FMRI Data Analysis II Advanced Data Analysis FMRI

  • Slides: 38
Download presentation
FMRI Data Analysis: II. Advanced Data Analysis FMRI Undergraduate Course (PSY 181 F) FMRI

FMRI Data Analysis: II. Advanced Data Analysis FMRI Undergraduate Course (PSY 181 F) FMRI Graduate Course (NBIO 381, PSY 362) Dr. Scott Huettel, Course Director FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Advanced Data Analyses • Complex modeling • Analyses of Connectivity – Functional Connectivity Analysis

Advanced Data Analyses • Complex modeling • Analyses of Connectivity – Functional Connectivity Analysis – Causality analysis – Across-subjects regularities • Independent Components Analysis • Prediction – Real-time analyses – Correlation techniques – Support Vector Machines FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Big Concept • Your analysis model should not determined by your stimuli. • It

Big Concept • Your analysis model should not determined by your stimuli. • It should be determined by your hypothesis about the underlying cognitive processes. You can construct and test an arbitrarily complex model, if that model is justified by the brain processes. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Daw and colleagues (2006) used a “four-arm bandit” gambling task. In this task, subjects

Daw and colleagues (2006) used a “four-arm bandit” gambling task. In this task, subjects sometimes exploit a winning arm, and sometimes explore to learn about new arms. They not only did analyses based on how much subjects won (top row), but also on how predictable was the subject’s decision (bottom row), which reflects how much reward the subject expected. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Suppose that our experimental was based on the game show “Deal or No Deal”.

Suppose that our experimental was based on the game show “Deal or No Deal”. How could we model the subject’s cognition? FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Data-Driven Analyses • Broadly considered, they examine the data to identify coherent patterns. •

Data-Driven Analyses • Broadly considered, they examine the data to identify coherent patterns. • Complement hypothesis-driven analyses (e. g. , GLM) • The primary challenge: interpretation FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Functional Connectivity Seed voxel in “b”. Colormap shows voxels with r > 0. 35.

Functional Connectivity Seed voxel in “b”. Colormap shows voxels with r > 0. 35. Active Task Resting State! Biswal et al. (1995) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Let’s just pick a voxel in the posterior cingulate and look at its connectivity.

Let’s just pick a voxel in the posterior cingulate and look at its connectivity. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Resting-state connectivity (positive) for the posterior cingulate cortex (PCC, arrow). Resting-state connectivity (negative) for

Resting-state connectivity (positive) for the posterior cingulate cortex (PCC, arrow). Resting-state connectivity (negative) for lateral prefrontal cortex. PCC correlation is similar during active task and resting state. FMRI – Week 10 – Analysis II Greicius et al. (2003) Scott Huettel, Duke University

Causality FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Causality FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Granger Causality “The basic "Granger Causality" definition is quite simple. Suppose that we have

Granger Causality “The basic "Granger Causality" definition is quite simple. Suppose that we have three terms, Xt, Yt, and Wt, and that we first attempt to forecast Xt+1 using past terms of Yt and Wt. We then try to forecast Xt+1 using past terms of Xt, Yt, and Wt. If the second forecast is found to be more successful, according to standard cost functions, then the past of Y appears to contain information helping in forecasting Xt+1 that is not in past Xt or Wt. … Thus, Yt would "Granger cause" Xt+1 if (a) Yt occurs before Xt+1 ; and (b) it contains information useful in forecasting Xt+1 that is not found in a group of other appropriate variables. ” - Clive Granger, 2003 Nobel Laureate in Economics. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Konya (2006) Do changes in the exports of a country (Granger) cause changes in

Konya (2006) Do changes in the exports of a country (Granger) cause changes in that country’s gross domestic product? That is, does export activity lead economic growth? FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Influence Delayed by <100 ms Simulated Activity (LFP) Fx y Fy x Fx, y

Influence Delayed by <100 ms Simulated Activity (LFP) Fx y Fy x Fx, y FMRI gives information about correct causality (blue), but also introduces spurious simultaneous influence (red). Roebroeck et al. (2005) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Red = Source Green = Inputs Blue = Targets Roebroeck et al. (2005) FMRI

Red = Source Green = Inputs Blue = Targets Roebroeck et al. (2005) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Intersubject Commonalities Hasson et al. (2004) FMRI – Week 10 – Analysis II Scott

Intersubject Commonalities Hasson et al. (2004) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Hasson et al. (2004) FMRI – Week 10 – Analysis II Scott Huettel, Duke

Hasson et al. (2004) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Hasson et al. (2004) FMRI – Week 10 – Analysis II Scott Huettel, Duke

Hasson et al. (2004) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Independent Components Analysis (ICA) Assumption: The observed data is the sum of a set

Independent Components Analysis (ICA) Assumption: The observed data is the sum of a set of inputs which have been mixed together in an unknown fashion. The goal of ICA is to discover both the inputs and how they were mixed. Mc. Keown, et al. (1998) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Value of Component 1 Principal Components Analysis (PCA) finds a set of components that

