fmri experimental design - mitweb.mit.edu/hst.583/www/course2001/lectures/hst583_lect...because fmri...
TRANSCRIPT
FMRI Experimental Design
Lila DavachiDepartment of Brain & Cognitive Sciences, M.I.T.
Because fMRI BOLD data is not anBecause fMRI BOLD data is not anabsolute measure of neuronal activity,absolute measure of neuronal activity,
all study designs must provide theall study designs must provide theopportunity to statistically contrast theopportunity to statistically contrast the
neuronal activity of interest with aneuronal activity of interest with asuitable rest or background conditionsuitable rest or background condition..
Thus, study design is Thus, study design is of paramount importance.of paramount importance.
Why?Why?HypothesisHypothesis
How?How?fMRI Study DesignfMRI Study Design
Where?Where?NeuroanatomyNeuroanatomy
What?What?BehaviorBehavior
Thanks to Chantal Stern
43 * 7 = ?
Key Points
• What can fMRI tell you?
• Always comparing across conditions
• Characteristics of the hemodynamic response(HRF) and how this affected the sequentialdevelopment of fMRI paradigms and influencesstudy design
• Sense of important design issues
What (good) is fMRI?
What it can tell you:
• Relative local “neural” activity (LFP’s ?)• NOT absolute neural activity• NOT excitation vs inhibition• NOT about necessity of a given region for a task• NOT fine-grained temporal information
Key Points
• What can fMRI tell you?
• Always comparing across conditions
• Characteristics of the hemodynamic response(HRF) and how this affected the sequentialdevelopment of fMRI paradigms and influencesstudy design
• Sense of important design issues
Subtraction ParadigmDonder’s method:
Ex: How to measure time of a mental transformation?
A random series of A’s and B’s presented and the subject must:
1. Respond whenever an event occurs (RTi)
2. Respond only to A not to B (RTii)
3. Respond X to A and Y to B (RTiii)
RTi = RT(detect) + RT(response)
RTii = RT(detect) + RT(discrimination) + RT(response)
RTiii = RT(detect) + RT(discrimination) + RT(choice) + RT(response)
THUS, RT(discrimination) = RTii - RTi
RT(choice) = RTiii - RTii
Criticisms of Subtraction Paradigm
1. That we already know what ‘counts’ as a single mentalprocess (i.e. choice is a single mental process?)
2. Assume that adding components does not affect otherprocesses (i.e. assumption of pure insertion)
THUS, one should pick tasks that differ along ONEdimension (either change the task OR the stimuli but notBOTH!)
And a resting baseline is good to include, however, theinterpretation should be taken lightly…(more later)
The loose task comparisonThe loose task comparison
Does Does notnot hold all variables constant BUT: hold all variables constant BUT:
(1) Uses a low level reference task(1) Uses a low level reference task
(2) Allows the data to be examined for predictable(2) Allows the data to be examined for predictablestimulus or response drivenstimulus or response driven activations activations
(3) Allows the more extensive activation pattern to(3) Allows the more extensive activation pattern tobe observedbe observed
The “loose” TaskThe “loose” TaskComparisonComparison
STISTISTIPREPREPRE
DOGDOGDOG
+++ +++ +++ +++
GREGREGRE
STISTISTIPREPREPRE
PAINTPAINTPAINT STRSTRSTR
STISTISTIPREPREPRE
EGGEGGEGGTASK 1TASK 1
TASK 2TASK 2
The tight task comparisonThe tight task comparison
Try to hold all variables constant including:Try to hold all variables constant including:
•• Stimulus display (nominally or statistically) Stimulus display (nominally or statistically)
•• Response and response selection characteristics Response and response selection characteristics
•• Performance level- especially if comparing cohorts Performance level- especially if comparing cohorts
•• Eye movements Eye movements
•• Emotional state (minimize anxiety and boredom) Emotional state (minimize anxiety and boredom)
The “tight” TaskThe “tight” TaskComparisonComparison
STISTISTIPREPREPRE
DOGDOGDOG
+++ +++ +++ +++
GREGREGRE
STISTISTIPREPREPRE
PAINTPAINTPAINT STRSTRSTR
STISTISTIPREPREPRE
EGGEGGEGGTASK 1TASK 1
TASK 3TASK 3
TASK 2TASK 2
22 mmiinnuuss 11TASK 1TASK 1 TASK 2TASK 2
BRAIN AREAS THAT DIFFERBRAIN AREAS THAT DIFFER
ALL ACTIVE BRAINALL ACTIVE BRAINAREASAREAS
Thanks to Randy Buckner
Example...
