Neural and metabolic basis of dynamic resting state fMRI

Abstract

Resting state fMRI (rsfMRI) as a technique showed much initial promise for use in psychiatric and neurological diseases where diagnosis and treatment were difficult. To realize this promise, many groups have moved towards examining “dynamic rsfMRI,” which relies on the assumption that rsfMRI measurements on short time scales remain relevant to the underlying neural and metabolic activity. Many dynamic rsfMRI studies have demonstrated differences between clinical or behavioral groups beyond what static rsfMRI measured, suggesting a neurometabolic basis. Correlative studies combining dynamic rsfMRI and other physiological measurements have supported this. However, they also indicate multiple mechanisms and, if using correlation alone, it is difficult to separate cause and effect. Hypothesis-driven studies are needed, a few of which have begun to illuminate the underlying neurometabolic mechanisms that shape observed differences in dynamic rsfMRI. While the number of potential noise sources, potential actual neurometabolic sources, and methodological considerations can seem overwhelming, dynamic rsfMRI provides a rich opportunity in systems neuroscience. Even an incrementally better understanding of the neurometabolic basis of dynamic rsfMRI would expand rsfMRI’s research and clinical utility, and the studies described herein take the first steps on that path forward.

1. Introduction

In 1995, Biswal and colleagues used functional magnetic resonance imaging (fMRI) to measure the blood-oxygenation level dependent (BOLD) signal without an explicit task or stimulus (Biswal et al., 1995). They discovered high correlation between signals measured from distant brain regions, “functional connectivity” which formed “networks.” The signals to be correlated were measured during no explicit task, known as “resting state fMRI” (rsfMRI). About ten years after Biswal and colleagues’ findings, numerous studies demonstrated that these networks were altered under various neuropsychological diseases where diagnosis and treatment had proved difficult, such as Alzheimer’s disease and schizophrenia (Garrity et al., 2007; Greicius et al., 2004; Rombouts et al., 2005; Sorg et al., 2007; Whitfield-Gabrieli et al., 2009). Unlike task or stimulus evoked fMRI, rsfMRI has no self-evident test of a perturbation versus a control. This provided methodological freedom which, along with the promise of being able to target previously unknown changes in disease, led to the explosion of rsfMRI methods we are currently experiencing (Soares et al., 2016).

While early work focused on stationary aspects of the rsfMRI signal, “static rsfMRI,” a combination of improved recording methods and limitations of existing analysis techniques have inspired many groups to examine both non-stationary aspects of rsfMRI and stationary aspects of rsfMRI on much shorter time scales than before, collectively called “dynamic rsfMRI” (for prior reviews see (Calhoun et al., 2014; Hutchison et al., 2013a; Keilholz, 2014; Tagliazucchi and Laufs, 2015)). However, even the earliest of these studies noted that it is difficult to determine whether the dynamic changes observed were physiologically meaningful, or merely effects of thermal noise, tissue type differences, subject motion, and so forth (Chang and Glover, 2010; Keilholz et al., 2013; Logothetis et al., 2009; Shmuel and Leopold, 2008). While dynamic rsfMRI can refer to changes measured over longer durations, most dynamic rsfMRI studies focus on time ranges from around 10s to two minutes (Thompson et al., 2013b), approximately spanning the frequency band where functional connectivity was observed in early work (Cordes et al., 2000).

Many studies of dynamic rsfMRI demonstrate that it can detect between-group differences beyond static rsfMRI, yet many other (and often, the same) studies stress the difficulty of concluding the measured differences are truly related to neurometabolic differences between the groups. Investigations of the neurometabolic basis of dynamic rsfMRI indicate that multiple mechanisms may be involved, potentially even manifesting as separable in the rsfMRI signal. Thus, understanding dynamic rsfMRI’s underlying mechanisms is critical not only to knowing whether it is being applied properly, but also may extend its application further.

In this review of the neurometabolic basis of dynamic rsfMRI, I first discuss both indirect evidence (from dynamic rsfMRI studies of how the brain’s state changes) and direct evidence (from correlational studies of dynamic rsfMRI versus alternate neural and metabolic measurements). However, hypothesis-driven work to find underlying mechanisms, particularly in animal models, will ultimately be needed. This understanding is critical not only to prove that dynamic rsfMRI is being used properly, but also to disentangle the different mechanisms which contribute to dynamic rsfMRI measurements.

