DCM: Advanced topics. Klaas Enno Stephan. SPM Course Zurich 06 February 2015

Transcription

1 DCM: Advanced topics Klaas Enno Stephan SPM Course Zurich 06 February 205

2 Overview DCM a generative model Evolution of DCM for fmri Bayesian model selection (BMS) Translational Neuromodeling

3 Generative model p( y, m) p( m) p( y, m). enforces mechanistic thinking: how could the data have been caused? 2. generate synthetic data (observations) by sampling from the prior can model explain certain phenomena at all? 3. inference about model structure: formal approach to disambiguating mechanisms p(y m) 4. inference about parameters p( y) 5. basis for predictions about interventions control theory

4 Bayesian system identification u(t) Design experimental inputs Neural dynamics dx dt f ( x, u, ) Define likelihood model Observer function y g( x, ) p( y, m) N( g( ), ( )) p(, m) N(, ) Specify priors Inference on model structure Inference on parameters p( y p( m) y, m) p( y, m) p( ) d p( y, m) p(, m) p( y m) Invert model Make inferences

5 Variational Bayes (VB) Idea: find an approximation q(θ) to the true posterior p θ y. Often done by assuming a particular form for q and then optimizing its sufficient statistics (fixed form VB). true posterior p θ y divergence KL q p best proxy q θ hypothesis class

6 Variational Bayes ln p(y) = KL[q p divergence 0 (unknown) + F q, y neg. free energy (easy to evaluate for a given q) ln p y KL[q p ln p y F q, y KL[q p F q, y is a functional wrt. the approximate posterior q θ. F q, y Maximizing F q, y is equivalent to: minimizing KL[q p tightening F q, y as a lower bound to the log model evidence When F q, y is maximized, q θ is our best estimate of the posterior. initialization convergence

7 Alternative optimisation schemes Global optimisation schemes for DCM: MCMC Gaussian processes Papers submitted in SPM soon

8 Overview DCM a generative model Evolution of DCM for fmri Bayesian model selection (BMS) Translational Neuromodeling

9 The evolution of DCM in SPM DCM is not one specific model, but a framework for Bayesian inversion of dynamic system models The implementation in SPM has been evolving over time, e.g. improvements of numerical routines (e.g., optimisation scheme) change in priors to cover new variants (e.g., stochastic DCMs) changes of hemodynamic model To enable replication of your results, you should ideally state which SPM version (release number) you are using when publishing papers.

10 Factorial structure of model specification Three dimensions of model specification: bilinear vs. nonlinear single-state vs. two-state (per region) deterministic vs. stochastic

11 y activity x (t) y activity x 2 (t) y activity x 3 (t) neuronal states BOLD y λ x hemodynamic model modulatory input u 2 (t) integration driving input u (t) The classical DCM: a deterministic, one-state, bilinear model t t Neural state equation endogenous connectivity modulation of connectivity direct inputs x ( j) ( A u B ) x Cu j x A x ( j) B u x C u j x x

12 bilinear DCM modulation non-linear DCM driving input driving input modulation Two-dimensional Taylor series (around x 0 =0, u 0 =0): dx f f f f x f ( x, u ) f ( x0,0) x u ux dt x u x u x 2 Bilinear state equation: Nonlinear state equation: dx dt A m i u B i ( i) x Cu dx dt m n ( i) A uib x i j j D ( j) x Cu

13 u Neural population activity u x x x fmri signal change (%) 0 0 Nonlinear Dynamic Causal Model for fmri dx dt A m n ( i) uib i j x j D ( j) x Cu Stephan et al. 2008, NeuroImage

14 u input Single-state DCM x Intrinsic (within-region) coupling Extrinsic (between-region) coupling N NN N N ij ij ij x x x ub A Cu x x Two-state DCM E x I N E N I E II NN IE NN EE NN EE NN EE N II IE EE N EI EE ij ij ij ij x x x x x ub A Cu x x ) exp( I x I E x x Two-state DCM Marreiros et al. 2008, NeuroImage

15 Stochastic DCM dx dt v ( A u B ) x Cv u ( v) j ( j) ( x) all states are represented in generalised coordinates of motion j random state fluctuations w (x) account for endogenous fluctuations, have unknown precision and smoothness two hyperparameters fluctuations w (v) induce uncertainty about how inputs influence neuronal activity can be fitted to "resting state" data Estimates of hidden causes and states (Generalised filtering) inputs or causes - V hidden states - neuronal hidden states - hemodynamic predicted BOLD signal excitatory signal flow volume dhb observed predicted - Li et al. 20, NeuroImage time (seconds)

