of fMRI studies<br />
Multiple stimuli of the same condition are grouped together in blocks<br />
[f ∗ g] (t) = ∫ ∞<br />
−∞<br />
f (τ)g(t − τ)dτ<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 4 / 77
Experiments Models Connectivity Prediction Summary<br />
fMRI Study Designs<br />
Block Designs<br />
Advantages:<br />
Relatively simplified data analysis<br />
Increased statistical power<br />
More robust results<br />
Often yields relatively large BOLD signal changes.<br />
Disadvantage:<br />
Many potential confounds<br />
Habituation<br />
<strong>An</strong>ticipation<br />
Other strategy effects.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 5 / 77
Experiments Models Connectivity Prediction Summary<br />
fMRI Study Designs<br />
Event-related Designs<br />
Allow different stimuli to be presented in arbitrary sequences.<br />
Permits r<strong>and</strong>omization of conditions.<br />
Two types of event-related designs:<br />
Spaced event-related designs<br />
Rapid event-related designs<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 6 / 77
Experiments Models Connectivity Prediction Summary<br />
fMRI Study Designs<br />
Event-related Designs<br />
Advantages:<br />
Detect transient variations in hemodynamic responses.<br />
Allow the temporal characterization of BOLD signal changes: the HRF.<br />
Reduce potential confounds.<br />
Disadvantages:<br />
Reduced statistical power to detect BOLD differences between different<br />
conditions.<br />
Lower SNR.<br />
More complex design <strong>and</strong> analysis. (Dale, 1999)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 7 / 77
Experiments Models Connectivity Prediction Summary<br />
fMRI Study Designs<br />
Resting-state fMRI Studies<br />
Rapid growth in resting-state studies<br />
Acquire scans while subjects are left to think for themselves.<br />
May reflect a natural <strong>and</strong> more common mode of neural processing.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 8 / 77
Experiments Models Connectivity Prediction Summary<br />
fMRI Study Designs<br />
Resting-state fMRI Studies<br />
Common analytic objectives:<br />
Resting-state connectivity<br />
Correlations (associations) in neural activity between distinct brain<br />
locations.<br />
Default mode network<br />
Raises numerous analytical challenges<br />
Lower SNR.<br />
More complex design <strong>and</strong> analysis.<br />
Mode of neural processing is inconsistent across subjects.<br />
Unable to make relative comparisons reflecting changes in localized<br />
neural processing.<br />
I.e., unable to determine average brain activity changes at each voxel.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 9 / 77
Experiments Models Connectivity Prediction Summary<br />
Data Example<br />
Working Memory in Schizophrenia Patients<br />
N=28 subjects: 15 schizophrenia patients <strong>and</strong> 13 healthy controls<br />
fMRI Tasks: Serial Item Recognition Paradigm (SIRP)<br />
Encoding set: Subjects asked to memorize 1, 3, or 5 target digits.<br />
Probing set: Subjects sequentially shown single digit probes <strong>and</strong> asked<br />
to press a button:<br />
with their index finger, if the probe matched<br />
with their middle finger, if not.<br />
6 runs per subject: (177 scans per run for each subject)<br />
3 runs of working memory tasks on each of 2 days<br />
Objective: Compare working memory-related brain activity between<br />
patients <strong>and</strong> controls<br />
Data from the Biomedical Informatics Research Network (BIRN) [1]: Potkin et al. (2002), Proc. 41st <strong>An</strong>nu. Meeting Am.<br />
College Neuropsychopharm.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 10 / 77
Experiments Models Connectivity Prediction Summary<br />
Notable fMRI Data Characteristics<br />
Massive Data Sets<br />
Roughly 300,000 brain voxels for each scan at time of analysis<br />
S = 177 scans per run<br />
6 runs (sessions) per subject<br />
T=1062 scans total per subject<br />
Over 300 million spatio-temporal data points per subject! Over 5<br />
billion for all subjects!!<br />
Temporal correlations<br />
Scan to scan correlations<br />
Between session correlations<br />
Complex spatial correlations<br />
Correlations between neighboring voxels<br />
Long-range correlations between different brain regions<br />
Roughly 45 billion voxel pairs<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 11 / 77
Experiments Models Connectivity Prediction Summary<br />
Common <strong>An</strong>alysis Framework<br />
Two-stage Model<br />
First, fit a linear model separately for each subject (at each voxel)<br />
Temporal correlations between scans: AR models (+ white noise)<br />
Pre-coloring/temporal smoothing [Worsley <strong>and</strong> Friston, 1995]<br />
Pre-whitening [Bullmore et al, 1996; Purdon <strong>and</strong> Weisskoff, 1998]<br />
Linear covariance structure<br />
Alternative structures available for PET [Bowman <strong>and</strong> Kilts, 2003]<br />
Second, fit linear model that combines subject-specific estimates<br />
A two-stage (r<strong>and</strong>om effects) model<br />
Simplifies computations*<br />
Sacrifices efficiency<br />
For Inference: Compute t-statistics at each voxel <strong>and</strong> threshold<br />
Consider a multiple testing adjustment (Bonferonni-type, FDR, RFT)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 12 / 77
Experiments Models Connectivity Prediction Summary<br />
