Abstract
We present a didactic introduction to spectral dynamic causal modeling (DCM), a Bayesian state-space modeling approach used to infer effective connectivity from noninvasive neuroimaging data. Spectral DCM is currently the most widely applied DCM variant for resting-state functional MRI analysis. Our aim is to explain its technical foundations to an audience with limited expertise in state-space modeling and spectral data analysis. Particular attention will be paid to cross-spectral density, which is the most distinctive feature of spectral DCM and is closely related to functional connectivity, as measured by (zero-lag) Pearson correlations. In fact, the model parameters estimated by spectral DCM are those that best reproduce the cross-correlations between all measurements—at all time lags—including the zero-lag correlations that are usually interpreted as functional connectivity. We derive the functional connectivity matrix from the model equations and show how changing a single effective connectivity parameter can affect all pairwise correlations. To complicate matters, the pairs of brain regions showing the largest changes in functional connectivity do not necessarily coincide with those presenting the largest changes in effective connectivity. We discuss the implications and conclude with a comprehensive summary of the assumptions and limitations of spectral DCM.
INTRODUCTION
Dynamic causal modeling (DCM) refers to the Bayesian fitting of state-space models to explain observed physiological signals in terms of hidden neuronal activity and connectivity (K. Friston, Harrison, & Penny, 2003; Triantafyllopoulos, 2021). The distinction between observed and hidden variables is particularly relevant in neuroscience because the signals recorded via noninvasive neuroimaging are not a direct measurement of neuronal states or connectivity. In fact, using observed recordings to infer unobserved neural interactions is the main purpose of DCM and the reason for its widespread adoption. It is also a distinctive feature that sets it apart from functional connectivity analysis, which simply characterizes statistical dependencies in observed time series. Over the last 20 years, the versatility offered by state-space models has seen DCM applications in most neuroimaging modalities (Frässle et al., 2021; K. Friston et al., 2003, 2019; K. J. Friston, Kahan, Biswal, & Razi, 2014; Jung, Kang, Chung, & Park, 2019; Kiebel, Garrido, Moran, & Friston, 2008; Moran et al., 2009; Tak, Kempny, Friston, Leff, & Penny, 2015), along with recent applications to epidemiology (K. J. Friston, Flandin, & Razi, 2022) and beyond (Bach, Daunizeau, Friston, & Dolan, 2010). Navigating the vast and technical DCM literature, however, is by no means a trivial task—especially to the novice learner. Happily, there are excellent introductory resources on individual- and group-level analysis using deterministic versions of DCM, which are designed for neuroimaging experiments involving behavioral tasks (Stephan, 2004; Stephan et al., 2010; Zeidman, Jafarian, Corbin, et al., 2019; Zeidman, Jafarian, Seghier, et al., 2019). A recent primer on variational Laplace explains how Bayesian inference is performed in DCM (Zeidman, Friston, & Parr, 2023) using the SPM software (https://www.fil.ion.ucl.ac.uk/spm). However, there is a lack of introductory material on DCM for resting-state data analysis, despite the remarkable growth of the resting-state paradigm and the widespread uptake of these methods. Here, we fill this gap with a didactic introduction to spectral DCM that aims to explain its technical aspects (K. J. Friston et al., 2014; Razi, Kahan, Rees, & Friston, 2015).
What distinguishes spectral DCM from other DCM versions, and when should we choose it? Firstly, spectral DCM employs random differential equations instead of deterministic ones. These are used to model spontaneous endogenous fluctuations in neuronal activity, enabling resting-state analysis in the absence of experimental inputs. But its distinctive feature is the focus on modeling the measured cross-spectral density, which is a second-order summary statistic of the time series data. This is closely related to Pearson’s correlation, another second-order statistic and the most widely used measure of functional connectivity in neuroimaging (see diagram in Figure 1; for mathematical relationships, also see K. J. Friston et al., 2014, Figure 1). In fact, the correlation is obtained by normalizing the covariance such that its values are restricted to the [−1, 1] interval. In turn, the covariance is a special case of the cross-covariance function between two time series, when there is no time lag between them. Finally, the Fourier transform of the cross-covariance function gives the cross-spectral density (under stationarity assumptions). In other words, the cross-spectral density is the equivalent representation of the cross-covariance function in the frequency domain instead of the time domain—an important relationship that we will unpack later in this article.