Interested in semantic processing and how it affectsmemory...
Parameters to specify in any experiment
1. Subjects: normal vs special populations
2. What part of brain look at? How many slices canyou have for your TR?
3. Choosing your TR: How often can you take a full setof pictures
4. What coil will you use?surface coils: higher SNR, only partial coveragehead coils: lower SNR, complete coverage
5. Toggle many times between conditions within a scan
6. Run as many scans as possible within a subject
Key Points
• What can fMRI tell you?
• Always comparing across conditions
• Characteristics of the hemodynamic response(HRF) and how this affected the sequentialdevelopment of fMRI paradigms and influencesstudy design
• Sense of important design issues
Visual Stimulation - 2 sec Flashes
0 10 20 30 40 50 60 70 80 90 100 110 120
TIME (seconds)
Per
cent
Sig
nal C
hang
e
0.0
1.0
2.0
3.0
4.0
-1.0
[Blamire, Ogawa et al., PNAS, 1992]
Visual Cortex During Brief Visual Stimulation
0.0
1.0
2.0
5 0 5 10 15
34
100
1000msec
TIME (seconds)
msec
msec
[Courtesy of Robert Savoy, Kathleen O’Craven MGH-NMR Center]
Blocked design fMRI
BLOCKED:
HORSE
(abstract or concrete?)
love
(upper or lowercase?)
“Blocked” fMRI:Memory Paradigm
STIPRE
HORSE
+ +
0 8 74 118 140 184 206 250
GRE
STIPRE
LOVE STR
STIPRE
HUMOR
TIME (SEC)52
+ +
STR
STIPRE
CHAIR
+
[Wagner et al., OHBM, 1998]
Typical Blocked-Design Response
-0.25
0
0.25
0.50
Per
cent
Sig
nal C
hang
e
0 20 40 60 80 100 120 140 160 180 200 220 240
Time
HORSE LOVE HUMOR CHAIR
13 SlicesPer
Brain Image
Thanks to Robert Savoy
80 brain imagesper 4 minute
run
Thanks to Robert Savoy
For purposes of illustration…….
Thanks to Robert Savoy
Examine thedata fromone slice
of the brainas a function
of time
Thanks to Robert Savoy
Thanks to Robert Savoy
Are these voxellevels statistically
different from each other?
Combinethese
Combinethese
Thanks to Robert Savoy
Combinethese
Combinethese
Consider EACH voxelacross all time points
Thanks to Robert Savoy
Typical Blocked-Design Response
-0.25
0
0.25
0.50
Per
cent
Sig
nal C
hang
e
0 20 40 60 80 100 120 140 160 180 200 220 240
Time
HORSE LOVE HUMOR CHAIR
Event-Related fMRI
BLOCKED:
SPACED EVENT-RELATED:
16 sec
“Spaced Event-Related” fMRI:Language Paradigm
[Buckner, Bandettini et al., PNAS, 1996]
TIME (seconds)
0 32 64 96
COU PRE TRA AFT STA PEL DRI
“Single-Trial” Response Across a Run“Single-Trial” Response Across a Run
TIME (SEC)TIME (SEC)
00 3232 6464 9696 128128 160160 192192 228228 256256
375375
376376
377377
378378
379379
380380
381381
ME
AN
MR
SIG
NA
LM
EA
N M
R S
IGN
AL
“Event-Related” Selectively Averaged Response
0
.5
0 1 2 3 4 5 6 7 8 9 101112131415
TIME (seconds)
PE
RC
EN
TS
IGN
AL
CH
AN
GE
375
376
377
378
379
380
381
ME
AN
MR
SIG
NA
L
374
375
376
377
378
379
0 2 4 6 8 10 12 14TIME (sec)
Broca’s Area During Language Paradigm
Thanks to Randy Buckner
“Rapid Event-Related” fMRI
BLOCKED:
SPACED EVENT-RELATED:
16 sec
RAPID EVENT-RELATED:2 sec
20 sec0 sec
1 secon
Assessing the Linearity Hypothesis
[Dale and Buckner, Hum. Brain Map., 1997]