Common <strong>An</strong>alysis Framework<br />
Properties<br />
Two-stage (r<strong>and</strong>om effects) model<br />
Simplifies computations<br />
Sacrifices efficiency<br />
Assumes independence between different brain locations<br />
Targets activation analyses, disregarding functional connectivity<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 13 / 77
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling<br />
General Linear Model: Stage I<br />
Y ig (v) = X igv β ig (v) + H iv γ ig (v) + ε ig (v)<br />
Y ig (v) S × 1 serial brain (BOLD) activity at voxel v.<br />
X igv S × q design matrix containing independent variables.<br />
β ig (v) q × 1 parameter vector linking experimental tasks.<br />
ε ig (v) S × 1 r<strong>and</strong>om error about ith subject’s mean.<br />
H iv S × m contains other covariates, e.g. high-pass filtering.<br />
ε ig (v) ∼ Normal(0, τgv 2 V).<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 14 / 77
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling<br />
General Linear Model: Stage II<br />
β ig (v) = µ g (v) + d ig (v)<br />
β ig (v) regression coefficients from stage I.<br />
µ g (v) group-level mean for all effects.<br />
d ig (v)<br />
r<strong>and</strong>om error.<br />
d ig (v) ∼ Normal(0, ω 2 (v)R).<br />
*Examine contrasts, e.g. A = [PostTx − PreTx] patients − [FU − BL] controls<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 15 / 77
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling<br />
General Linear Model: Test Statistic<br />
t(v) =<br />
Cˆµ g (v)<br />
ŝe(Cˆµ g (v))<br />
C is a user-specified contrast matrix (or vector).<br />
A threshold is applied to the test statistic t(v) for every voxel,<br />
v = 1, . . . , V .<br />
Adjustments are made for multiplicity<br />
Bonferroni-type corrections: divide significance level by the number of<br />
tests performed.<br />
R<strong>and</strong>om field theory: account for spatial dependence, as captured by<br />
the maximum distribution.<br />
False discovery rate: expected proportion of erroneous rejections,<br />
among all rejections.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 16 / 77
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling<br />
Mixed Model Equivalent<br />
Y ig (v) = X igv β ig (v) + H iv γ ig (v) + ε ig (v)<br />
β ig (v) = µ g (v) + d ig (v)<br />
=⇒ Y ig (v) = X igv µ g (v) + X igv d ig (v) + H iv γ ig (v) + ε ig (v)<br />
X igv µ g (v)<br />
X igv d ig (v)<br />
fixed effects (group-level).<br />
r<strong>and</strong>om effects (subject-specific effects).<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 17 / 77
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling<br />
Assumptions<br />
BOLD signals follow a Gaussian<br />
distribution.<br />
The same effects are present for each<br />
voxel.<br />
Relationship between independent<br />
variables <strong>and</strong> the BOLD response are<br />
captured by the general linear model.<br />
Estimation at one voxel is unrelated to<br />
estimation at other voxels, i.e. no<br />
spatial correlations.<br />
Stage II model does not always<br />
address correlations between multiple<br />
effects estimated for each individual.<br />
Quadratic Linear<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 18 / 77
Experiments Models Connectivity Prediction Summary<br />
Statistical Modeling<br />
Other Considerations<br />
How well does my model fit? For which voxel?<br />
Do I have sufficient power? For which voxel?<br />
Multiplicity: Are you really conducting hundreds of thous<strong>and</strong>s of<br />
tests?<br />
Largely addressed by r<strong>and</strong>om field theory.<br />
How do I best portray the uncertainty in my images of estimates?<br />
Confidence intervals for images.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 19 / 77
Experiments Models Connectivity Prediction Summary<br />
Spatial Correlations<br />
Distances<br />
Functional<br />
Physical (Geometric)<br />
<strong>An</strong>atomical<br />
(a) Functional<br />
(b) Geometric<br />
The complex<br />
neuroanatomy <strong>and</strong><br />
neurophysiology make<br />
basic assumptions of<br />
many spatial methods<br />
questionable for<br />
neuroimaging<br />
Figure: Alternative measures of<br />
distance.<br />
Bowman (2007), JASA.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 20 / 77
Experiments Models Connectivity Prediction Summary<br />
Spatial Modeling Framework<br />
Bayesian Spatial Model for Activation <strong>and</strong> Connectivity (BSMac):<br />
Stage I: Individual level - addresses temporal correlations.<br />
Stage II: Group level<br />
Define brain regions using<br />
neuroanatomic parcellation (e.g.<br />
Brodmann or AAL)<br />
Spatial correlations<br />
Within regions<br />
Between regions<br />
Inferences<br />
Voxel-level<br />
Regional<br />
Bowman et al. (2008), NeuroImage<br />
Zhang et al. (2011), Journal of Neuroscience Methods<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 21 / 77
Experiments Models Connectivity Prediction Summary<br />
BSMac<br />
Stage II: Bayesian Spatial Model for Activation <strong>and</strong> Connectivity (BSMac):<br />
Y igj | µ gj , α igj , σ 2 gj ∼ Normal(µ gj + 1α igj , σ 2 gjI)<br />
µ gj | λ 2 gj ∼ Normal(1µ 0gj , λ 2 gjI)<br />
σ −2<br />
gj<br />
∼ Gamma(a 0 , b 0 )<br />
α ij | Γ j ∼ Normal(0, Γ j )<br />
λ −2<br />
gj<br />
∼ Gamma(c 0 , d 0 )<br />
Γ −1<br />
j<br />
∼ Wishart { (h 0 H 0j ) −1 }<br />
, h 0<br />