Spectral DCM fits the parameters of a linear, continuous-time model to the observed cross-spectral density. The estimated parameters are those that best reproduce the cross-correlations between all variables, at all time lags. In particular, the estimated effective connectivity also reproduces the zero-lag correlations between the observed time series—the most common measure of functional connectivity in the literature. This would be appealing to researchers who are interested in both effective and functional connectivity. The nuanced relationship between effective and functional connectivity is explored in the Effective and Functional Connectivity section. Prior to that, we introduce and explain the various components of the generative model, that is, the model that generates the cross-spectral density given a set of parameters. These basic building blocks are used routinely in signal processing and control theory and are often presented only briefly in the DCM literature. Here, we adopt an inclusive and slower pace for those who are not familiar with state-space models and spectral data analysis. That said, we count on the reader to fill in the gaps and look up concepts such as the Fourier transform (Oppenheim, Willsky, & Nawab, 1997; Smith, 2002) or convolution (Smith, 2002), if needed. Moving from theory to practice, a step-by-step guide to running spectral DCM on a real resting-state functional MRI (fMRI) dataset is provided in Chapter 38 of the SPM12 manual (Ashburner et al., 2020).
A final reason for choosing spectral DCM is its computational advantage compared with stochastic DCM (Li et al., 2011). It is important to note that the lower computational complexity and the resulting increase in speed rely on the assumption that the statistics of endogenous neuronal fluctuations are conserved over the experimental time window, making spectral DCM suitable for resting-state neuroimaging experiments. Experimental inputs can also be included via an additional term in the model, although applications to task experiments are infrequent in the literature. Introducing even stronger assumptions leads to even faster schemes, such as regression DCM, which can analyze hundreds of brain regions in minutes (Frässle et al., 2021). However, this method forgoes the strict separation between hidden and observed variables that is typical of state-space modeling and that we have used to define DCM herein. As the name suggests, regression DCM is more akin to the Bayesian fitting of a multivariate autoregressive model in the frequency domain.
The key assumptions made in spectral DCM are summarized in the Assumptions and Limitations section.
BUILDING THE GENERATIVE MODEL, ONE ELEMENT AT A TIME
The signals recorded via noninvasive neuroimaging are not a direct measurement of neuronal activity. In the case of fMRI, the observed blood-oxygen-level-dependent (BOLD) signal captures changes in blood oxygenation that indirectly reflect neuronal activity. For this simple reason, spectral DCM models the neuronal and the observed variables separately (denoted by x and y, respectively). Such a distinction represents both the main strength and the challenge of the DCM framework.
Neuronal Model
Effective Connectivity
(Note: The reason why DCM studies often report positive values on the diagonal of the effective connectivity matrix A is that the self-connections are transformed using the logarithmic function log(−2a) by SPM. This convention is technically motivated by the use of log-normal priors to enforce positivity or negativity on certain parameters; here, to enforce recurrent or self inhibition. A reported zero value for a self-connection corresponds to −0.5 Hz, which is the default prior self-connectivity value in SPM12. Negative values correspond to slower decay rates in the (−0.5, 0) range, while positive values correspond to faster decays, that is, <−0.5 Hz.)
Power Spectral Density and Cross-Spectral Density
Admittedly, the cross-spectral density definition as a covariance in the frequency domain is quite abstract. A better intuition will develop after exploring the close relationship between cross-spectral density and functional connectivity. In the Cross-Spectral Density and Functional Connectivity section, we’ll see how the cross-spectral density is the Fourier transform of the cross-covariance function, which captures both the correlation matrix and its time-lagged extensions. For now, the power spectral density definition given above is sufficient to understand how endogenous fluctuations are modeled in spectral DCM.