2.1 Background of dynamic rsfMRI

2.1.1 Why is understanding rsfMRI on shorter time scales necessary?

There are many reasons to examine shorter time periods of rsfMRI. First, even if rsfMRI is primarily stationary, shorter time periods may allow its use during non-task periods of event-driven protocols (Thompson et al., 2013a), may allow shorter scanning protocols, and may also capture differing aspects of the underlying signal (Laumann et al., 2016; Thompson et al., 2015). Second, if rsfMRI is able to measure the known non-stationary states of the brain (Ioannides, 2007), this information may be critical in painting a full picture of the healthy versus diseased brain’s activity, which could otherwise average out to be the same during entire runs (Calhoun et al., 2014).

2.1.2 Dynamic rsfMRI methods

There are numerous different approaches to characterizing the rsfMRI signal on a shorter time scale. Most of these methods attempt to produce a measure of some metric varying over time. Often, these are the same methods as were used in static analysis such as correlation between mean signals in different brain regions (Chang and Glover, 2010), methods in the spectral domain (such as coherence) (Yaesoubi et al., 2015), independent component analysis (Allen et al., 2014; Kiviniemi et al., 2011), number of functional connections (Tomasi et al., 2016), or signal magnitude (Eichele et al., 2008). Such methods typically use a “sliding window” moved across the time series, calculated at several lags, or inherent time-frequency decompositions such as wavelet transformations (Chang and Glover, 2010). Each window’s result can be considered a discrete correlation state and clustering of these states may measure transient cognitive states which the brain is moving through (Calhoun et al., 2014) (which may also be separated based on spectral content (Yaesoubi et al., 2017) or relationship to multimodal measurements (Allen et al., 2017)). Windows can be of fixed length, or of varying length using data-driven methods (Jin et al., 2017; Ou et al., 2014), and alternatives to windowing using multivariate volatility models (Lindquist et al., 2014) or hidden Markov Models (Vidaurre et al., 2017) have also been used.

Alternatively, reproducible patterns of spatiotemporal activations can be extracted from brief instances of the rsfMRI signal through spatiotemporal averaging (Majeed et al., 2011; Majeed et al., 2009), partial least squares (Grigg and Grady, 2010), a point process to represent instantaneous activations (Tagliazucchi et al., 2012a), other signal-based triggering methods (Chen et al., 2015a; Liu and Duyn, 2013), or cross-covariance-based methods (Mitra et al., 2015). These methods can often find “large-scale waves” of activity moving across the brain if they are used to examine multiple time lags (Liang et al., 2015; Majeed et al., 2011; Mitra et al., 2015).

These approaches are not mutually exclusive as, for example, strength of patterns extracted from brief instances can be examined over time either continuously or windowed (Majeed et al., 2011). Different examples of each type of approach are shown in Fig. 1.

Figure 1.

Examples of different types of dynamic observed in resting state fMRI. Results are shown in coronal slices from a dexmedetomidine-anesthetized rat. (A) Filmstrip of a quasi-periodic pattern (QPP), a type of large-scale wave seen reproducibly during the resting state. (B) Anatomical image for reference. (C) Relative strength of the QPP over time. (D) Correlation between two regions of interest either for electrophysiology (black line and light gray line) or BOLD (dark gray line) in a 50 s sliding window, slid in 1 s increments across a 500 s time course. (Reprinted with permission from (Thompson et al., 2015))

2.1.3 Hypotheses for neural basis of dynamic rsfMRI

Despite differences in goal and approach, a common assumption is that synchrony in the measured BOLD signal represents synchrony in the underlying neural activity and physiology on a similar time scale. One could hypothesize that rsfMRI is merely a biomarker, wherein correlations in the BOLD signal over sufficiently long periods mark where neural synchrony also occurs. However, the more common hypothesis is that the instantaneous changes in BOLD (that create these correlations) correlate with instantaneous changes in physiology (cerebral metabolic rate of oxygenation consumption (CMRO₂), cerebral blood volume (CBV) or cerebral blood flow (CBF) (Hyder and Rothman, 2012; Logothetis, 2008)), some of which are due to instantaneous changes in neural electrical activity.