Y igj = (Y igj1 , . . . , Y igjVg ) ′ , µ gj = (µ gj1 , . . . , µ gjVg ) ′ , <strong>and</strong><br />
α ij = (α i1j , . . . , α iGj ) ′<br />
Bowman et al. (2008), NeuroImage<br />
Zhang et al. (2011), Journal of Neuroscience Methods<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 22 / 77
Experiments Models Connectivity Prediction Summary<br />
BSMac MATLAB Toolbox<br />
GUI Interface: Available at www.sph.emory.edu/bios/CBIS/<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 23 / 77
Experiments Models Connectivity Prediction Summary<br />
BSMac MATLAB Toolbox<br />
Interactive Activation Maps<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 24 / 77
Experiments Models Connectivity Prediction Summary<br />
BSMac MATLAB Toolbox<br />
Task-Related Connectivity Maps: Schizophrenia Patients, WM Load 5<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 25 / 77
Experiments Models Connectivity Prediction Summary<br />
BSMac MATLAB Toolbox<br />
Task-Related Connectivity Maps: Healthy Controls, WM Load 5<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 26 / 77
Experiments Models Connectivity Prediction Summary<br />
BSMac Properties<br />
BSMac framework<br />
Considers activation objectives <strong>and</strong> task-related FC<br />
Models spatial correlations in brain activity<br />
Within <strong>and</strong> between defined neuroanatomic regions<br />
Yields easily interpretable posterior probabilities of activation<br />
Permits voxel-level <strong>and</strong> region-level inferences<br />
Can apply FDR-like concepts<br />
Limitations<br />
Does not account for temporal dependence between multiple sessions<br />
Fairly simple intra-regional correlation model<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 27 / 77
Experiments Models Connectivity Prediction Summary<br />
Extensions: Spatial Modeling<br />
Between session correlations: (a) captures correlations between<br />
multiple scanning sessions (e.g. days or treatment periods), (b) does<br />
NOT model between-region spatial correlations<br />
[Derado et al. (2010), Biometrics]<br />
Model correlations between scanning sessions, between-regions, <strong>and</strong><br />
locally within regions*<br />
Use imaging data to predict future neural responses<br />
[Derado et al. (2012), Statistical Methods in Medical Research]<br />
Model spatial correlations: Between AAL regions (unstructured<br />
covariance matrix), borrow strength between subregions within each<br />
AAL region (CAR model), between voxels within each subregion<br />
(exchangeable structure)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 28 / 77
Experiments Models Connectivity Prediction Summary<br />
Other Extensions<br />
Unspecified onsets of stimuli: meditation or pain tolerance.<br />
[Robinson et al. (2010), NeuroImage]<br />
Spectral modeling: Fourier coefficients are approximately uncorrelated<br />
across frequencies; wavelets have decorrelating properties.<br />
[Kang et al. (2012), JASA; Ombao et al. (2008), Statistica Sinica]<br />
Rician regression model to characterize noise contributions in fMRI<br />
along with associated estimation <strong>and</strong> diagnostic procedures.<br />
[Zhu et al. (2009), JASA]<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 29 / 77
Experiments Models Connectivity Prediction Summary<br />
Brain Connectivity<br />
Determine functional <strong>and</strong> anatomical<br />
organization of the brain<br />
<strong>An</strong>atomical<br />
Structural links between regions (axonal<br />
fiber tracts)<br />
Millions of axonal connections<br />
Functional<br />
Functional Connectivity: Correlations in<br />
brain activity between regions<br />
Task-induced<br />
Resting state<br />
Effective Connectivity: Influence that one<br />
region exerts on another<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 30 / 77
Experiments Models Connectivity Prediction Summary<br />
Structural Connectivity<br />
Structural linkages through white-matter.<br />
Humans have estimated 86 billion neurons.<br />
Axons are neuron fibers that form bundles<br />
<strong>and</strong> serve as lines of transmission in the CNS<br />
Millions of fiber bundles in the white-matter<br />
These bundles directly link some brain<br />
structures<br />
Association bundles joining cortical areas<br />
within the same hemisphere<br />
Commissural bundles linking cortical areas<br />
in separate hemispheres<br />
Projection fibers uniting areas in the<br />
cerebral cortex to various subcortical<br />
structures<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 31 / 77
Experiments Models Connectivity Prediction Summary<br />
Structural Connectivity<br />
Diffusion Tensor Imaging<br />
Identifies major white matter (axonal)<br />
fiber tracts in the brain<br />
Reflects white matter integrity by<br />
measuring restricted water diffusion in<br />
tissue<br />
Diffuse more quickly in the direction of<br />
fibers than in a perpendicular direction<br />
White matter fiber bundles (axons) are<br />
tracked from tensor field information<br />
combined across voxels.<br />
Reconstruct 3-D fibers (colors indicate<br />
direction): Red: Left-Right, Green:<br />
<strong>An</strong>t.-Post., Blue: Inferior-Superior.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 32 / 77
Experiments Models Connectivity Prediction Summary<br />
Structural Connectivity<br />
Diffusion Tensor Imaging<br />
The tensor<br />
Represented by matrix D<br />