Endogenous Fluctuations
Observation Function
Putting It All Together
We have now fully described the forward model used in spectral DCM. The predicted cross-spectral density can be computed using the following model parameters:
- (a)
the effective connectivity parameters in the A matrix;
- (b)
the power-law parameters (i.e., the amplitude and the exponent) describing the spectrum of the endogenous fluctuations and the observation noise (Equations 11 and 14); and
- (c)
the observation function parameters, such as the biophysical parameters of the BOLD balloon model.
Crucially, neither the neuronal state variables X(ω) nor the endogenous fluctuations V(ω) appear in Equation 22, only the parameters describing their cross-spectral densities. This parameterization allows spectral DCM to infer the model parameters listed above without inferring the hidden neuronal states. Inferring the state variables (neuronal time series) is a computationally harder problem addressed by stochastic DCM (Li et al., 2011).
Simulated and Empirical Cross-Spectral Density
As before, the strength of the directed connection from x1 to x2 is set via the parameter a21. The third neuronal state variable receives an inhibitory influence from x1 and an excitatory influence from x2. All three state variables have the same negative self-connections represented by the diagonal elements of A. The resulting cross-spectral density is plotted in Figure 5, assuming the same canonical HRF defined in Equation 25 and the same power-law parameters as in Equation 24. The figure shows only the special case where a21 = 0, but different values of a21 would generate different sets of nine plots.
It is instructive to qualitatively compare the simulated cross-spectral density plots in Figure 5 with the empirical plots in Figure 6, obtained by fitting the spectral DCM model to a real resting-state dataset (Razi et al., 2015). Note that the empirical plots correspond to a system with four neuronal variables instead of three. Here, a four-region default mode network is modeled with posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), and bilateral inferior parietal cortices (IPC) as its nodes. We used a fully connected model where each state variable (or node) is connected to every other state variable. The figure is reproduced from Chapter 38 of the SPM12 manual (Ashburner et al., 2020), which provides the link to the data and a step-by-step tutorial to replicate the DCM specification and estimation results using the SPM graphical interface. Despite the clear differences due to different data and parameter settings, the empirical cross-spectral density plots also feature a single large peak at low frequencies, followed by a decay at larger frequencies (with smaller fluctuations). If the data are too noisy, owing to head motion and various physiological artifacts, additional large peaks may appear at higher frequencies. These empirical plots can be visualized using the review function (spm_dcm_fmri_csd_results()) in SPM. Another DCM diagnostics function (spm_dcm_fmri_check()) in SPM also reports the percentage of variance explained (R2), which is a useful performance metric to judge the quality and success of the model fit to the data (see Figure S2 in the Supporting Information).
EFFECTIVE AND FUNCTIONAL CONNECTIVITY
This shows that a local variation in a single effective connectivity parameter in the A matrix can have a global impact across all functional connectivity values in the network. Let’s examine each pair to unpack some of the many nuances involved. First, the symmetric nature of the correlation matrix requires that ρkj = ρjk, even if akj ≠ ajk. This is a key difference between functional and effective connectivity: Only the latter is directed and is able to differentiate between two bilateral connections.
Both correlations ρ21 and ρ31 increase monotonically with a21, but this is not the case for ρ32. Therefore, there is no one-to-one mapping between effective and functional connectivity that holds in general (Park & Friston, 2013). This poses a challenge for spectral DCM because it makes model fitting an ill-posed problem with multiple potential solutions. Technically, this problem is mitigated by using the cross-spectral density that implicitly contains information about functional connectivity over all lags (we will unpack this below) and by using priors on the solutions implicit in the functional form of the DCM. Although model fitting remains an ill-posed problem, these two additional constraints allow spectral DCM to find better solutions, such that the model can reproduce a larger set of statistical relationships between the observed time series. One way to appreciate the amount of additional information and constraints provided by the cross-spectral density over the zero-lag correlation is to inspect Figure 5 again. Zero-lag correlation measures only the area under the curves in the nine plots, regardless of their detailed shapes. On the other hand, spectral DCM fits all values taken by the cross-spectral density curves at different frequencies. This requirement narrows the parameter space in a useful way (note that the area under the curve is still reproduced as a consequence).