Response to Averaged Single Trials
-1
0
1
2
3
4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (sec)
PE
RC
EN
T M
R S
IGN
AL
Split-half reliability
Thanks to Randy Buckner
20 sec0 sec
20 sec0 sec 5 sec
Assessing the Linearity Hypothesis:5 Second ITI
Thanks to Randy Buckner
-1
0
1
2
3
4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (sec)
PE
RC
EN
T M
R S
IGN
AL
Response to Averaged Double Trials
Thanks to Randy Buckner
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
TIME (SEC)0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
TIME (SEC)
Assessing the Linearity Hypothesis:Separation of Responses
Thanks to Randy Buckner
-1
0
1
2
3
4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (sec)
Assessing the Linearity Hypothesis:Separation of Responses
FIRST TRIAL
ESTIMATEDSECOND TRIAL
Thanks to Randy Buckner
20 sec0 sec
0 sec2 sec 20 sec
0 sec2 sec 20 sec4 sec
Assessing the Linearity Hypothesis:2 Second ITI
Thanks to Randy Buckner
-1
0
1
2
3
4
5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (SEC)
-1
0
1
2
3
4
5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (SEC)
RAW DATA ESTIMATED RESPONSES
Responses to Multiple Rapidly Intermixed Trials
THREE TRIAL
TWOTRIAL
ONETRIAL
FIRST TRIAL
ESTIMATEDSECOND TRIAL
ESTIMATEDTHIRD TRIAL
Thanks to Randy Buckner
Structuring Event-Related Trial Presentations
[Burock et al., Neuroreport 1998]
Fixed Interval Presentation
Randomized Presentation
ITI = 16 sec 3 sec 1 sec
Nor
mal
ized
% S
igna
l Cha
nge
8
6
4
2
0
8
6
4
2
00 100 200 300 400
Time (sec)
Variance Associated with Fixed Interval Designs
Seven unknowns, BUTonly three independentequations
INDIVIDUALBOLD
RESPONSE
TIME
h7 + h4 + h1
MEASUREDBOLD
RESPONSE
h6 + h3
h5 + h2
h4 + h1 + h7
h3 + h6 h2 + h5
h1 + h4 + h7
= TR
6 12 18 24 30
Thanks to Randy Buckner
Variance Associated with Jittered Designs
INDIVIDUALBOLD
RESPONSE
TIME
Seven unknowns, ANDmore than sevenindependent equations
MEASUREDBOLD
RESPONSE
h7 + h5 + h3
h6 + h4 + h2
h5 + h3 + h1
h4 + h2
h3 + h1 + h7 h2 + h6
h1 + h5
h7 + h4
h6 + h3
Thanks to Randy Buckner
Does the neural correlate of priming vary with the lagbetween the first and second episode within asemantic task?
PEACE+
ANVIL
PEACE+
TABLE
+ENERGY
ANVIL~2 min
~25 hr
ABSTRACT or CONCRETE?
Sorting Based onExperimenter Determined Conditions
[Wagner et al., J. Cognitive Neuroscience 2000]
Rest
Rest
Novel
Long Prime
Short Prime
Long Prime
Short Prime
Novel
Rest
Rest
Novel
Long Prime
Short Prime
Long Prime
Short Prime
Novel
Group datatogether
Shorter Lags Yield Greater Neural Priming
Anterior IFG
-0.1
0
0.1
0.2
0.3
% SignalChange
0 2 4 6 8 10 12 14Time (s)
ShortLongNovel
Thanks to Anthony Wagner
TRUTH
HOUSE
PEACE
ANVIL
TAPEREMEMBERED
FORGOTTEN
Sorting Based on Subject Behavior:Subsequent Memory Performance
[Wagner et al., Science 1998]
Neural Regions Predicting Subsequent Memory
ForgottenRemembered
Inferior Prefrontal Gyrus
-1
01234
0 2 4 6 8 10 12 14Time (s)
-101234
0 2 4 6 8 10 12 14Time (s)
B
Left Posterior Parahippocampus
Thanks to Anthony Wagner
Key Points
• What can fMRI tell you?