D xx , D yy , D zz reflect molecular<br />
mobility along x, y, z axes.<br />
D xy , D xz , D yz reflect correlations<br />
between molecular displacements in<br />
perpendicular directions.<br />
Eigenvalues λ 1 , λ 2 , λ 3 indicate<br />
main diffusion direction<br />
⎡<br />
D = ⎣<br />
D xx D xy D xz<br />
⎤<br />
D xy D yy D yz<br />
⎦<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 33 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Functional Specialization (Localization):<br />
Specific mental processes link to particular<br />
brain regions [Gall, 1796; Broca, 1861;<br />
Brodmann, 1909]<br />
Functional Integration:<br />
Interactions between specialised neuronal<br />
populations<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 34 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Functional Specialization (Localization):<br />
Specific mental processes link to particular<br />
brain regions [Gall, 1796; Broca, 1861;<br />
Brodmann, 1909]<br />
Functional Integration:<br />
Interactions between specialised neuronal<br />
populations<br />
=⇒ Functional connectivity<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 34 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Seed Voxel (or Region) Approaches<br />
Calculate the correlation (or partial<br />
correlation) between a selected voxel<br />
(region) <strong>and</strong> all other voxels (regions).<br />
ρ ij = min<br />
u∈U ρ ij(u)<br />
∑ T −u<br />
= min { t=1 [y i(t + u) − ȳ i ][y j (t) − ȳ j ]<br />
}<br />
u∈U ˆσ i ˆσ j<br />
Should be strongly hypothesis driven<br />
May miss important associations due to<br />
limited number of voxels (regions) assessed<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 35 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Whole-Brain Region to Region Approaches<br />
Calculate the correlation (or partial<br />
correlation) between a all pairs of regions).<br />
∑ T −u<br />
ρ ij = min { t=1 [y i(t + u) − ȳ i ][y j (t) − ȳ j ]<br />
}<br />
u∈U ˆσ i ˆσ j<br />
Requires parcellation of the brain into<br />
regions OR selection of a subset of regions.<br />
May miss important associations due to this<br />
selection process<br />
May select finer or coarser map of regions<br />
depending on objectives<br />
Finer map may prompt the assessment of<br />
whole brain topological properties (graph<br />
methods)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 36 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Partitioning Approaches: Independent Components <strong>An</strong>alysis (ICA)<br />
Y = × . +<br />
Goal: Decompose observed fMRI data into a linear combination of<br />
spatio-temporal source signals.<br />
Y(T × V ): observed fMRI measurements<br />
A(T × q): mixing matrix (columns contain latent time series for ICs)<br />
S(q × V ): statistically independent spatial signals (in rows)<br />
E(T × V ): noise or variability not explained by the ICs, e v ∼MVN<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 37 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Partitioning Approaches: ICA<br />
Goal: Decompose observed fMRI data into a linear combination of<br />
spatio-temporal source signals.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 38 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Partitioning Approaches: ICA<br />
A Zen meditation study<br />
Zen meditation involves disciplined regulation of attention <strong>and</strong> bodily<br />
posture<br />
24 subjects: Zen meditation group (n=12) <strong>and</strong> control group (n=12).<br />
fMRI scans acquired at 520 time points.<br />
There was a meditative condition interspersed with a lexical decision<br />
task<br />
Mediative condition: Subjects were instructed to focus on breathing<br />
<strong>and</strong> to refocus attention to breathing when distracted by a lexical task.<br />
Lexical task: Subjects were r<strong>and</strong>omly shown word or nonword letter<br />
strings <strong>and</strong> instructed to press a button indicating whether a word<br />
appeared, ”yes” (index finger) or ”no“” (middle finger).<br />
Objective: whether training in Zen meditation modifies the neural<br />
activity underlying attentional control <strong>and</strong> behavior switching.<br />
Pagnoni et al., 2008, PLoS ONE<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 39 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Partitioning Approaches: ICA<br />
A Zen meditation study<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 40 / 77
Experiments Models Connectivity Prediction Summary<br />
Functional Connectivity<br />
Partitioning Approaches: Cluster <strong>An</strong>alysis<br />
Goal is to partition the brain into<br />
clusters of regions, which exhibit<br />
similar patterns of brain function<br />
within.<br />
Each region begins as a single cluster<br />
Calculate functional distances between<br />
all brain regions, e.g. d ij = 1 − ρ ij<br />
Iteratively join most similar regions<br />
(clusters) until a single cluster remains<br />
Examine each stage of clustering tree<br />
to determine the optimal number of<br />
clusters, e.g. assess within <strong>and</strong><br />
between cluster variability<br />
Motor<br />
Visual:<br />
Auditory<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 41 / 77
Experiments Models Connectivity Prediction Summary<br />
Multimodal Connectivity<br />
<strong>An</strong>atomically-weighted Functional Connectivity (awFC)<br />
[Bowman et al., 2012, NeuroImage]<br />