The sign of ρ31 (representing a positive or a negative correlation) either matches or contradicts the sign of the underlying effective connectivity a31 (representing an excitatory or an inhibitory connectivity), depending on how we set a21. There is a value of a21 such that ρ31 = 0, even though the underlying effective connectivity a31 is negative (i.e., inhibitory). Again, this shows that each pairwise functional connectivity value is a summary statistic (global property) of the system: Given a pair of variables j and k, the correlation value ρkj depends not only on the corresponding effective connectivity parameter akj but also, potentially, on all the entries of the A matrix.
While here we have discussed its dependence on only effective connectivity, functional connectivity also depends on the parameters characterizing the endogenous fluctuations, the observation function, and the observation noise. This is because all these parameters appear in the forward model for the cross-spectral density derived in Equation 22 and, in turn, determine the correlation matrix of the system. In fMRI, specific and reproducible spatial patterns of functional connectivity could simply arise from specific and reproducible variations in the hemodynamic response across brain regions—even in the absence of interregional effective connectivity (Rangaprakash, Wu, Marinazzo, Hu, & Deshpande, 2018). When comparing groups, differences in BOLD functional connectivity may reflect differences in the vasculature rather than in effective connectivity (or both), for example because of aging or a neurodegenerative or psychiatric disease (Tsvetanov, Henson, & Rowe, 2020). This important point is further discussed in K. J. Friston (2011), where it is shown that the correlation values are also influenced by different levels of observation noise:
One can see a change in correlation by simply changing the signal-to-noise ratio of the data. This can be particularly important when comparing correlations between different groups of subjects. For example, obsessive compulsive patients may have a heart rate variability that differs from normal subjects. This may change the noise in observed hemodynamic responses, even in the absence of neuronal differences or changes in effective connectivity. (p. 22)
In fact, separating these confounding factors from the effective connectivity was one of the main motivations for the development of DCM. This is not to say that DCM is without issues: Fitting a large model with many parameters—using the limited amount of data available in typical fMRI studies—is not guaranteed to produce optimal estimates. We’ll discuss this and other limitations in the Assumptions and Limitations section.
On the other hand, functional connectivity has proven valuable for fingerprinting, that is, identifying an individual based on their brain activity (Finn et al., 2015). It is also useful in studies aiming to differentiate between two groups or conditions (by detecting statistically significant changes in correlation between observed time series), rather than in identifying which effective connections between brain regions underlie that change. Moreover, being a global property of the system, each pairwise functional connectivity value naturally captures higher order interactions, which is a topic of growing interest in complex systems, network science, and neuroimaging (Benson, Gleich, & Leskovec, 2016; Rosas et al., 2022).
So far, we have examined only the zero-lag correlation matrix because it is widely used to quantify the functional connectivity in neuroimaging. However, spectral DCM doesn’t explain just the zero-lag correlation but also the cross-correlations between all variables at all time lags. To understand this point, in the next section, we will invoke the elegant Wiener-Khinchin theorem, which links the cross-covariance function to the cross-spectral density.
Autocovariance and Cross-Covariance Functions
Let’s briefly revisit the role of the self-connections in the presence of the endogenous fluctuations. As discussed in the Effective Connectivity section, a self-connection determines the rate of decay of a variable. It can also be understood as quantifying the memory of a variable, that is, whether inputs have a short-lived or long-lasting impact on its activity. Large negative self-connections reflect short memory: The variable “forgets” its past quickly and responds promptly to any new inputs. On the contrary, small negative values reflect long memory: The neuronal variable integrates and smooths out the endogenous fluctuations and other inputs, resulting in slower oscillations and lower frequencies.
