Frontotemporal networks for hierarchical predictive coding: Convergent evidence from dynamic causal modelling of human electrocorticography and magnetoencephalography

PHILLIPS HOLLY; BLENKMANN ALEJANDRO; HUGHES LAURA; BEKINSCHTEIN TRISTAN; KOCHEN SILVIA; ROWE JAMES

Washington DC

Conferencia; Annual Meeting Society for Neuroscience 2014; 2014

Society for Neuroscience

Unexpected sensory events engage automatic local and distributed processes of the mismatch negativity (MMN, Näätänen et al. 2007). Event related potentials (ERP) to standard and deviant auditorystimuli show difference waveforms between 100-250ms, from temporal and frontal cortical sources (Hughes et al. 2013). Neurocomputationalmodels (Lieder et al. 2013) and dynamic causal modelling (DCM) of the MMN in electroencephalography (EEG, Garrido et al. 2009)and magnetoencephalography (MEG, Hughes et al.2013) indicate a hierarchy of feedback sensory prediction and feedforward prediction errors,between primary auditory cortex (A1), superior temporal gyrus (STG) and prefrontal cortex (PFC). We posited that an internal pacemaker may provide conditional expectations to PFC that support its rolein predicting auditory events, and obtained evidence for this using MEG and direct human neurophysiological recordings. We used an auditoryMMN paradigm (Näätänen et al. 2004), alternating standard tones with deviant tones (differing by frequency, intensity, location, duration or a silent gap). We recorded MEG data in healthy adults and electrocorticography (ECoG) in patient candidates for epilepsy surgery, covering temporal and frontal cortices. ECoG electrode locations were normalised from patient inter-operative CT. MEG and ECoG recordings were processed using SPM8 with artefact removal using EEGLAB. We compared 12 dynamic causal models (SPM8-DCM10) of networks among A1, STG and PFC with neural mass models and mean fields fitted to MEG-sources (bilateral) andECoG-local field potentials (unilateral) during the first 250ms. Bayesian model selection used the free energy estimate F of log-model evidence to compare models and families of models with shared features.Results show that MEG and ECoG models had greater evidence with the features of: modulated frontotemporal feedforward and feedbackconnections and a pacemaker signal to PFC. Further, the deviant stimuli defined by temporal structure differences (duration and gap) had greatest evidence for network models including a pacemaker signal, whereas other deviants had greater evidence without it. This demonstrates different neural mechanisms are used, dependent on the particular violation to alearnt regularity. ECoG and MEG are complimentary methods that balance generalisation to larger populations (MEG) against precise anatomicallocalisation with direct recording of cortical field potentials (ECoG). They provide convergent evidence for the hierarchical interactions in frontotemporal networks and new evidence for an internal pacemaker influencing prefrontal cortex.