Functional connectivity relies on signaling along structural white<br />
matter pathways.<br />
Our main objective is to reliably detect networks of functionally<br />
connected regions<br />
We leverage supplementary information from the SC strength between<br />
brain regions<br />
SC does not imply FC<br />
Absence of SC does not preclude FC<br />
Indirect SC (e.g. via a remote third brain region)<br />
Potential Advantages:<br />
Help determine FC in the presence of fMRI noise<br />
Defaults to st<strong>and</strong>ard clustering techniques when no structural<br />
weighting is considered<br />
Provides additional information about connectivity relationships<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 42 / 77
Experiments Models Connectivity Prediction Summary<br />
Subregion Selections <strong>and</strong> Summaries<br />
Whole-brain analysis with subregions<br />
for each AAL region<br />
For fMRI,<br />
Resting-state subregions selected in<br />
GM to be representative of larger<br />
regions<br />
Task-based subregions selected<br />
according to other criterion, e.g.<br />
centered on maximally activated<br />
voxel in region (across all tasks)<br />
Generate summary temporal signal<br />
for each subregion<br />
For DTI,<br />
Subregions centered in WM proximal<br />
to GM regions<br />
Probabilistic tractography used to<br />
summarize region-to-region SC<br />
fMRI<br />
DTI<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 43 / 77
Experiments Models Connectivity Prediction Summary<br />
Determining Functional Networks<br />
Functional Clustering with DTI-based <strong>An</strong>atomical Weighting:<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 44 / 77
Experiments Models Connectivity Prediction Summary<br />
<strong>An</strong>atomically Weighted Functional Distance<br />
(<br />
d ij =<br />
1 − π ij<br />
λ<br />
)<br />
f ij<br />
f ij is a measure of functional dissimilarity between regions i <strong>and</strong> j<br />
E.g. 1 − ρ ij<br />
We allow lag-u associations, u ∈ [−2, 2]<br />
π ij is the probability of structural connectivity<br />
π ij ∈ [0, 1)<br />
We consider second-order connections<br />
Adjust for physical distances between regions<br />
λ is a parameter used to potentially attenuate the impact of<br />
anatomical weighting<br />
λ determined empirically<br />
λ ∈ [1, ∞)<br />
d ij → f ij , as λ → ∞<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 45 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC<br />
<strong>An</strong>atomically weighted Functional Connectivity (awFC):<br />
Perform subject-specific clustering using average linkage method<br />
applied to distances d ij<br />
Empirically optimize λ using objective function based solely on the<br />
functional data<br />
h(λ, G) = log<br />
(<br />
FCwi<br />
FC tot<br />
)<br />
Generate group cluster solution from subject-specific networks<br />
Conduct hypothesis testing to determine the statistical significance of<br />
the correlation between each pair of regions within the identified<br />
networks<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 46 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC<br />
(a)<br />
(b)<br />
6<br />
0.16<br />
0.14<br />
G=9<br />
0.12<br />
5.5<br />
λ=7.5<br />
0.1<br />
Σ i<br />
∆ i<br />
(λ)<br />
5<br />
∆(G)<br />
0.08<br />
0.06<br />
0.04<br />
0.02<br />
4.5<br />
10 20 30 40 50 60 70 80 90 100<br />
λ<br />
0<br />
0 5 10 15 20 25 30 35<br />
G (lag 1)<br />
(a) Empirical optimization of the attenuation parameter λ<br />
(b) Determination of the number of cluster, G, for subject 1<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 47 / 77
Experiments Models Connectivity Prediction Summary<br />
Data Example<br />
Auditory Task Study<br />
N=17 healthy female subjects<br />
3 fMRI runs (483 scans total for each subject)<br />
fMRI Auditory Task<br />
Cue: Hear an auditory tone (2000 Hz) in one ear.<br />
Target: Subsequently hear a target tone (1000 Hz) in either ear.<br />
Instructed to click a button with right (a) index finger if the target is<br />
from the left earphone OR (b) middle finger if from right earphone.<br />
50% of the trials have the cue <strong>and</strong> target tones in the same ear.<br />
Subjects fixated at a cross throughout the entire scanning session.<br />
DTI data also collected<br />
Objective: Study had a broad set of objectives (Mayer et al., 2009).<br />
We consider an analysis that identifies different functional networks.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 48 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC: <strong>An</strong>alysis of Task Data<br />
Motor Motor Visual<br />
Visual<br />
Auditory<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 49 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC Results: Auditory Data<br />
(a)<br />
(b)<br />
1<br />
0.9<br />
Network Correlations median<br />
In−network<br />
Cross−network<br />
1<br />
0.9<br />
Network SC Proportion<br />
In−network<br />
Cross−network<br />
0.8<br />
0.8<br />
0.7<br />
0.7<br />
Median Correlation<br />
0.6<br />
0.5<br />
0.4<br />
Proportion Connected<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.3<br />
0.2<br />
0.2<br />
0.1<br />
0.1<br />
0<br />
0 5 10 15 20<br />
subject<br />
0<br />
0 5 10 15 20<br />
subject<br />
(a) Median correlations within-networks <strong>and</strong> between-networks<br />