Cross-Spectral Density and Functional Connectivity
ASSUMPTIONS AND LIMITATIONS
State-Space Modeling
The foundational hypothesis of the DCM framework is that the system of interest can be modeled using a state-space approach. This formulation separates the equations describing the temporal evolution of unobserved variables (e.g., the neuronal activity) from those describing the observed variables (e.g., the BOLD). Such a separation is important because some of the model parameters have a direct interpretation in terms of effective connectivity between unobserved neuronal populations rather than statistical dependency between observed variables. A subtle and important consequence of the distinction between functional effective connectivity is that one can only estimate recurrent or self-inhibition using a state-space model. This is because the correlation of a time series with itself is always 1 (and the variance is always positive). Assessing self-connectivity in terms of excitability or disinhibition of a node can be empirically important (e.g., in estimating changes in excitation-inhibition balance or condition-specific changes in the “gain” of a particular source or region).
Continuous-Time Formulation
Like most DCM approaches, spectral DCM treats time as a continuous quantity. Representing time as a sequence of discrete steps may seem more natural, especially for fMRI, where the observations are recorded at evenly spaced time intervals with a relatively low sampling rate. However, DCM also models the neuronal activity, which unfolds at a much faster timescale and in an asynchronous manner. These two features can be naturally modeled using differential equations in continuous time. The discrete formulation usually converges to the continuous one with faster acquisition times, as is the case for EEG and MEG.
Separation of Timescales and Macroscopic Modeling
Spectral DCM assumes that a single macroscopic neuronal variable can capture the essential dynamical properties of a population of neurons. This is not just a pragmatic way to avoid billions of equations representing individual neurons. It is an approach based on the separation of timescales that has a long history in dynamical systems theory (Carr, 1981; Haken, 1983). The assumption is that the neuronal activity can be separated into slow and fast modes: The fast modes decay quickly so that the long-term behavior of the system can be described by a few slow modes, or even a single one. Mathematically, the assumption (often satisfied in real systems) is that only a small number of eigenvalues of the Jacobian are near zero, while the rest are large and negative (K. J. Friston et al., 2021; K. J. Friston, Li, Daunizeaue, & Stephan, 2011). Macroscopic modeling also relies on the mean-field assumption that the dynamics of one region are determined by the mean activity in another (Deco, Jirsa, Robinson, Breakspear, & Friston, 2008). Since the mean activity is dominated by the slow modes, it is possible to build a compact macroscopic model where only the slow modes are communicated among brain regions, whereas the fast endogenous fluctuations affect only the local activity. This is precisely the neuronal model in Equation 1. It is important to note that, despite being grounded in dynamical and complex systems theory, this model is an abstraction. Biological details are necessarily omitted to enable the analytic treatment and faster numerical computations. A detailed critical review of the biophysical and statistical foundations of DCM is provided in Daunizeau, David, and Stephan (2011).
Stationarity
Among the DCM variants, spectral DCM has the most direct conceptual and analytical links to functional connectivity, which we have examined in the Effective and Functional Connectivity section. Both methods can be read as assuming that the observed processes are weakly stationary, that is, that their covariance remains unchanged over the length of the experiment (Liégeois, Laumann, Snyder, Zhou, & Yeo, 2017). In the case of functional connectivity, this assumption arises because the covariance is used as a proxy for connectivity, after being normalized to obtain the correlation matrix. Therefore, assuming weak stationarity is equivalent to assuming that the functional connectivity remains unchanged. Similarly, in spectral DCM, the stationarity assumption allows one to interpret the effective connectivity as remaining unchanged over the length of the time series (K. J. Friston et al., 2014; Moran et al., 2009). This is because the effective connectivity is inferred from the observed cross-spectral density, that is, the Fourier transform of the cross-covariance function (see the Cross-Spectral Density and Functional Connectivity section), which remains unchanged under stationarity assumptions. However, in contrast to functional connectivity, spectral DCM only treats the cross-covariance of the observed time series as a means to an end, where the end is to infer the effective connectivity between neuronal variables (and various other model parameters). In other words, spectral DCM looks “under the hood” of functional connectivity and beyond the observed variables.