Concluding that synchronized BOLD reflects synchronized neural activity may seem obvious when comparing to static rsfMRI, since “taken together, [studies of the neural correlates of static rsfMRI] strongly indicate that the BOLD signal reflects contributions from multiple frequency bands and possibly from multiple brain processes” (p770 (Keilholz, 2014)). Thus, a default expectation could be for this to translate to dynamic rsfMRI. However, many caveats make this conclusion particularly risky for dynamic rsfMRI. Most studies use correlation coefficients (r, Z or T) which are inherently unit-less, and which are affected by the signal to noise ratio (SNR) regardless of other considerations. Lacking units means that differentiating correlations in small, noise-level fluctuations from correlations in large, physiologically relevant fluctuations is not possible by raw scores alone. Thus, co-varying noise, such as motion (Power et al., 2012) is a major concern. SNR varies by tissue type and due to the magnetic field properties during a particular scan. Thus, even neurologically valid networks could appear in BOLD despite underlying fluctuations not reflecting underlying neural activity. For example, differences in cortical thickness organize into networks which are similar to those found in to rsfMRI correlations (Chen et al., 2008; He et al., 2007).

The influence of structural connectivity (Dawson et al., 2013; Wang et al., 2013) and cortical layer dependence (Baek et al., 2016; Sotero et al., 2015) also needs to be considered. Evidence exists that dynamic connectivity may be shaped by structural connections (Liao et al., 2015; Liegeois et al., 2016). Computational simulations indicate that measurements on longer time scales may more strongly relate to structural connections while shorter time scales may be less constrained by it, exhibiting spontaneous changes in hub locations (Cabral et al., 2017; Honey et al., 2007). Hypothetically reduction of neural activity under anesthesia may cause dynamic rsfMRI to follow structural connections more than under the less constrained awake condition (Barttfeld et al., 2015) (though studies in human subjects do not demonstrate this (Tagliazucchi et al., 2016)).

Thus, dynamic rsfMRI results could be correlated with brain structure because structure serves as a constraint for neural communication. However, within in-vivo data it remains hard to distinguish SNR-related differences due to structure from purely functional changes in the brain. A network could be found under both BOLD and direct measurement of neural electrical activity, but the correlations between brain regions in BOLD could be unrelated to the correlations between brain regions in electrical activity.

Debate also exists how to link cognition to dynamic rsfMRI; recent arguments have focused on whether transient “states” of functional connections reflect discrete cognitive episodes. E.g., Allen and colleagues demonstrate that these states are linked to discrete patterns observed in electroencephalographic (EEG) measurements (Allen et al., 2017), but conversely Laumann and colleagues demonstrate that similar connectivity patterns can be generated from data which preserves stationary correlation relationships but disrupts when they would occur (Laumann et al., 2016). While this argument is not necessarily relevant to large-scale waves or rest-task relationships (see Laumann and colleagues’ discussion section “Stable RSFC is Compatible with ‘Dynamic’ BOLD Features and the Possibility that Ongoing BOLD Activity Influences Behavior”), its resolution is critical to the use of state-finding for disease diagnosis (e.g. schizophrenia and bipolar disorder (Rashid et al., 2014)).

Furthermore, even if we assume that the rsfMRI signal is time-invariant (over sufficiently long periods of time), if rsfMRI has neurological meaning then this stationarity must ultimately be the result of numerous individual neural events. These events influence CMRO₂, CBF and CBV which combine in a non-linear fashion (Hoge et al., 1999). Understanding the neural basis of dynamic rsfMRI thus becomes necessary to understand how these events combine to form the observed networks. However, this is a more difficult problem than finding stationary features, as over longer time periods numerous individual events (not necessarily originally normally distributed (Beggs and Plenz, 2003)), plus noise, convolved with hemodynamic responses, create normal-like distributions that are amenable to standard null hypothesis testing. Even determining whether measured changes are actually dynamic can be difficult depending upon the techniques used. It has been shown since early studies that correlation distributions from sliding window techniques are hard to distinguish between actual data and randomly mismatched data (Hutchison et al., 2013b; Keilholz et al., 2013). If actual state changes occur, these techniques may be unable to detect them (Hindriks et al., 2016; Leonardi and Van De Ville, 2015; Lindquist et al., 2014; Shakil et al., 2016), and noise may make stationary processes appear nonstationary (Hlinka and Hadrava, 2015; Laumann et al., 2016).