16 Spectral DCM deterministic model that generates predicted crossed spectra in a distributed neuronal network or graph finds the effective connectivity among hidden neuronal states that best explains the observed functional connectivity among hemodynamic responses advantage: replaces an optimisation problem wrt. stochastic differential equations with a deterministic approach from linear systems theory computationally very efficient disadvantages: assumes stationarity Friston et al. 204, NeuroImage

17 cross-spectra = inverse Fourier transform of crosscorrelation cross-correlation = generalized form of correlation (at zero lag, this is the conventional measure of functional connectivity) Friston et al. 204, NeuroImage

18 Overview DCM a generative model Evolution of DCM for fmri Bayesian model selection (BMS) Translational Neuromodeling

19 Model comparison and selection Given competing hypotheses on structure & functional mechanisms of a system, which model is the best? Which model represents the best balance between model fit and model complexity? For which model m does p(y m) become maximal? Pitt & Miyung (2002) TICS

20 p(y m) Bayesian model selection (BMS) Model evidence: Gharamani, 2004 p( y m) p( y, m) p( m) d accounts for both accuracy and complexity of the model y all possible datasets allows for inference about model structure Various approximations, e.g.: - negative free energy, AIC, BIC McKay 992, Neural Comput. Penny et al. 2004a, NeuroImage

21 Approximations to the model evidence in DCM Logarithm is a monotonic function Maximizing log model evidence = Maximizing model evidence Log model evidence = balance between fit and complexity log p( y m) accuracy ( m) complexity( m) log p( y, m) complexity( m) No. of parameters Akaike Information Criterion: Bayesian Information Criterion: AIC BIC log p( y, m) log p( y, m) p p 2 log N No. of data points Penny et al. 2004a, NeuroImage

22 The (negative) free energy approximation F Neg. free energy is a lower bound on log model evidence: log p( y m) F KL q, p y, m ln p y m KL[q p F q, y Like AIC/BIC, F is an accuracy/complexity tradeoff: F log p( y, m) KL q, p m accuracy complexity

23 The complexity term in F In contrast to AIC & BIC, the complexity term of the negative free energy F accounts for parameter interdependencies. KL q( ), p( m) 2 ln C 2 ln C y 2 T C y y The complexity term of F is higher the more independent the prior parameters ( effective DFs) the more dependent the posterior parameters the more the posterior mean deviates from the prior mean

24 Bayes factors To compare two models, we could just compare their log evidences. But: the log evidence is just some number not very intuitive! A more intuitive interpretation of model comparisons is made possible by Bayes factors: B 2 p( y p( y Kass & Raftery classification: Kass & Raftery 995, J. Am. Stat. Assoc. m m 2 ) ) positive value, [0;[ B 2 p(m y) Evidence to % weak 3 to % positive 20 to % strong 50 99% Very strong

25 Fixed effects BMS at group level Group Bayes factor (GBF) for...k subjects: GBF ij BF ij k ( k ) Average Bayes factor (ABF): ABF ij K k BF ( k ) ij Problems: - blind with regard to group heterogeneity - sensitive to outliers

26 Random effects BMS for heterogeneous groups Dirichlet parameters = occurrences of models in the population r ~ Dir( r; ) Dirichlet distribution of model probabilities r m y k ~ p( mk p) mk ~ p( mk p) mk ~ p( mk p) m ~ Mult( m;, r ~ p( y m ) y ~ p( y m ) y2 ~ p( y2 m2 ) y ~ p( y m ) ) Multinomial distribution of model labels m Measured data y Model inversion by Variational Bayes (VB) or MCMC Stephan et al. 2009a, NeuroImage Penny et al. 200, PLoS Comp. Biol.