The stationarity assumption can certainly be challenged, and it is generally untenable in task experiments; nonetheless, it is widely adopted in resting-state studies. Stationarity is typically assumed not just in functional connectivity and spectral DCM, but also in Granger causality analysis (Granger, 1969; Seth, Chorley, & Barnett, 2013), transfer entropy (Bossomaier, Barnett, Harré, & Lizier, 2016), and autoregressive modeling (Liégeois et al., 2017). That said, Granger causality and information-theoretic methods have been adapted for nonstationary processes (Dhamala, Rangarajan, & Ding, 2008; Gómez-Herrero et al., 2015; Lizier, Prokopenko, & Zomaya, 2008; Paluš, 2018; Wollstadt, Martínez-Zarzuela, Vicente, Díaz-Pernas, & Wibral, 2014), and time-varying approaches to functional connectivity analysis have been rapidly gaining popularity (Lurie et al., 2020; Novelli & Razi, 2022). A time-varying extension of spectral DCM has also been developed (Park, Friston, Pae, Park, & Razi, 2018). All of these methods relax the stationarity assumption and allow the statistical properties of the system to change over the course of the experiment. Practically, in DCM, one appeals to something called an adiabatic approximation: that effective connectivity is constant over a small timescale but can change at longer timescales. This means that one can apply spectral DCM to short segments of data and then examine (or model) slow fluctuations in effective connectivity (Jafarian, Zeidman, Wykes, Walker, & Friston, 2021; Rosch et al., 2019; Zarghami & Friston, 2020).
Linearity
The second assumption that spectral DCM (partially) shares with functional connectivity is linearity, although nonlinear extensions of both methods exist (Kraskov, Stögbauer, & Grassberger, 2004; Stephan et al., 2008). In DCM for fMRI, the use of a linear random differential equation to model the neuronal activity is motivated by the separation of timescales, whereby the neuronal variables represent the slow modes of the system, which are assumed to be linearly coupled, while the endogenous random fluctuations represent the fast modes (K. J. Friston et al., 2011). Spectral DCM also extends the linearity assumption to the observation function, which can be a nonlinear function of time and frequency (as in Equation 25) but is linearly convolved with the neuronal activity (Equation 13). In the Endogenous Fluctuations section, we have observed that deterministic linear models have a limited scope since they can only describe a system that converges to equilibrium or generates a sequence of identical oscillations. Adding stochastic terms to the linear differential equations allows for a richer repertoire of behaviors.
Stochastic Fluctuations
The addition of a stochastic term to a dynamical system is traditionally used to model noise, often assumed to be white and temporally uncorrelated. The underlying assumption is that the noise is due to physical processes operating at a much faster timescale than the state variables, such as microscopic thermal fluctuations. Spectral DCM relaxes this assumption and allows the endogenous fluctuations to be temporally correlated, with a spectrum following a power-law decay in the frequency domain (in the Endogenous Fluctuations section, we saw how this form includes white noise as a special case). This is easily motivated by noting that the endogenous fluctuations are themselves generated by dynamical processes within the source or region of interest. The ensuing temporal autocorrelation makes the endogenous fluctuations differentiable and smooth, in line with their characterization as mixtures of fast modes of a dynamical system (K. J. Friston et al., 2011).
Gaussianity
In line with most DCM approaches, spectral DCM assumes a Gaussian distribution for the prior over model parameters (the nonnegative parameters are transformed using the natural logarithm and are assumed to follow a log-normal distribution). This enables a fast Bayesian inversion scheme called variational Laplace (K. Friston et al., 2007; Zeidman et al., 2023). As for most of the assumptions listed in this section, the Gaussian hypothesis can also be relaxed and other (albeit more computationally intensive) inversion schemes can be used instead, such as Markov chain Monte Carlo methods (Aponte et al., 2022; K. Friston et al., 2007; Xie, Zhang, & Zhao, 2023).
