Finn ohbm2017 ed_workshop

Reliable individual functional networks and their relationship to behavior

Emily S. Finn, PhDSection on Functional Imaging MethodsLaboratory of Brain & Cognition, [email protected]

Taking Connectivity to a Skeptical Future Educational Workshop | OHBM Annual MeetingJune 25, 2017

@esfinn

Outline

Introduction:Why study individual differences?

1 How to get good brain data

2 How to get good behavioral data

3

Outline



2

How to relate brain to behavior

4

How to get good behavioral data

3

Outline



2


4

Learn more5How to get good behavioral data

3

Why study individual differences?

Controls

Patients


Controls

Patients

• Basic neuroscience: mapping from individual brains to individual behaviors


Controls

Patients


• Eventual clinical applications:


Categorical approach

Controls

Patients




Categorical approach

Controls

Patients



p < 0.05*


Categorical approach Dimensional approach

Controls

Patients



p < 0.05*

Why study individual differences?Motor

Mt

EmotionEm

LanguageLg

Working memoryWM

RestingR1

+

RestingR2

+

Why study individual differences?Motor

Mt

EmotionEm

LanguageLg

Working memoryWM

RestingR1

RestingR2


Finn et al., Nat Neurosci (2015)


0.93 0.84 0.63

0.72 0.79 0.67

0.64 0.60 0.54

R1

WM

Mt

R2 Lg Em

Data

base

Target

ID rate

0.5 1.0



You always look most like yourself, regardless of what you’re doing

0.93 0.84 0.63

0.72 0.79 0.67

0.64 0.60 0.54

R1

WM

Mt

R2 Lg Em

Data

base

Target

ID rate

0.5 1.0



You always look most like yourself, regardless of what you’re doing

0.93 0.84 0.63

0.72 0.79 0.67

0.64 0.60 0.54

R1

WM

Mt

R2 Lg Em

Data

base

Target

ID rate

0.5 1.0

Individual differences in FC predict individual differences in behavior


Outline



2


4


3

Q. Do you need HCP-quality data?

Q. Do you need HCP-quality data?A. Not really


ID is fairly robust even at more standard spatial & temporal resolutions:


Airan et al., Hum Brain Mapp (2016)





Courtesy of Jason Druzgal


• More nodes —> higher identification rate






‣ Parcellation method (random vs. functional) did not matter






‣ Parcellation method (random vs. functional) did not matter‣ Caution: Higher resolution may amplify effects of anatomical diffs/registration error






‣ Parcellation method (random vs. functional) did not matter‣ Caution: Higher resolution may amplify effects of anatomical diffs/registration error‣ Parcellations in the 200-300 node range seem like a good compromise




Q. What about amount of data?

Q. What about amount of data?A. Scan duration matters!


Longer acquisitions are better:


Longer acquisitions are better:• higher reliability within subjects

Birn et al., NeuroImage (2013)



• higher identifiability across subjects






‣ higher sampling rate (shorter TR) cannot make up for shorter scan duration








Shah et al., Brain & Behav (2016)



Q. Does scan condition matter?

Q. Does scan condition matter?A. Yes!

Q. Does scan condition matter?A. Yes!Rest has become the default condition for FC & individual differences, but tasks may increase signal-to-noise


0.93 0.84 0.63

0.72 0.79 0.67

0.64 0.60 0.54

R1

WM

Mt

R2 Lg Em

Data

base

Target

ID rate

0.5 1.0


Finn et al., NeuroImage (2017)


‣ for a set scan duration, tasks may be more reliable than rest‣ tasks may converge faster on a subject’s “true” profile


Q. Is rest best?

Q. Is rest best?A. Probably not


Tasks may stabilize individuals’ functional architecture, increase SNR:


‣ some task pairs give better ID than the two rest sessions




Target

Data

base

Day 1 Day 2

ID ra

te

Day

1D

ay 2





Target

Data

base

Day 1 Day 2

ID ra

te

Day

1D

ay 2

‣ consider a combination of rest and task



Q. Is rest best? A. Probably not


Consider naturalistic tasks:



Inscapes: Vanderwal et al., NeuroImage 2015headspacestudios.org

http://headspacestudios.org



‣ ID rate is just as good as (if not better than) rest



Session 1

Rest Inscapes Ocean’s 11

Session 2




Session 1

Rest Inscapes Ocean’s 11

Session 2


Vanderwal et al., NeuroImage (2017)


Tasks also have purely practical advantages:

‣ increase subject compliance (i.e., decrease head motion), especially in certain populations




Huijbers et al., NeuroImage (2017)




Child

ren

Adul

ts

Vanderwal et al., NeuroImage (2015)Huijbers et al., NeuroImage (2017)



Outline



2


4


3

How to choose behavior

How to choose behaviorIs it stable?


• Trait vs. state• State variables may be better

suited to within-subject analysis




Betzel et al., Sci Rep (2017)


Does it show a good distribution in your population?





Does it show a good distribution in your population?

Is it something you expect to be reflected in brain function?