27 m 2 LD LD LVF m MOG FG FG MOG MOG FG FG MOG LD RVF LD LVF LD LD LG LG LG LG RVF stim. LD LVF stim. RVF stim. LD RVF LVF stim. Subjects m 2 m Data: Stephan et al. 2003, Science Models: Stephan et al. 2007, J. Neurosci Log model evidence differences

28 5 p(r >0.5 y) = m 2 p r r 99.7% 2 m 3 p(r y) r % r r 84.3% Stephan et al. 2009a, NeuroImage

29 When does an EP signal deviation from chance? EPs express our confidence that the posterior probabilities of models are different under the hypothesis H that models differ in probability: r k /K does not account for possibility "null hypothesis" H 0 : r k =/K Bayesian omnibus risk (BOR) of wrongly accepting H over H 0 : protected EP: Bayesian model averaging over H 0 and H : Rigoux et al. 204, NeuroImage

30 Overfitting at the level of models #models risk of overfitting solutions: regularisation: definition of model space = setting p(m) family-level BMS Bayesian model averaging (BMA) posterior model probability: p m y p y m p( m) p y m p( m) m BMA: p m y p y, m p m y

31 FFX RFX 80 log GBF Summed log evidence (rel. to RBML) nonlinear linear CBM N CBM N (ε) RBM N RBM N (ε) CBM L CBM L (ε) RBM L RBM L (ε) m 2 Model space partitioning: comparing model families p(r >0.5 y) = m alpha p(r y) p r r % 0 CBM N CBM N (ε) RBM N RBM N (ε) CBM L CBM L (ε) RBM L RBM L (ε) 2.5 Model space partitioning alpha m m 2 * 4 k k 0.5 r2 26.5% r 73.5% r nonlinear models * 2 8 k k5 linear models Stephan et al. 2009, NeuroImage

32 Bayesian Model Averaging (BMA) abandons dependence of parameter inference on a single model and takes into account model uncertainty represents a particularly useful alternative when none of the models (or model subspaces) considered clearly outperforms all others when comparing groups for which the optimal model differs single-subject BMA: p m y p y, m p m y group-level BMA: p m n y.. N p y, m p m y n n.. N NB: p(m y..n ) can be obtained by either FFX or RFX BMS Penny et al. 200, PLoS Comput. Biol.

33 Schmidt et al. 203, JAMA Psychiatry

34 Schmidt et al. 203, JAMA Psychiatry

35 Prefrontal-parietal connectivity during working memory in schizophrenia 7 ARMS, 2 first-episode (3 non-treated), 20 controls Schmidt et al. 203, JAMA Psychiatry

36 definition of model space inference on model structure or inference on model parameters? inference on individual models or model space partition? inference on parameters of an optimal model or parameters of all models? optimal model structure assumed to be identical across subjects? comparison of model families using FFX or RFX BMS optimal model structure assumed to be identical across subjects? BMA yes no yes no FFX BMS RFX BMS FFX BMS RFX BMS FFX analysis of parameter estimates (e.g. BPA) RFX analysis of parameter estimates (e.g. t-test, ANOVA) Stephan et al. 200, NeuroImage

37 Overview DCM a generative model Evolution of DCM for fmri Bayesian model selection (BMS) Translational Neuromodeling

38 Model-based predictions for single patients model structure DA BMS parameter estimates generative embedding

39 Synaesthesia projectors experience color externally colocalized with a presented grapheme associators report an internally evoked association across all subjects: no evidence for either model but BMS results map precisely onto projectors (bottom-up mechanisms) and associators (top-down) van Leeuwen et al. 20, J. Neurosci.

40 Generative embedding (unsupervised): detecting patient subgroups Brodersen et al. 204, NeuroImage: Clinical

41 Generative embedding: DCM and variational Gaussian Mixture Models Supervised: SVM classification Unsupervised: GMM clustering 7% number of clusters number of clusters 42 controls vs. 4 schizophrenic patients fmri data from working memory task (Deserno et al. 202, J. Neurosci) Brodersen et al. (204) NeuroImage: Clinical

42 Detecting subgroups of patients in schizophrenia Optimal cluster solution three distinct subgroups (total N=4) subgroups differ (p < 0.05) wrt. negative symptoms on the positive and negative symptom scale (PANSS) Brodersen et al. (204) NeuroImage: Clinical

43 A hierarchical model for unifying unsupervised generative embedding and empirical Bayes Raman et al., submitted

44 Dissecting spectrum disorders Differential diagnosis model validation (longitudinal patient studies) BMS Generative embedding (unsupervised) Computational assays Generative models of behaviour & brain activity BMS Generative embedding (supervised) optimized experimental paradigms (simple, robust, patient-friendly) initial model validation (basic science studies) Stephan & Mathys 204, Curr. Opin. Neurobiol.