Many-to-One Mapping
There is no one-to-one mapping between effective and functional connectivity (or cross-spectral density) that holds in general. This is a challenge for spectral DCM because it makes model inversion an ill-posed problem with multiple potential solutions, in the absence of any constraints on the way data are generated. As with all ill-posed problems, this issue is addressed by placing prior constraints on the explanations in the form of a model and prior densities over model parameters. When one does not know which priors to use, a weighted average of plausible priors is often performed in DCM analysis using Bayesian model averaging (Hoeting, Madigan, Raftery, & Volinsky, 1999), where each set of priors corresponds to a separate model. A related challenge is that one cannot always use functional connectivity to identify the regions of interest to study using DCM. As we saw in the Effective and Functional Connectivity section, not even the regions showing the largest differences in correlation are guaranteed to coincide with the regions with a change in effective connectivity.
Computational Complexity
Despite the simplifying assumptions mentioned above, the computational complexity of spectral DCM limits the number of brain regions that can be studied in reasonable time. Selecting the regions of interest requires more upfront work than in functional connectivity analysis, which can be quickly performed across the whole brain. Given the large size of the parameter space, the specification of the model has been noted as a conceptual issue for DCM in the past (K. J. Friston, Moran, & Seth, 2013). That said, a more recent theoretical advance now enables the exploration of a large model space using Bayesian model reduction (Seghier & Friston, 2013). Alternatively, one can introduce further assumptions to winnow the parameter space. Using functional connectivity to place prior constraints on the eigenvectors of the effective connectivity matrix enables spectral DCM analyses with dozens of brain regions (Razi et al., 2017). In fMRI, ignoring the spatial variability of the hemodynamics and removing the separation between hidden and observed variables leads to the regression DCM scheme, which can analyze hundreds of regions in minutes (Frässle et al., 2021). However, this method forgoes the state-space formulation and can be understood as a Bayesian multivariate regression in the frequency domain.
ACKNOWLEDGMENTS
Adeel Razi is affiliated with the Wellcome Centre for Human Neuroimaging supported by core funding from Wellcome (203147/Z/16/Z). Adeel Razi is a CIFAR Azrieli Global Scholar in the Brain, Mind & Consciousness Program. Karl Friston is funded by the Wellcome Centre for Human Neuroimaging (Ref: 205103/Z/16/Z) and a Canada-UK Artificial Intelligence Initiative (Ref: ES/T01279X/1).
SUPPORTING INFORMATION
Supporting information for this article is available at https://doi.org/10.1162/netn_a_00348.
AUTHOR CONTRIBUTIONS
Leonardo Novelli: Conceptualization; Formal analysis; Investigation; Methodology; Software; Visualization; Writing – original draft; Writing – review & editing. Karl Friston: Conceptualization; Methodology; Writing – review & editing. Adeel Razi: Conceptualization; Funding acquisition; Methodology; Supervision; Writing – review & editing.
FUNDING INFORMATION
Leonardo Novelli, Australian Research Council, Award ID: DP200100757. Adeel Razi, Australian Research Council, Award ID: DE170100128. Adeel Razi, Australian Research Council (https://dx.doi.org/10.13039/501100000923), Award ID: DP200100757. Adeel Razi, Australian National Health and Medical Research, Award ID: 1194910. Karl Friston, Wellcome Centre for Human Neuroimaging, Award ID: 205103/Z/16/Z. Karl Friston, Canada-UK Artificial Intelligence Initiative, Award ID: ES/T01279X/1.
TECHNICAL TERMS
- SPM:
The Statistical Parametric Mapping open software is widely used for dynamic causal modeling.
- Endogenous fluctuations:
Intrinsic stochastic fluctuations that serve as a proxy for thoughts or mind-wandering-like processes during resting-state brain activity.
- Effective connectivity:
The directed effect of one brain region on another, measured as a rate of change.
- Forward generative model:
A model that generates the data feature of interest using a set of biologically plausible parameters. In spectral DCM, the generative model produces the BOLD cross-spectral density using neuronal and hemodynamic parameters.
- Stochastic process:
A sequence of random variables, typically indexed by time.
REFERENCES
Competing Interests
Competing Interests: The authors have declared that no competing interests exist.
Author notes
Handling Editor: Olaf Sporns