Thus, coordinated neural activity and physiology (coming from an unknown, underlying mechanism) combined with noise sources that are non-neurophysiological (e.g. coil-based magnetic field inhomogeneity) or unrelated to the experiment (e.g. motion) influence our measurements. If we want to use our understanding of dynamic rsfMRI, e.g. to develop drugs for disease treatment, we need to be able to pick apart these influences and understand the underlying mechanisms of the coordination observed (Fig. 2).

Figure 2.

Attempting to find the neural basis of dynamic rsfMRI. We measure dynamic changes in BOLD at rest and compare it to behavior variation and disease. These changes are influenced by neural activity and physiology, but are also influenced by factors which may not be of interest to the experiment (e.g. subject motion) or may not be even come from a biological source (e.g. SNR differences due to the radiofrequency coil used). In turn, the behavior variation and disease observed may be linked to these changes (e.g. certain groups may move more than other groups). If we want to use dynamic rsfMRI for disease diagnosis, or to understand the neurometabolic changes in disease so that we can develop drugs to treat them, we need to understand the underlying mechanisms which create the coordination.

Such problems make many techniques necessary for determining if dynamic rsfMRI actually reflects underlying neural and metabolic activity (Calhoun et al., 2014; Hutchison et al., 2013a). In the following sections I will discuss indirect evidence of a neural basis for dynamic rsfMRI from shifting brain states, direct evidence from correlations measured in multimodal studies, and finally the beginnings of evidence toward finding causative mechanisms.

2.2 Dynamic rsfMRI changes with the shifting states of the brain

Comparing state changes in the brain, dynamic rsfMRI has frequently been able to find differences where static rsfMRI has not. The states studied have included stimulation versus rest, controlled changes due to a task or anesthesia, and inherent changes due to healthy structural differences or disease. The work reviewed in this section, covering such state changes, indicates that dynamic rsfMRI may be measuring underlying differences in neurometabolism between groups. However, further application will require understanding the mechanism(s) underlying these between-group differences.

2.2.1 Rest-stimulus interaction

Early activation-based fMRI studies in animal models used anesthesia to perturb the baseline metabolism of the brain. These frequently showed a strong relationship between the baseline (a rest condition) and the magnitude of stimulus activation immediately following (Hyder et al., 2002; Maandag et al., 2007; Pasley et al., 2007; Smith et al., 2002). During-rest dynamicity and this rest-stimulus interaction suggests that dynamic resting state changes may be behaviorally relevant (Northoff et al., 2010), in part because dynamics appear modulated by static networks (Di and Biswal, 2015), and activity in static networks is altered during stimulation (Emerson et al., 2015; Greicius and Menon, 2004; Vogt et al., 2016).

2.2.2 State changes during healthy cognition

Controlled changes to the brain’s state have shown promise in understanding dynamic rsfMRI. Early work with human subjects (using stimulus detection tasks and event-driven fMRI paradigms) linked performance on tasks to resting BOLD amplitude in various brain regions (Boly et al., 2007; Eichele et al., 2008; Sadaghiani et al., 2009). Using dynamic rsfMRI, fluctuations in correlation between brain regions or networks have shown similar effects within a variety of tasks (Cassidy et al., 2016; Gonzalez-Castillo et al., 2015; Jia et al., 2014; Madhyastha et al., 2015; Mooneyham et al., 2017; Sadaghiani et al., 2015; Telesford et al., 2016; Thompson et al., 2013a; Weissman et al., 2006), as have amplitudes within resting state networks themselves (Wohlschlager et al., 2016). Presentation of stimuli (Nummenmaa et al., 2014; Raz et al., 2012), as well as spontaneous state changes including eye closures (Wang et al., 2016), daydreaming (Kucyi and Davis, 2014), sleep state (Tagliazucchi and Laufs, 2014; Wilson et al., 2015) and deprivation (Kaufmann et al., 2016) all alter dynamic rsfMRI measurements.