(b) Proportion of region pairs exhibiting SC, i.e. Pr(SC) > 0.5,<br />
within-networks <strong>and</strong> between-networks<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 50 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC Results: Auditory Data<br />
Regional Time-series from 90 Brain Regions (Subject 2)<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
0 10 20 30 40 50 60 70 80 90<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 51 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC Results: Auditory Data<br />
Time-series for awFC Networks (Subject 2)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 52 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC: <strong>An</strong>alysis of Resting-state Data<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 53 / 77
Experiments Models Connectivity Prediction Summary<br />
awFC Results: Resting-state Data<br />
(a)<br />
(b)<br />
1<br />
0.9<br />
Network Correlations<br />
In−network<br />
Cross−network<br />
1<br />
0.9<br />
Network SC<br />
In−network<br />
Cross−network<br />
0.8<br />
0.8<br />
0.7<br />
0.7<br />
Median Correlation<br />
0.6<br />
0.5<br />
0.4<br />
Proportion Connected<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.3<br />
0.2<br />
0.2<br />
0.1<br />
0.1<br />
0<br />
1 2 3 4 5 6<br />
subject<br />
0<br />
1 2 3 4 5 6<br />
subject<br />
(a) Median correlations within-networks <strong>and</strong> between-networks<br />
(b) Proportion of region pairs exhibiting SC, i.e. Pr(SC) > 0.5,<br />
within-networks <strong>and</strong> between-networks<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 54 / 77
Experiments Models Connectivity Prediction Summary<br />
Comparative <strong>An</strong>alysis<br />
Simulation Results: awFC versus St<strong>and</strong>ard (Unweighted) Clustering<br />
Table: Percent accuracies <strong>and</strong> st<strong>and</strong>ard errors (in parentheses). G=13 clusters<br />
used for all results, <strong>and</strong> σ 2 set using the experimental data.<br />
Noise-level Accuracy awFC Unweighted<br />
σ 2 Overall 97.35 (1.02) 94.61 (1.59)<br />
In-network 89.74 (4.61) 76.82 (7.78)<br />
Cross-network 98.70 (0.58) 97.78 (0.72)<br />
1.25σ 2 Overall 96.72 (1.15) 93.28 (1.73)<br />
In-network 87.01 (5.46) 70.60 (8.48)<br />
Cross-network 98.46 (0.61) 97.32 (0.80)<br />
1.50σ 2 Overall 96.13 (1.29) 92.33 (1.77)<br />
In-network 84.38 (6.01) 66.28 (8.52)<br />
Cross-network 98.23 (0.67) 96.97 (0.84)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 55 / 77
Experiments Models Connectivity Prediction Summary<br />
Other Connectivity Methods<br />
Model-based methods (Bayesian):<br />
Chen et al., 2013, submitted.<br />
Xue <strong>and</strong> Bowman, 2013, submitted.<br />
Patel et al., 2006a, Human Brain Mapping.<br />
Patel et al., 2006b, Human Brain Mapping.<br />
Whole-brain complex networks (graph theoretic methods)<br />
Simpson et al., 2013, Statistics Surveys.<br />
Effective Connectivity<br />
Granger causality<br />
Structural equation modeling<br />
Dynamic causal modeling<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 56 / 77
Experiments Models Connectivity Prediction Summary<br />
Prediction Studies<br />
Emerging direction in neuroimaging<br />
Increase clinical applicability<br />
Use imaging data to predict clinical outcomes (e.g. to distinguish<br />
treatment responders <strong>and</strong> non-responders)<br />
We address intermediate objective of predicting neural responses<br />
Forecast neural representations of disease progression<br />
Predict neural responses to various treatments<br />
We develop a spatial modeling framework within this prediction<br />
context<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 57 / 77
Experiments Models Connectivity Prediction Summary<br />
Motivating Data Example<br />
From the Alzheimer’s Disease Neuroimaging Initiative (ADNI)<br />
database http://www.loni.ucla.edu/ADNI/.<br />
Goal of ADNI: to develop biomarkers of AD in elderly subjects.<br />
Study participants receive PET scans at: baseline, 6 months, 12<br />
months <strong>and</strong> 24 months.<br />
In our analysis, we used<br />
Baseline <strong>and</strong> month 6 scans.<br />
Subjects classified as: AD patients or healthy controls (HC).<br />
Training data set: 40 AD <strong>and</strong> 40 HC subjects; Testing data set: 33 AD<br />
<strong>and</strong> 33 HC subjects.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 58 / 77
Experiments Models Connectivity Prediction Summary<br />
Bayesian spatial hierarchical model<br />
[Derado et al., 2012, Statistical Methods in Medical Research].<br />
Notation<br />
We developed a Bayesian spatial hierarchical model for predicting<br />
follow-up neural activity based on baseline functional neuroimaging<br />
data <strong>and</strong> other patient characteristics.<br />
Model borrows strength from spatial correlations present in the data.<br />
Let i = 1, . . . , n denote subjects, v = 1, . . . , V voxels, g = 1, . . . , G<br />
regions.<br />
Let Y (v) denote the regional cerebral blood flow (rCBF) (a proxy for<br />
brain activity) at voxel v.<br />
(<br />
) T<br />
Let Y ig (v) = Y ig (v) (1) , Y ig (v) (2) , (1)=baseline, (2)=follow-up.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 59 / 77
Experiments Models Connectivity Prediction Summary<br />
Spatial dependence: 3D neighborhood<br />
For each voxel in the analysis, we define a 3D neighborhood as the 26<br />
immediate neighboring voxels.<br />
