Data analyses

Behavioural analyses examined mean reaction time for correct Go and incorrect NoGo responses, and response accuracy (arsine transformed) using IBM SPSS Statistics 22.0^®. Reaction times and accuracy rates are presented in Table 2. Independent two-sample t-tests compared reaction times of the patients to controls, and Mann-Whitney U-tests were used to compare response accuracy (due to non-Gaussian distribution). Greenhouse-Geisser correction was used to correct for non-sphericity where necessary. Cohen’s d effect size is also reported. We supplement classical frequentist classical statistics with a Bayesian analysis of group differences using JASP software to test the hypothesis that patients are slower in reaction times and less accurate. Thresholds for interpretation are Bayes factors 3, 20 and 150 representing weak, strong and very strong evidence, respectively.

Mean reaction times and accuracy rates (arcsin transformed in radians, and non-transformed mean accuracy %) for Go (correct trials) and NoGo (commission errors) trials

Standard errors are in parentheses.

An index of clinical behavioural disinhibition was calculated from the CBI, including the sum of all items from the disinhibited, challenging, motor, eating and insight subscales, and the euphoria items from the mood subscale (Hughes et al., 2015). These specific types of behaviours have been previously shown to robustly quantify the syndrome of behavioural ‘disinhibition’ (Borroni et al., 2012). This index was used in correlations with the time-frequency data.

Time-frequency power spectra were computed for frequency bands between 4–80 Hz across the whole epoch using Morlet wavelets with a factor of 5. The transformed data were baseline corrected using a log ratio of power and scaled to dB. Only the accurate trials were used in the data analysis, as a measure of the physiological thresholds for movement (Go Trials) and withholding movement (NoGo trials). The number of accurate trials (but not inaccurate trials) was sufficient for good signal-to-noise ratio and power for MEG analyses.

Statistical analysis was performed on 2D images of frequency by time, averaging across the root mean squared value of gradiometer sensor pairs. These images were entered into a 2 × 2 ANOVA, to test the differences and interactions between the conditions and groups. The statistical maps were thresholded with a cluster-based family-wise error (FWE) correction P < 0.05 (after P < 0.001 voxel-wise height threshold).

DCM for MEG is an approach that explains the statistical dependencies between sources in terms of causal mechanisms, by inversion of generative models of brain networks to the neurophysiological observations. For the analysis of induced responses, DCM provides a phenomenological model of the time-dependent changes in spectral density. Importantly, this method captures how the frequency dynamics in one source affect the same or different frequency dynamics in another source, thus revealing both linear (within frequency) and non-linear (cross-frequency) couplings.

A detailed explanation of this method is described by Chen et al. (2008, 2009). In summary, DCM involves three main stages of analysis. The first step is the architectural specification of the neuronal network model, in which the neuronal sources of interest are identified and defined by MNI coordinates and the connections between sources varied to create a set of models to compare. The number of sources included in DCM is limited due to computations becoming intractable with large numbers. The second step is the inversion of the model to the observed data, and the time-frequency decomposition. The sources are modelled with equivalent current dipoles and then spectral density is calculated using a Morlet wavelet transform. For computational efficiency data are reduced to a number of modes from a singular value decomposition of the spectral power. The coupling dynamics between regions (i.e. time-dependent changes in spectral energy for each connection) are estimated using linear state equations. These estimates are represented in ‘A’ and ‘B’ matrices for each connection. If the model is specified as only ‘linear’, then each matrix will represent within frequency coupling, if the model is specified as ‘non-linear’, then each matrix will include all frequency couplings, including within and cross-frequency couplings. The ‘A’ matrix describes the coupling strength between the source and target frequencies, dependent on exogenous inputs, for all trials (i.e. the Go and NoGo accurate trials). The ‘B’ matrix describes the coupling strength between the source and target frequencies, dependent on the experimental manipulation (i.e. the NoGo versus the Go trials). The last step in the DCM is to identify the optimal model that best supports the observed data using Bayesian model selection based on free-energy estimates of the log model evidence.