Selected studies including task, stimulus, and intrinsic state changes generally in healthy cognition are summarized in Table 1. While methods are very heterogeneous, frequent themes include dynamic rsfMRI predicting behavior better than static rsfMRI, stimuli inducing changes in dynamic rsfMRI (Mooneyham et al., 2017; Raz et al., 2012), and particular types of dynamic rsfMRI measured prior to a stimulus or task predicting the evoked activation patterns or performance (Sadaghiani et al., 2015; Thompson et al., 2013a), often better than static rsfMRI (Cassidy et al., 2016; Madhyastha et al., 2015).

Table 1.

Selected studies related to dynamic rsfMRI and healthy cognition. Columns include citation, the method used for calculating changes in the fMRI signal, the task, stimulus or spontaneous condition used in the study, the time scale for comparison between resting and task blocks, and selected observations. (Some studies using BOLD amplitude changes over time are listed as well, as their results are very similar to studies using windowed correlation.)

Citation	Method	Task	Time scale	Selected observations
Weisman et al., 2006	BOLD amplitude	Visual attention	During task	Less deactivation of default region linked to errors
Boly et al., 2007	BOLD amplitude	Somatosensory detection	3 s prestimulus	Increases or decreases (based on region) linked to performance
Eichele et al., 2008	BOLD amplitude	Flanker	6–30 s prestimulus	Increases or decreases (based on region) linked to performance
Sadaghiani et al., 2009	BOLD amplitude	Auditory detection	Up to 9s prestimulus	Increases or decreases (based on region) linked to performance
Thompson et al., 2013a	BOLD amplitude, Windowed correlation (12 s)	Vigilance	Up to 18s prestimulus	Greater anticorrelation or difference indicates better performance
Jia et al., 2014	Windowed correlation (7–72 s)	75 different tasks	During task	Windowed connectivity more related to behavior than static rsfMRI, greater dynamicity indicates better performance
Gonzalez-Castillo et al., 2015	Windowed correlation (23–180 s)	Math, 2-back working memory, Visual attention	Preceding scans (3 mins) and during task	Windowed connectivity more related to behavior than static rsfMRI, shorter windows better
Madhyastha et al., 2015	Windowed correlation (41 s)	Attention network task	Preceding scans (12 mins)	Dynamic rsfMRI more reliable and predicts different factors than static
Sadaghiani et al., 2015	Windowed correlation (15–25 s)	Auditory detection	Up to 25s prestimulus	Different connectivity patterns for hits vs. misses
Telesford et al., 2016	Windowed coherence (25–600 s)	Memory, attention	During task	Strong effect of window size, less effect of task
Wohlschlager et al., 2016	BOLD amplitude	Visual detection	2–4s prestimulus, during task	Primary visual field at rest linked to conscious detection of stimulus
Cassidy et al., 2016	Linear regression between network amplitudes	N-back working memory	During task	Dynamic connectivity changes predicted performance better than activation
Mooneyham et al., 2017	Windowed correlation (75 s)	Mindful breathing	During task	Specific states linked to mindfulness, training increases these states

		Stimulus
Raz et al., 2012	Windowed correlation (30 s)	Sadness-inducing films	During stimuli	Connectivity changes dynamically with emotion
Nummenmaa et al., 2014	Instantaneous phase synchrony	Unpleasant, neutral or pleasant narratives audio	During stimuli	Dynamic changes in different regions depending on the emotion

		Spontaneous states
Kucyi and Davis, 2014	Windowed correlation (30 s)	Daydreaming	Separate resting state and pain stimulation scans	Positive correlation between daydreaming and dynamicity between most nodes, but not during pain
Tagliazucchi and Laufs, 2014	Windowed correlation (120 s)	Sleep state	Nighttime (7pm) and standard resting state scans	Sleep states classifiable with dynamic rsfMRI, some stages higher or lower connectivity than wakefulness
Wilson et al., 2015	Windowed correlation (8–300 s)	Sleep state	During wakefulness and sleep	Default mode network correlation decreases with sleep stage
Wang et al., 2016	Windowed correlation (40 s) clustered into states	Eye closures	Separate resting state and auditory vigilance task scans	States linked to task performance correspond to fewer spontaneous eye closures in resting state
Kaufmann et al., 2016	Windowed correlation (120 s)	Sleep deprivation	Separate resting state and attention task scans	Morning to evening connectivity changes reverted by sleep, not by sleep deprivation