Borrow strength locally.<br />
We consider only within-region neighbors.<br />
This information is saved in a connectivity matrix W.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 60 / 77
Experiments Models Connectivity Prediction Summary<br />
Model<br />
Y ig (v)|β g , φ g , α ig , γ gv , Ω g ∼ N(β g (v) + φ g (v) + α ig + X ig γ g , Ω g )<br />
φ g (v)|φ g (v ′ ), v ≠ v ′ , ρ, Σ ∼ N ( ρ ∑ w vv ′<br />
w v+<br />
φ g (v ′ 1<br />
),<br />
w v+<br />
Σ ) ( MCAR(ρ, Σ) )<br />
β gj (v)|λ 2 gj ∼ N(β 0gj , λ 2 gj) (λ vgj = λ gj , ∀v ∈ region g)<br />
Ω −1<br />
g ∼ Wishart ( (c 1 Z 1 ) −1 )<br />
, c 1<br />
Σ −1 ∼ Wishart ( (c 2 Z 2 ) −1 , c 2<br />
)<br />
α ij |Γ j ∼ N(0, Γ j )<br />
(α ij = α (j)<br />
i<br />
Γ −1<br />
j<br />
∼ Wishart((h j H j ) −1 , h j ) j = 1, 2<br />
λ −2<br />
gj<br />
∼ Gamma(a j , b j )<br />
γ gjq |τ 2 gjq ∼ N(0, τ 2 gjq)<br />
τ −2<br />
gjq ∼ Gamma(e 0, f 0 )<br />
= (α (j)<br />
i1 , . . . , α(j) iG )′ )<br />
q = 1, . . . , Q (covariates)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 61 / 77
Experiments Models Connectivity Prediction Summary<br />
Estimation <strong>and</strong> Prediction<br />
Estimation is performed using MCMC methods (Gibbs)<br />
Prediction:<br />
For region g, we can write<br />
Y g = (Y T g,1 , YT g,2 )T ∼ N ( (µ T g,1 , µT g,2 )T , Ψ g<br />
)<br />
, where Ψg = Σ g ⊗ I Vg .<br />
Then E(Y i ∗ g,2|Y i ∗ g,1) = b i ∗ g , where<br />
b i ∗ g = µ i ∗ g,2 + Ψ T g,12Ψ −1<br />
g,11 (Y i ∗ g,1 − µ i ∗ g,1)<br />
<strong>and</strong> µ i ∗ g = β g + φ g + 1 Vg ⊗ α i ∗ g + 1 Vg ⊗ X i ∗ gv γ gv .<br />
Estimated conditional mean ˆb i ∗ g obtained from inputting the posterior<br />
means of the corresponding parameters<br />
The follow-up brain activity Y i ∗ g,2 is predicted using the estimated<br />
conditional mean ˆb i ∗ g .<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 62 / 77
Experiments Models Connectivity Prediction Summary<br />
Figure: Individualized observed <strong>and</strong> predicted 6 month follow-up rCBF<br />
measurements for a test subject in the AD group (axial slice 40).<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 63 / 77<br />
Results: Prediction for PET study of AD<br />
Subject 1: observed Y (2) Subject 1: predicted Y (2)
Experiments Models Connectivity Prediction Summary<br />
Figure: Individualized observed <strong>and</strong> predicted 6 month follow-up rCBF<br />
measurements for a test subject in the AD group (axial slice 40).<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 64 / 77<br />
Results cont.<br />
Subject 19: observed Y (2) Subject 19: predicted Y (2)
Experiments Models Connectivity Prediction Summary<br />
Prediction Error<br />
We evaluate the prediction error using a scale-free (squared error) loss<br />
function, which adjusts for local magnitude of brain activity<br />
√<br />
1<br />
∑<br />
(2) ) N<br />
(v)}, {Ŷ i (v)} =<br />
N i=1 [Ŷ (2) (2)<br />
i (v) − Y<br />
stPMSE ( {Y (2)<br />
i<br />
1<br />
N<br />
∑ N<br />
i=1 Y (2)<br />
i<br />
(v)<br />
i<br />
(v)] 2<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 65 / 77
Experiments Models Connectivity Prediction Summary<br />
Comparative <strong>An</strong>alyses<br />
Model BSPM BSMac GLM<br />
Aver. error 0.083 0.154 (85.5% increase) 0.157 (89.2% increase)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 66 / 77
Experiments Models Connectivity Prediction Summary<br />
Simulation Study<br />
We simulated 100 data sets for 15 subjects.<br />
Selected 5 AAL regions with sizes ranging from 234 to 4,655 voxels.<br />
Specified the true values for β g , φ g , α i , <strong>and</strong> γ g .<br />
The rest of the (hyper)parameters drawn from their prior distributions.<br />
Param.<br />
Region<br />
1 2 3 4 5<br />
True Est. True Est. True Est. True Est. True Est.<br />
Ω g,11 323.26 321.09 160.61 159.99 11.46 11.46 3.44 3.44 3.57 3.57<br />
Ω g,22 120.75 120.46 45.30 45.11 37.24 37.19 10.45 10.45 3.38 3.38<br />
Ω g,12 -41.95 -41.25 -33.70 -33.40 16.32 16.30 4.08 4.08 2.40 2.40<br />
Table: Summary of the simulation results for the parameters in the covariance<br />
matrix Ω g .<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 67 / 77
Experiments Models Connectivity Prediction Summary<br />
NIH Parkinson’s Disease Biomarker Program<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 68 / 77
Experiments Models Connectivity Prediction Summary<br />
Parkinson’s Disease<br />
PD is a chronic, progressive movement disorder affecting at least one<br />
half million patients in the U.S.<br />
No diagnostic test to validate PD determination<br />
Diagnosis: medical history, neurological exam, <strong>and</strong> treatment response<br />
Brain scans <strong>and</strong> lab tests possibly used to rule out other diseases<br />
Symptoms<br />
Tremor, rigidity, bradykinesia (slowness of movement), <strong>and</strong> impaired<br />
balance <strong>and</strong> coordination<br />
With progression, patients may have difficulty walking, talking, or<br />
completing other simple tasks<br />
Non-motor symptoms: cognition, mood, sleep, <strong>and</strong> autonomic function<br />
PD usually affects people over 50 years of age<br />
Drug (e.g. levodopa combined with carbidopa) <strong>and</strong> surgical (DBS)<br />
treatments available, but none halt progression<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 69 / 77
Experiments Models Connectivity Prediction Summary<br />
Parkinson’s Disease Biomarkers<br />