In this study, the model architecture included three predefined regions based on the well-described response inhibition network: the left motor cortex, preSMA, and right IFG (Montreal Neurological Institute, MNI coordinates: −37 −25 64; −4 4 60; 48 18 −2). The coordinates for preSMA and motor cortex were based on a functional MRI meta-analysis of right hand motor control (Mayka et al., 2006). The right IFG coordinate was based on a recent NoGo inhibition study (Ye et al., 2014).

The set of models designed included two families defined by the possibility for linear only or combined linear and non-linear coupling. Within each family, seven model architectures were examined to test alternate hypotheses about the contribution of power couplings between regions. All models assumed reciprocal coupling between the three regions for the exogenous task-related perturbations. For the condition-dependent perturbations the model space was varied, allowing either: all reciprocal connections, two sets of connections, or just one set of connections to be modulated by condition. All self-connections were modulated in all models. The cortical driving input was applied to the sources in the IFG and preSMA (Rae et al., 2015).

The DCM included only the accurate Go and NoGo trials and the gradiometer channels for the dipole model fit. The time-frequency decomposition used similar processes to the sensor space analysis: Morlet wavelets with a factor of 5 were computed across a frequency bandwidth 4–80 Hz. The data were reduced to four frequency modes obtained from a singular value decomposition of the spectral power. In our dataset these four modes explained 96% of the variance in both groups [standard deviation (SD): Controls = 2.3%; bvFTD = 9.7%]. The models predicted frequency dynamics across a window of −50 to 1200 ms to include the ERD and ERS. Onset priors, reflecting stimulus input, were set to the default 60 ms post-stimulus presentation with a standard deviation of 16 ms, adjusted during the process of model fitting with default variance. The conditional modulation of the model (as represented by the ‘B’ matrix) compared accurate NoGo versus accurate Go trials.

Bayesian model selection was used to identify the model with the best fit to the data using free-energy estimates of the log model evidence. A two-step procedure was used. First, to identify whether cross-frequency coupling was an important feature, the seven linear and seven non-linear models were compared using a family-based model comparison procedure (Penny et al., 2010). Second, the seven individual models within the winning family were compared to identify the best model architecture. The subject-specific frequency-frequency parameter estimations of the winning model were entered into two flexible factorial ANOVAs, one for the connection parameters of all trials (A matrix) and one for the parameters of NoGo versus Go trials (B matrix). The effects of interest were the interactions between group and connection (two groups × nine connections), and also the difference between controls and patients.

A T₁-weighted structural image (magnetization prepared rapid acquisition gradient echo, MPRAGE) was obtained from each subject (repetition time 2250 ms, echo time 2.99 ms, flip angle 9°, inversion time 900 ms, 256 × 256 × 192 isotropic 1 mm voxels) to co-register the MEG data and to enable subject-specific modelling of the lead field for the DCM analyses. These images were also included in the voxel-based morphometry (VBM).

The VBM analysis used SPM 12 (www.fil.ion.ucl.ac.uk/spm) and the DARTEL toolbox (Ashburner, 2007), and followed the steps suggested by Ashburner (2007). The T₁ images for each participant were segmented into grey, white, and CSF tissue classes and together used to create a study-specific group template to improve intersubject alignment during normalization. The template was registered to MNI space and used to generate Jacobian scaled modulated grey and white matter images from each subject that were spatially normalized to MNI space and smoothed with an 8-mm full-width at half-maximum (kernel.

Two general linear models were used to examine the differences in the grey and white matter images between patients and controls. For each general linear model, total intracranial volumes from each subject were included as nuisance covariates to correct for intersubject differences in global brain volume. Age was also included as a covariate. Statistical maps were thresholded with a cluster-based familywise error correction P < 0.05 (after P < 0.001 voxelwise uncorrected threshold). A Bayesian estimation of the same contrast was performed and then subjected to a null hypothesis test, providing a statistical map of the posterior probability at each voxel. Voxels considered significant exceeded 95% confidence threshold and had a volume threshold greater than 0.7%. These voxels represented regions that had strong evidence for normal cortical volume.