These studies are supported indirectly by evidence from static rsfMRI studies which relate the resting state to subsequent task performance, including on a within-individual basis (Cohen et al., 2014; Gao et al., 2013; Magnuson et al., 2015; Piccoli et al., 2015; Tambini et al., 2010; Thompson et al., 2013a; Wang et al., 2012b). Even though these measurements are static, the varying state at rest suggests dynamic measurements are possible. Other indirect evidence comes from use of resting state network regions as targets for biofeedback (“real time fMRI”) suggests the potential of altering network activity to alter behavior (Garrison et al., 2013a; Garrison et al., 2013b) and functional connectivity can be affected by this biofeedback (Haller et al., 2013; Zhang et al., 2015).

2.2.3 State changes due to exogenous agents

Anesthesia is also a method to control the brain’s baseline state, this has shown an effect on dynamic rsfMRI measurements. Changing the anesthesia level alters the ability to detect frequency, strength and structure of co-activation and large-scale waves observed during the resting state (Amico et al., 2014; Liang et al., 2015; Magnuson et al., 2014b; Thompson et al., 2014b). Anesthesia level also affects correlation and related measures in sliding windows (Barttfeld et al., 2015; Bettinardi et al., 2015; Hudetz et al., 2015; Hutchison et al., 2014; Kafashan et al., 2016; Ma et al., 2017; Tagliazucchi et al., 2016). Other exogenous agents such as psychedelic drugs (Tagliazucchi et al., 2014) and caffeine (Rack-Gomer and Liu, 2012) also impact dynamic rsfMRI.

Selected studies of dynamic rsfMRI under exogenous agents are summarized in Table 2. Most of these studies used windowed correlation and frequently observed that either dynamicity or strength of correlation patterns, or the repertoire of states of these patterns, was reduced under anesthesia (Kafashan et al., 2016; Tagliazucchi et al., 2016). Interestingly, depending on the region, large-scale waves frequently differed little in spatial pattern, but did differ in speed and occurrence rate (Amico et al., 2014; Magnuson et al., 2014b).

Table 2.

Selected studies related to effect of exogenous agents on dynamic rsfMRI. Columns include citation, species used, the method used for calculating changes in the fMRI signal, the anesthesia or other exogenous agent, and selected observations.

Citation	Species	Method	Anesthesia	Selected observations
Amico et al., 2014	Humans	Coactivation	Propofol	Extent of coactivation altered or unaltered under anesthesia, depending on region
Thompson et al., 2014b	Rats	Wave detection (QPP, 6.5–11.25 s)	Isoflurane or dexmedetomidine	Rate of waves changes based on anesthetic
Magnuson et al., 2014b	Rats	Wave detection (QPP, 4.5 s)	Isoflurane and/or dexmedetomidine	Waves become more visible under prolonged anesthesia
Hutchison et al., 2014	Monkeys	Windowed correlation (30 s)	Isoflurane with ketamine	Complex states rarer under deeper anesthesia
Liang et al., 2015	Rats	Coactivation, wave detection (CAP, 15 s)	None, isoflurane	Extent of coactivation altered or unaltered under anesthesia, depending on region
Bettinardi et al., 2015	Rats	Windowed correlation (10 mins)	Medetomidine with ketamine	Connectivity increases and shifts to a higher frequency as anesthesia fades
Hudetz et al., 2015	Rats	Windowed regional homogeneity (20–200 s)	Propofol	Fewer states observed under deeper anesthesia
Barttfeld et al., 2015	Monkeys	Windowed correlation (84 s) clustered into states	None, Propofol	Less dynamic states under anesthesia, closer to structural connections
Kafashan et al., 2016	Humans	Windowed correlation (11 s)	None, sevoflurane	Weaker connections and states are lost under anesthesia, stronger connections and states are preserved
Tagliazucchi et al., 2016	Humans	Detrended fluctuations (25–74 s), windowed correlation (49 s), wavelets to Hurst exponents	None, Propofol	Reduced dynamism under anesthesia and reduced range of functional networks
Ma et al., 2017	Rats	Windowed correlation (60 s)	None, isoflurane	Some connectivity patterns occur more under light anesthesia, others occur more under deep anesthesia
			Other drugs
Tagliazucchi et al., 2014	Humans	Windowed correlation (10–100 s)	None, Psilocybin	Wider repertoire of states under psilocybin
Rack-Gomer and Liu, 2012	Humans	Windowed correlation (10–100 s)	None, Caffeine	Greater variability in correlation under caffeine