Motivation<br />
The optimal window for<br />
neuroprotecion in PD is during<br />
the premotor period before<br />
most neuronal death occurs.<br />
Neurology 2009;72(Suppl 2):S21-S26.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 70 / 77
Experiments Models Connectivity Prediction Summary<br />
Multimodal PD biomarker detection<br />
Suggested links between PD <strong>and</strong> single<br />
genetic, imaging, <strong>and</strong> biologic factors.<br />
Many of these are non-specific or<br />
insensitive.<br />
Single modality biomarkers may not fully<br />
address the complexity of PD.<br />
We regard PD as a complex, systems-level,<br />
multi-dimensional disorder with discrete, but<br />
functionally integrated manifestations.<br />
We are developing methods to define<br />
multimodal PD biomarkers from a massive<br />
number of hypothesis driven <strong>and</strong> exploratory<br />
c<strong>and</strong>idate markers.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 71 / 77
Experiments Models Connectivity Prediction Summary<br />
Data<br />
Multimodal Parkinson’s Disease Imaging Studies<br />
The data will include 81 subjects<br />
across three studies<br />
38 Parkinson’s disease patients<br />
32 Healthy control subjects<br />
11 Alzheimers disease patients<br />
Potential Biomarkers<br />
Imaging (MRI, fMRI, DTI,<br />
NM-MRI, CSI)<br />
Genetic<br />
Neurocognitive Testing<br />
Questionnaire-Derived Scores<br />
Clinical<br />
CSF Neuroinflammation<br />
CSF Catecholamine Metabolites<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 72 / 77
Experiments Models Connectivity Prediction Summary<br />
Methods<br />
High-dimensional Data Methods<br />
Develop statistical models, which pool predictive strength across<br />
multiple data modalities<br />
Proposed statistical techniques<br />
Penalized likelihood approaches:<br />
l p (β, λ) = −l(β) + ∑ k λ kP k (β k )<br />
l(β) = ∑ i [y ilogπ i + (1 − y i )log(1 − π i ]<br />
Variable selection or shrinkage<br />
Novel joint multimodal probability models<br />
Model development, training, testing, <strong>and</strong> validation<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 73 / 77
Experiments Models Connectivity Prediction Summary<br />
Bayesian Spatial Prediction Model<br />
We blindly classify subjects as either PD or HC<br />
from multimodal imaging data (<strong>and</strong> other patient<br />
information)<br />
Two-level parcellation: AAL Regions → Subregions<br />
Spatial correlations:<br />
Between AAL regions: Unstructured covariance<br />
matrix<br />
Borrow strength between subregions within each<br />
AAL region: CAR model<br />
Between voxels within each subregion:<br />
Exchangeable structure<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 74 / 77
Experiments Models Connectivity Prediction Summary<br />
Bayesian Spatial Prediction Model<br />
Example of modelling framework using ALFF <strong>and</strong> VBM data<br />
Let D i ∈ {0, 1} represent disease status (e.g. 1=PD, 0=HC)<br />
Let X ilg (v) <strong>and</strong> Z ilg (v) respectively denote the ALFF <strong>and</strong> VBM for<br />
subject i at voxel v in subregion l (in region g), <strong>and</strong><br />
B i = [X ilg (v), Z ilg (v)]<br />
Construct a Bayesian joint (multimodal) probability model<br />
f (X ilg (v), Z ilg (v)|D i ) ∝ f 1 (X ilg (v)|Zilg(v), D i ) × f 2 (Z ilg (v)|D i )<br />
Generate predictions for a new subject n + 1 based on<br />
Pr(D n+1 = k|B n+1 , {B i , D i } n i=1 )<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 75 / 77
Experiments Models Connectivity Prediction Summary<br />
Summary<br />
Statistical methods play important roles in the analysis of<br />
neuroimaging data<br />
identify biomarkers of disease <strong>and</strong> treatment response.<br />
detect disease- or treatment-related alterations in brain function<br />
determine brain networks<br />
Multimodal analyses may yield more accurate <strong>and</strong> robust results than<br />
single genetic, imaging, biological, or clinical markers.<br />
Leveraging available information in the data (e.g. borrowing strength<br />
across brain areas) may yield more precise estimates.<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 76 / 77
Experiments Models Connectivity Prediction Summary<br />
Acknowledgements<br />
Research Team<br />
Lijun Zhang, PhD, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
Jian Kang, PhD, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
Ying Guo, PhD, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
Gordana Derado, PhD, CDC<br />
Shuo Chen, PhD, Univ. of Maryl<strong>and</strong>, Epidemiology <strong>and</strong> Biostatistics<br />
Daniel Huddleston, MD, Kaiser Permanente Georgia<br />
Xiaoping Hu, PhD, Emory/GA Tech, Biomedical Engineering<br />
Helen Mayberg, MD, Emory, Psychiatry <strong>and</strong> Behavioral Sciences<br />
Doctoral Students:<br />
Wenqiong Xue, MS, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
<strong>An</strong>thony Pileggi, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
Phebe L. Brenne, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
Qing He, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
Yize Zhao, Emory, Biostatistics <strong>and</strong> Bioinformatics<br />
NINDS U18 NS082143-01 (Bowman) <strong>and</strong> the PDBP<br />
NIMH R01-MH079251 (Bowman)<br />
F. D. Bowman (Emory University) <strong>SAMSI</strong>: NDA RTP, NC 77 / 77