45 Methods papers: DCM for fmri and BMS part Brodersen KH, Schofield TM, Leff AP, Ong CS, Lomakina EI, Buhmann JM, Stephan KE (20) Generative embedding for model-based classification of fmri data. PLoS Computational Biology 7: e Brodersen KH, Deserno L, Schlagenhauf F, Lin Z, Penny WD, Buhmann JM, Stephan KE (204) Dissecting psychiatric spectrum disorders by generative embedding. NeuroImage: Clinical 4: 98- Daunizeau J, David, O, Stephan KE (20) Dynamic Causal Modelling: A critical review of the biophysical and statistical foundations. NeuroImage 58: Daunizeau J, Stephan KE, Friston KJ (202) Stochastic Dynamic Causal Modelling of fmri data: Should we care about neural noise? NeuroImage 62: Friston KJ, Harrison L, Penny W (2003) Dynamic causal modelling. NeuroImage 9: Friston K, Stephan KE, Li B, Daunizeau J (200) Generalised filtering. Mathematical Problems in Engineering 200: Friston KJ, Li B, Daunizeau J, Stephan KE (20) Network discovery with DCM. NeuroImage 56: Friston K, Penny W (20) Post hoc Bayesian model selection. Neuroimage 56: Friston KJ, Kahan J, Biswal B, Razi A (204) A DCM for resting state fmri. Neuroimage 94: Kiebel SJ, Kloppel S, Weiskopf N, Friston KJ (2007) Dynamic causal modeling: a generative model of slice timing in fmri. NeuroImage 34: Li B, Daunizeau J, Stephan KE, Penny WD, Friston KJ (20). Stochastic DCM and generalised filtering. NeuroImage 58: Marreiros AC, Kiebel SJ, Friston KJ (2008) Dynamic causal modelling for fmri: a two-state model. NeuroImage 39: Penny WD, Stephan KE, Mechelli A, Friston KJ (2004a) Comparing dynamic causal models. NeuroImage 22:57-72.

46 Methods papers: DCM for fmri and BMS part 2 Penny WD, Stephan KE, Mechelli A, Friston KJ (2004b) Modelling functional integration: a comparison of structural equation and dynamic causal models. NeuroImage 23 Suppl :S Penny WD, Stephan KE, Daunizeau J, Joao M, Friston K, Schofield T, Leff AP (200) Comparing Families of Dynamic Causal Models. PLoS Computational Biology 6: e Penny WD (202) Comparing dynamic causal models using AIC, BIC and free energy. Neuroimage 59: Rigoux L, Stephan KE, Friston KJ, Daunizeau J (204). Bayesian model selection for group studies revisited. NeuroImage 84: Stephan KE, Harrison LM, Penny WD, Friston KJ (2004) Biophysical models of fmri responses. Curr Opin Neurobiol 4: Stephan KE, Weiskopf N, Drysdale PM, Robinson PA, Friston KJ (2007) Comparing hemodynamic models with DCM. NeuroImage 38: Stephan KE, Harrison LM, Kiebel SJ, David O, Penny WD, Friston KJ (2007) Dynamic causal models of neural system dynamics: current state and future extensions. J Biosci 32: Stephan KE, Weiskopf N, Drysdale PM, Robinson PA, Friston KJ (2007) Comparing hemodynamic models with DCM. NeuroImage 38: Stephan KE, Kasper L, Harrison LM, Daunizeau J, den Ouden HE, Breakspear M, Friston KJ (2008) Nonlinear dynamic causal models for fmri. NeuroImage 42: Stephan KE, Penny WD, Daunizeau J, Moran RJ, Friston KJ (2009a) Bayesian model selection for group studies. NeuroImage 46: Stephan KE, Tittgemeyer M, Knösche TR, Moran RJ, Friston KJ (2009b) Tractography-based priors for dynamic causal models. NeuroImage 47: Stephan KE, Penny WD, Moran RJ, den Ouden HEM, Daunizeau J, Friston KJ (200) Ten simple rules for Dynamic Causal Modelling. NeuroImage 49: Stephan KE, Mathys C (204). Computational approaches to psychiatry. Current Opinion in Neurobiology 25:

47 Thank you