Value of Component 1 Principal Components Analysis (PCA) finds a set of components that are uncorrelated. The first principal component gives the direction of maximal variance in the data. Value of Component 1 The assumption of temporal noncorrelation can be violated by some forms of structure in the data. The assumption of spatial non-correlation is violated when a given voxel contributes to more than one process. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

“The brain is not orthogonal!” Cf. Makeig, others. FMRI – Week 10 – Analysis

“The brain is not orthogonal!” Cf. Makeig, others. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Visual Cortex Heartbeat Breathing Head Motion? Vascular Oscillations? Mc. Keown et al, (2003) FMRI

Visual Cortex Heartbeat Breathing Head Motion? Vascular Oscillations? Mc. Keown et al, (2003) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

MELODIC (FSL’s version of ICA) FMRI – Week 10 – Analysis II Scott Huettel,

MELODIC (FSL’s version of ICA) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

An Example MELODIC ICA output FMRI – Week 10 – Analysis II Scott Huettel,

An Example MELODIC ICA output FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Limitations of ICA • Cannot test hypotheses • Provides no criterion for significance •

Limitations of ICA • Cannot test hypotheses • Provides no criterion for significance • Relies on interpretations drawn by the researchers FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Predicting Behavior and Thoughts FMRI – Week 10 – Analysis II Scott Huettel, Duke

Predicting Behavior and Thoughts FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Brain Pong The specific method of training varies from subject to subject. "But it

Brain Pong The specific method of training varies from subject to subject. "But it can be really different from one person to the next. We have one subject, a musician, who can vividly imagine the sight and sound of a concert, and that's a very specific brain region, " says Sorger. If the musician wants the ping-pong bat to move up the screen, he adds more and more musicians to his mental orchestra, increasing the intensity of his vision to a crescendo. To move the bat back down the screen, he clears his mind of such thoughts until the bat rests at the base of the screen. It is just like visualizing a volume control, says Goebel. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Real-Time f. MRI Caria et al. , (2007) FMRI – Week 10 – Analysis

Real-Time f. MRI Caria et al. , (2007) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Pattern Classification Kamitani & Tong, (2006) FMRI – Week 10 – Analysis II Scott

Pattern Classification Kamitani & Tong, (2006) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Simple Correlations Goal: To determine whethere are category-specific spatial maps. Procedure: Look at brain

Simple Correlations Goal: To determine whethere are category-specific spatial maps. Procedure: Look at brain response to category in odd runs, then see how well that pattern is replicated in even runs. Why is this a limited approach? Haxby et al. , (2001) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

(Logistic) Regression for Behavior Knutson et al. , (2007) FMRI – Week 10 –

(Logistic) Regression for Behavior Knutson et al. , (2007) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Support Vector Machines (SVM) SVM approaches train on a (large) set of voxels to

Support Vector Machines (SVM) SVM approaches train on a (large) set of voxels to develop a multidimensional classifier that predicts behavior. Cox & Savoy (2003) Norman et al (2006) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Key Steps of a Prediction Model • Voxel Selection: Whole brain or ROI? •

Key Steps of a Prediction Model • Voxel Selection: Whole brain or ROI? • Set aside part of your data • Train the classifier (on part of the data) – E. g. , runs 1 -5 (of 6) • Cross-validate the classifier – Test how well a classifier trained on 1 -4 predicts 5 • Determine an optimal classifier • Pray • Test that classifier on the omitted data FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Cox & Savoy (2003) FMRI – Week 10 – Analysis II Scott Huettel, Duke

Cox & Savoy (2003) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Using a large number of voxels in a classifier provides potentially very good predictive

Using a large number of voxels in a classifier provides potentially very good predictive power. The pattern of voxels provides much more information than ROI-based approaches. Cox & Savoy (2003) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Looking in visual cortex, Kamitani and Tong could predict the direction of moving dots

Looking in visual cortex, Kamitani and Tong could predict the direction of moving dots with great precision. Kamitani & Tong, (2006) FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Haynes et al. (2007) looked at medial prefrontal activation while people were making a

Haynes et al. (2007) looked at medial prefrontal activation while people were making a decision. They could predict, with >70% accuracy, what people would choose. FMRI – Week 10 – Analysis II Scott Huettel, Duke University

The pattern of voxel activation in just two brain regions (within the parietal cortex)

The pattern of voxel activation in just two brain regions (within the parietal cortex) can identify whether a person is making a decision under risk or under delay with>80% accuracy. Courtesy of John Clithero and Mc. Kell Carter FMRI – Week 10 – Analysis II Scott Huettel, Duke University

Data-Driven Analyses: Beautiful or Seductive? Data-Driven Hypothesis-Driven Non-Predictive FMRI – Week 10 – Analysis

Data-Driven Analyses: Beautiful or Seductive? Data-Driven Hypothesis-Driven Non-Predictive FMRI – Week 10 – Analysis II Predictive Standard GLM with GLM-Based Blinding, Split. Approaches Samples, etc. Exploratory Predictive Data-Driven Approaches Scott Huettel, Duke University