Measuring Activation and Causality using multiple Prior Information Pedro A. Valdés-Sosa Cuban Neuroscience Center
Mapping
Brain Maps
Multimodal T1-Nissl-cryotomy-PET-myelin stain
Localization versus Connectivity Hiparcus ( II BC) Jackson US air Traffic Luria
Localization versus Connection AnatomicalPhysiological Localization Morphometry: Voxel based Region based Cortical thickness Activation: EEEG/MEG fMRI Connection Anatomical Connectivity Diffusion Weighted Functional Connectivity Effective Connectivity
Brain Tomographies Physical Model of some brain characteristic Prediction of measurement Direct Problem Image of some brain characteristic measurement Inverse Problem a Priori Information
EEG/MEG Forward Problem Primary current j EEG/MEG v
EEG inverse Problem Bayesian Inference!!
Methods for Regression Data VARETA LORETA ICA Non Negative Matrix Factorization In fact can be unified or combined
L0 Norm “Sparsness” AIC, BIC, TIC, RIC “subset selection”“Matching Purusit”“Dipoles”
L1 Norm “Sparseness” “Lasso”“Basis Pursuit”“FOCUSS” Connection with ICA
Fast LARS Algorithm (Friedman, Hastie, Tibshirani) Regularization path for diabetes data
L2 Norm “Minimum Norm” “Ridge”“Frames”“Minimum Norm”
Simplest EEG inverse Problem Bayesian Inference!!
Multiple Priors SparsenessMinimal Norm Non smoothDipoles=FOCUSSMinimum Norm SmoothVARETALORETA
Which inverse solution to choose?: let the data decide combining all solutions
Bayesian Model Averaging For 69 compartments
Simulations with Bayesian Model Averaging
BMA during concurrent EEG/fMRI
Combining Priors Fused Lasso VARETA-LORETA
Combining penalties (L1,I) (L1,L)
Between LORETA and VARETA LORETA VARETA Solution Chosen
Further Combination: Multiple Priors plus (semi) Non Negative Matrix Factorization Non Negative Matrix Factorizations used for data reduction Equivalent to Cluster Analysis
Multiple Priors plus (semi) Non Negative Matrix Factorization
Fast Non-negative LARS Algorithm (Morup) Regularization paths for diabetes data
Results for a Simulation 64 Channels, 1 Patch complex time series BIC Regularization path
Results of a Simulation
Localization versus Connection AnatomicalPhysiological Localization Morphometry: Voxel based Region based Cortical thickness Activation: EEEG/MEG fMRI Connection Anatomical Connectivity Diffusion Weighted Functional Connectivity Effective Connectivity
Effective vs. Functional Connectivity (Karl Friston)
Statistical Analysis of Causal Modeling "Beyond such discarded fundamentals as 'matter' and 'force' lies still another fetish amidst the inscrutable arcana of modern science, namely, the category of cause and effect.“ Karl Pearson (1911)
Granger (Non) Causality for TWO time series Granger Non Causality t t-1 t =1,…,N
Granger Causality of EEG signals Freiwald et al. (1999) J. Neurosci. Methods. 94: C3 C4 t t-1
What happens when you have a LOT of time series? 12…p12…p … t t-1 t =1,…,N Long history: Bressler, Baccala, Kaminski, Eichler, Goebel
Problems with the Multivariate Autoregressive Model for Brain Manifolds p→∞p→∞ # of parameters likelihood
Regions of Interest Alemán-Gómez Y. et al. PS0103
Point influence Measures is the simple test
38 Spike and Wave
39 Spike and Wave
What happens when you have a LOT of time series? 12…p12…p … t t-1 t =1,…,N Long history: Bressler, Baccala, Kaminski, Eichler, Goebel
a) Teat CG as a Random Field Concept applied to correlation fields by Worsley Usual SPM: RF is the brain New Idea RF is Cartesian product of Brain by Brain = = X
Granger Causality must be measured on a MANIFOLD
Influence Measures defined on a Manifold An influence field is a multiple test and all for a given
Discretization of the Continuos AR Model
Influence Fields and Bayesian Estimation Influence field likelihood prior
Influence Fields Outield Infield
Priors for Influence Fields maximal SMOOTHNESS Valdés-Sosa PA Neuroinformatics (2004) 2:1-12 Valdés-Sosa PA et al. Phil. Trans R. Soc. B (2005) 360: Minimum norm I Minimum spatial laplacian L prior
vs FFA Amigdala Fear Static + Fear Dynamic Neutral Neural basis of emotional expression processing
Emotional Network (Dipole)
Cuban Neuroscience Center Concurrent EEG-fMRI recordings Fine time scale
Cuban Neuroscience Center Concurrent EEG-fMRI ( Rhythm)
Basis of concurrent EEG/MEG-fMRI analysis-voxel level Trujillo et al. IJBEM (2001) BOLD Vasomotor Feed Forward Signal VFFS Ensemble of Postsynaptic Potentials ePSP net Primary Current Density nPCD EEG/MEG
EEG/MEG-fMRI-voxel Inverese solution Association BOLD VFF S ePSP nPCD EEG/MEG
correlationlog BOLD-log j
First order Autoregressive Model for fMRI and EEG
Estimated A for fMRI-EEG (f,s) using L1 regularizer
EEG-fMRI influence Fields Maximal Evidence dipole MN non smooth smooth nonsmooth+smooth dipole+MN
/
Localization versus Connection AnatomicalPhysiological Localization Morphometry: Voxel based Region based Cortical thickness Activation: EEEG/MEG fMRI Connection Anatomical Connectivity Diffusion Weighted Functional Connectivity Effective Connectivity