• Always comparing across conditions
• Characteristics of the hemodynamic response(HRF) and how this affected the sequentialdevelopment of fMRI paradigms and influencesstudy design
• Sense of important design issues
Critical issues in paradigm designCritical issues in paradigm design
•• Poorly defined Poorly defined neuroanatomical neuroanatomical hypothesishypothesis
•• Poorly controlled baseline Poorly controlled baseline
•• Attentional Attentional effects effects
•• Learning effects Learning effects•• Stimulus habituation or sensitization Stimulus habituation or sensitization
•• System and physiological drift System and physiological drift
Baseline, what is it?
Ex: if want to say something about verb generation andcompare it only to reading aloud..
BUT, still do not know if these regions are involved inreading only (thus can include a low level reading condition..)
No inherent “0” baseline for cognition,i.e. what are subjects doing when asked to do nothing?
• Ans: they are doing a lot• how interpret deactivations?
Issues: Generality vs SpecificityHypothesis: Region X is involved in process Y.
Evidence: Region X is activated when subjects do an instanceof process Y
Problem: Without running several further conditions, we can’ttell whether region X might instead be involved in somethingeither more SPECIFIC or more GENERAL than Y.
Example:
Hypothesis Space: Function of Region X
hands
feeteyes
2-toned faces
facesProfileFaces
FacesW/ohair
IDphotos
Any human body part
Anything animate
Thanks to Nancy Kanwisher
Issues: Attentional ConfoundsA given region might respond more strongly in condition A thancondition B simply because A is more interesting/attention-capturing than B.
Solutions:
1. Double Dissociations, i.e. faces versus objects?
2. Test conditions with opposite attentional predictionsi.e. passive viewing vs 1-back task
Thanks to Nancy Kanwisher
Issues: Statistical Significance vs.Theoretical Significance
P levels alone are not sufficientFor example, the FFA may respond significantly more to pineapplesthan watermelons, but the response to pineapples might nonethelessbe much lower than the response to faces.
Solutions:Quantity effect size, e.g. with percent signal change
Provide “benchmark” conditions within the same scan to give thesemagnitudes meaning
Objects Watermelon Pineapple Faces
0.6
0.6
0.7
0.7
0.9
1.8
2.0
2.0Thanks to Nancy Kanwisher
Acknowledgements
Anthony Wagner
Nancy Kanwisher
Randy Buckner
Robert Savoy
Randy Gollub
Chantal Stern
Data AnalysisA. General Issues• Individual vs Group Analyses: brains are very different BUT wantTo make a general claim ANS: do both if can
• Multiple Comparisons….if doing 20,000 T-tests, better not acceptp< .05
B. Methods• Simple comparisons, is X > Y ?, look in each voxel..
• Conjunction Analyses, are any voxels significant for both X>Yand A>B?
• Regression Analyses, obtaining weights for different regressors
• ROI-based Analyses
[Boyton et al., J. Neuroscience 1996]
Hemodynamic Response Summation:Linear Systems Approach
The fMRI response to a stimulus lasting a duration of NT is roughly alinear summation of N temporally shifted responses to a stimulus lastinga duration of T
“Mixed” fMRI
BLOCKED:
RAPID EVENT-RELATED:2 sec
MIXED:
[Chawla et al., Nat. Neuroscience 1999; Donaldson et al., NeuroImage, 2001]
TIME (SEC)
= OLD WORD = NEW WORD
MEMORY
FIXATION
MEMORYA
B
FIXATION FIXATION
Mixed Blocked/Event-relatedDesign
Donaldson et al., NeuroImage, 2001 (see also Chawla et al., Nature Neurosci., 1999)
TIME
TRANSIENT BOLDRESPONSETO EVENTS
MEASUREDBOLD
RESPONSE
SUSTAINED BOLDRESPONSE
TO SET
TIME
“Mixed” fMRI:Trial Separation with Task Blocking
Task 1 Task 2 Task 1
Analysis Strategies:
– Event-related analyses
• Task 1 trials ( ) vs. Task 2 trials ( )
• Trial type A ( ) vs. Trial type B ( )
[Badre et al., in prep.]
“Mixed” fMRI:Trial Separation with Task Blocking
Task 1 Task 2 Task 1
Analysis Strategies:
– Event and State effects• Same event contrasts
• Also model NULL components within blocks to explore “state” effects
CAVEAT: correlation between event and state regressors
vs vs
Task 1 vs. Task 2 Nulls