Behavior: Mitigating confounds

Behavior: Mitigating confoundsMany behaviors/phenotypes are correlated with head motion!


Siegel et al., Cerebral Cortex (2016)

Negatively: Positively:


Siegel et al., Cerebral Cortex (2016)

Negatively: Positively:

Geerligs et al., Hum Brain Mapp (2017)

Age, vascular health:


‣ Check for correlation in your sample


‣ Check for correlation in your sample‣ Consider excluding particularly high-motion subjects


‣ Check for correlation in your sample‣ Consider excluding particularly high-motion subjects‣ Choose appropriate preprocessing techniques


‣ Check for correlation in your sample‣ Consider excluding particularly high-motion subjects‣ Choose appropriate preprocessing techniques

Ciric et al., NeuroImage (2017)


‣ Check for correlation in your sample‣ Consider excluding particularly high-motion subjects‣ Choose appropriate preprocessing techniques‣ Use motion as an explicit covariate

Ciric et al., NeuroImage (2017)

Outline



2


4


3

Brain ↔ behavior: Cross-validate!

Brain ↔ behavior: Cross-validate!Find a brain-behavior relationship in one sample, see if it holds in another sample


‣ leave-one-subject-out (within dataset)

Connectome-based Predictive ModelingShen et al., Nature Protocols (2017)




2

7.5

4

6.3




Pred

icte

d be

hav

Observed behav

2

7.5

4

6.3



‣ even better: cross-dataset

Sustained attention model


Pred

icte

d be

hav

Observed behav

2

7.5

4

6.3

Rosenberg et al., Nature Neuroscience (2016)Rosenberg et al., Trends in Cog Sci (2017)






Pred

icte

d be

hav

Observed behav

2

7.5

4

6.3


Observed behav

Pred

icted

beh

av

n = 25 adults






Pred

icte

d be

hav

Observed behav

2

7.5

4

6.3


Observed behav

Pred

icted

beh

av

n = 25 adults

Observed ADHD score

n = 113 children

Pred

icted

ADH

D sc

ore

Q. What is the best brain state?

Q. What is the best brain state?A. Maybe it depends on your behavior


Certain task conditions generate better predictions of behavior:


n = 716, 10-fold cross-validationConnectome-based Predictive Modeling (CPM; Shen et al., Nat Protocols 2017)

Mod

el in

put d

ata

Target behavior



n = 716, 10-fold cross-validationConnectome-based Predictive Modeling (CPM; Shen et al., Nat Protocols 2017)

Mod

el in

put d

ata

Target behavior


Consider tailoring scan condition to behavior of interest

➡“Stress test”?

Take-home points




3


4

Learn more5

Take-home points




3


4

Learn more5

• Data quality ≠ critical

Take-home points




3


4

Learn more5

• Data quality ≠ critical• Longer scans are better!

Take-home points




3


4

Learn more5

• Data quality ≠ critical• Longer scans are better!• Consider using tasks

Take-home points




3


4

Learn more5

• Data quality ≠ critical• Longer scans are better!• Consider using tasks ‣ improve compliance

Take-home points




3


4

Learn more5

• Data quality ≠ critical• Longer scans are better!• Consider using tasks ‣ improve compliance‣ increase sensitivity

Take-home points




3


4

Learn more5


• Choose an interesting behavior with a good distribution

Take-home points




3


4

Learn more5



• Consider potential confounds and take steps to mitigate them

Take-home points




3


4

Learn more5




• Cross-validate whenever possible

Take-home points




3


4

Learn more5




• Cross-validate whenever possible

• Consider tailoring your scan condition to behavior of interest

Open data sets with brain & behavior

➡ Use these on their own or in combination with your own data to generate or test hypotheses, see if a finding generalizes, etc

Philadelphia Neurodevelopmental Cohort

Further reading

Building a science of individual differences from fMRIDubois & Adolphs, Trends in Cognitive Sciences (2016)

From regions to connections and networks: new bridges between brain and behaviorMisic & Sporns, Current Opinion in Neurobiology (2016)

Can brain state be manipulated to emphasize individual differences in functional connectivity?Finn et al., NeuroImage (2017)

Prediction as a humanitarian and pragmatic contribution from human cognitive neuroscienceGabrieli, Ghosh & Gabrieli, Neuron (2015)

Selected reviews:

Learn more at OHBM 2017

Symposium: Collect your thoughts: Individual differences in the networks underlying intelligenceTues 8:00-9:15am

Symposium:Relating connectivity to inter- and intra-individual differences in attention and cognitionWeds 8:00-9:15am

Poster #4042: Can brain state be manipulated to emphasize individual differences in functional connectivity? Finn et al.

Poster #4040: Large-scale functional connectivity networks predict attention fluctuations Rosenberg et al.

Poster #2110: Functional connectivity-based predictors of naturalistic reading comprehension Jangraw et al.

Poster #4029: FMRI connectivity is differentially associated with performance across tasks in a multi-task study Topolski et al.

@esfinn

Finn ohbm2017 ed_workshop

Science

Transcript of Finn ohbm2017 ed_workshop