400 likes | 673 Views
PPI. Friston (1997) Gitelman (2003). Basic fMRI refreshments. Friston et al. 1997. Friston 1997. Introduction Aim: define PPI Address interpretation Basic idea: Correlation between areas changes as context changes. Effective connectivety efficasy and contributions.
E N D
PPI Friston (1997) Gitelman (2003)
Friston et al 1997
Friston 1997 • Introduction • Aim: • define PPI • Address interpretation • Basic idea: • Correlation between areas changes as context changes.
Effective connectivetyefficasy and contributions • Functional specialization • Functional integration • Functional connectivity • (correlation) • Effective connectivity • (taking into account full model)
Effective connectivityefficacy and contributions Test on : H0: ik=0 i.e., test correlation between regions Note if more regions, towards effective connectivity.
Factorial designs and Psychological interactions • Imagine 1 task (gr), two conditions (ga) Note gr and ga are mean corrected
Physiological interaction • Imagine 2 areas (gr, and ga) ga BRAIN gr
Physiological interaction • Example in paper: • gr=PP • ga=V1 • Responding area: V5 • Note this is interaction and not only due to PP, PP activity is a confound
Non linear models SKIP
Psychophysiological interaction • xk : source region (V1) • gp : task (-1 or +1 label)
xk : V1 gp : task (attention)
fit V1 V5 V1 V5 attention No attention
Once more be aware ? V1 V1 V5 V5
Summary • Psychophysiological interaction • Predict activity in area B by area A as a function of context • PPIeffective connectivity • PPI=contribution (c.f. correlation) • Note on interpretations. • Connection AB influenced by task • Influence TaskB is modified by activity in A • No guarantee that connections are direct.
Gitelman et al 2003 (where Friston went “wrong”)
Aim • Show importance of deconvolution • How to deconvolve properly
Introduction • Don’t analyze interactions on raw BOLD signal. (using SEM PPI etc) • “veridical models of neuronal interactions require the neural signal or at least a well-constrained approximation to it. “
A simulation (see examples) Time shift (0-8 s)
Deconvolve A&B Interaction Reconvolve
Conclusion Interaction with the convolved signal Interaction at neural representation + convolution
conclusion • Noise has more effect on HRF interactions • Deconvolution reduces noise
Conclusion • There is an effect for event related designs. • Not so strong as simulations.
Conclusion • Effect on BLOCK design data is not dramatic. • In short: • Calculating interactions at neural representation pays especially for ER designs. • Friston was wrong, but not that far off because of block design in his experiment.
Theoryinteraction on the convoluted signal(i.e. BOLD signal)
How to obtain xA from yA NOTE 112 columns Basis set
How to obtain xA from yA X has too many columns over determined matrix not one unique solution
Solution • Biased estimation. (bayesian stat.) • I start to get lost…..
What I do understand • High frequencies are a problem in deconvolution. • Convolution is low pass filter. high frequency information is losthigh frequency estimates are unstable/unreliable. • High frequencies were also the most troubling in interactions based on BOLD signal (cf ER & BLOCK designs) • High frequencies are regularized using bayesian stat.