DCM for evoked responses Ryszard Auksztulewicz SPM for M/EEG course, 2016

? Does network XYZ explain my data better than network XY? Which XYZ connectivity structure best explains my data? Are X & Y linked in a bottom-up, top-down or recurrent fashion? Is my effect driven by extrinsic or intrinsic connections? Which neural populations are affected by contextual factors? Which connections determine observed frequency coupling? How changing a connection/parameter would influence data? context input

The DCM analysis pathway Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

The DCM analysis pathway Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

Phillips et al., 2016

Data for DCM for ERPs / ERFs Downsample Filter (e.g. 1-40Hz) Epoch Remove artefacts Average Per subject Grand average Plausible sources Literature / a priori Dipole fitting / 3D source reconstruction Data should be processed like this to leave ERPs for each condition

The DCM analysis pathway Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

The DCM analysis pathway ‘Hardwired’ model features Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

Models I’m going to talk about these models here

Neuronal (source) model   Kiebel et al., 2008

NEURAL MASS MODEL L2/3 Inhib Inter Spiny Stell L4 Pyr L5/6 spm_fx_erp

Canonical Microcircuit Model (‘CMC’) Bastos et al. (2012) Pinotsis et al. (2012)

CANONICAL MICROCIRCUIT NEURAL MASS MODEL CANONICAL MICROCIRCUIT Pyr L2/3 Inhib Inter xv Spiny Stell Spiny Stell mv L4 Inhib Inter Pyr Pyr L5/6 spm_fx_erp spm_fx_cmc

Canonical Microcircuit Model (‘CMC’) Supra- granular Layer Granular Layer Infra-granular Layer

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Inhibitory Interneurons Superficial Pyramidal Cells Supra-granular Layer Spiny Stellate Cells Granular Layer Infra-granular Layer Deep Pyramidal Cells Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Voltage change rate: f(current) Current change rate: f(voltage,current) Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Voltage change rate: f(current) Current change rate: f(voltage,current) H, τ Kernels: pre-synaptic inputs -> post-synaptic membrane potentials [ H: max PSP; τ: rate constant ] S Sigmoid operator: PSP -> firing rate David et al., 2006; Pinotsis et al., 2012

Canonical Microcircuit Model (‘CMC’) Supra-granular Layer Granular Layer Infra-granular Layer Pinotsis et al., 2012

The DCM analysis pathway ‘Hardwired’ model features Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

5 4 3 2 1

5 4 3 2 1 Input

5 4 3 2 1 Input

5 4 3 2 1 Input

5 4 3 2 1 Input

Factor 1 5 4 3 2 1 Input

Factor 1 Factor 2 5 4 3 2 1 Input

The DCM analysis pathway Fixed parameters Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

Fitting DCMs to data In theory this should take 1s of work and 45mins drinking coffee.

Fitting DCMs to data H. Brown You should end up with something like this. This is nicely fitted data because….. H. Brown

Fitting DCMs to data H. Brown But sometimes this happens. This is poorly fitted data because… H. Brown

Fitting DCMs to data Check your data H. Brown Do you have a clear ERP which differs between conditions? If not, check your artefact rejection and other sources of noise in the data. H. Brown

Fitting DCMs to data Check your data Check your sources H. Brown Are they in locations of greatest source activity as identified by source localisation? Does the fit improve a lot if you allow the source locatiosn to be optimised? H. Brown

Fitting DCMs to data Check your data Check your sources Model 1 V4 IPL A19 OFC Fitting DCMs to data Check your data Check your sources Check your model e.g. does the model have enough explanatory power or enough areas? What about the input? IPL IPL V4 V4 Model 2 H. Brown

Fitting DCMs to data Check your data Check your sources Check your model Re-run model fitting Sometimes you know the model is sound but a few datasets are being stubborn. Might be due to the free energy landscape. If you’re fitting individual subject data, try re-initialising with the posteriors form a previous subject. H. Brown

The DCM analysis pathway Fixed parameters Build model(s) Fit your model parameters to the data Pick the best model Make an inference (conclusion) Collect data

Phillips et al., 2016

? Does network XYZ explain my data better than network XY? Which XYZ connectivity structure best explains my data? Are X & Y linked in a bottom-up, top-down or recurrent fashion? Is my effect driven by extrinsic or intrinsic connections? Which connections/populations are affected by contextual factors? context input

Example #1: Architecture of MMN Garrido et al., 2008

Example #2: Role of feedback connections Garrido et al., 2007

Example #3: Group differences Boly et al., 2011

Example #4: Factorial design & CMC Attention cf. Feldman & Friston, 2010 FORWARD PREDICTION ERROR L2/3 p x x xx L4 s m mx A1 STG L5/6 BACKWARD PREDICTIONS Bastos et al., Neuron 2012 Auksztulewicz & Friston, 2015

2x2 design: Attended vs unattended Standard vs deviant (Only trials with 2 tones) N=20 Auksztulewicz & Friston, 2015

Attention Expectation Flexible factorial design Thresholded at p<.005 peak-level Corrected at a cluster-level pFWE<.05 Auksztulewicz & Friston, 2015 Contrast estimate A1E1 A1E0 A0E1 A0E0

Auksztulewicz & Friston, 2015 Flexible factorial design Thresholded at p<.005 peak-level Corrected at a cluster-level pFWE<.05 Contrast estimate A1E1 A1E0 A0E1 A0E0 Auksztulewicz & Friston, 2015

Connectivity structure input input Extrinsic modulation input input Intrinsic modulation SP Inh Int input input

xv mv xx mx Superficial pyramidal cells: signal prediction errors to higher areas Brown & Friston, 2010 Feldman & Friston, 2013 Inhibitory interneurons: prediction errors of hidden states Bastos et al., 2012 attentional effects of ACh -> depression of I.I. activity Xiang et al., 1998 Buia and Tiesinga, 2006 attentional gamma spike-LFP phase locking Vinck et al., 2013 xv mv xx mx SPM CMC

Auksztulewicz & Friston, 2015

Motivate your assumptions! Rubbish data Perfect model Rubbish results Perfect data Rubbish model Rubbish results

Thank you! Karl Friston Gareth Barnes Andre Bastos Harriet Brown Jean Daunizeau Marta Garrido Stefan Kiebel Vladimir Litvak Rosalyn Moran Will Penny Dimitris Pinotsis Bernadette van Wijk