The reduction of dynamicity under anesthesia (Barttfeld et al., 2015; Tagliazucchi et al., 2016), and the increase in dynamics under caffeine (Rack-Gomer and Liu, 2012) suggests that shifting arousal levels, or differing levels of vigilance may underlie many of the state changes observed in other studies. This may in particular take effect during studies of varying behavior over time such as are shown in Table 1.

Future work could use transcranial magnetic stimulation, which can modify neural activity on long enough time scales where measuring the alterations using subsequent fMRI is possible (Gollo et al., 2017), and optogenetic stimulation is also possible within scanners themselves (Airan et al., 2013). Good controls are needed in both cases, particularly for scans long enough to measure the “baseline” in each state.

2.2.4 Inherent group differences

Using shorter time windows for calculating correlation has been shown to improve disease detection. Some of the earliest studies of sliding window dynamic rsfMRI were done using patients with Schizophrenia (Sakoglu et al., 2010), and dynamic rsfMRI methods have shown much promise in that disease (Cassidy et al., 2016; Damaraju et al., 2014; Du et al., 2016; Ma et al., 2014; Miller et al., 2016; Rashid et al., 2014; Shen et al., 2014; Yu et al., 2015). Other diseases where dynamic rsfMRI has shown promise have included bipolar disorder (Rashid et al., 2014), Parkinson’s disease (Madhyastha et al., 2015), Alzheimer’s disease (Jones et al., 2012), depression (Kaiser et al., 2016), chronic headache (Wang et al., 2017), autism (Falahpour et al., 2016; Price et al., 2014), attention deficit hyperactivity disorder (ADHD) (Ou et al., 2014), PTSD (Jin et al., 2017; Li et al., 2014b), mild cognitive impairment (Wee et al., 2016), and epilepsy (Douw et al., 2015; Laufs et al., 2014; Liao et al., 2014; Liu et al., 2017; Morgan et al., 2015; Nedic et al., 2015; Ridley et al., 2017). Brain differences in healthy subjects have also been better discriminated using dynamic rsfMRI, including taxi drivers versus non-drivers (Shen et al., 2016), age (Qin et al., 2015), and males versus females (Yaesoubi et al., 2015).

Selected studies of these inherent group differences are summarized in Table 3. The majority of these studies used some form of windowed measurement, particularly windowed correlation, and many studies clustered the measurements in each window into several states for analysis. A frequent observation is that dynamic rsfMRI results in better disease diagnosis than static rsfMRI (Sakoglu et al., 2010; Shen et al., 2016). In some cases, dynamic rsfMRI clarifies details of how the correlations are changed, which would be unclear with static rsfMRI (Falahpour et al., 2016; Ridley et al., 2017). The successful state mapping seen in many studies (Damaraju et al., 2014; Miller et al., 2016) may indicate that the error due to window length mismatch (Shakil et al., 2016) is smaller than inherent group differences. Alternately, dividing the data into windows may improve SNR in measuring an underlying stationary process (Laumann et al., 2016). In support of the non-stationary interpretation, Ou and colleagues produced extremely high accuracy in disease diagnosis by using Bayesian statistics to attempt to classify the state changes rather than a static window (Ou et al., 2014), and Jin and colleagues observed better classification using dynamically varying window length (Jin et al., 2017). This may indicate that alternate methods which avoid fixed window lengths (e.g. (Leonardi and Van De Ville, 2015; Lindquist et al., 2014; Vidaurre et al., 2017)) may be useful in future studies. However, a better understanding of underlying mechanisms is a requirement to know if the underlying processes are actually stationary or nonstationary.

Table 3.

Selected studies related to effect of intrinsic state on dynamic rsfMRI. Columns include citation, the method used for calculating changes in the fMRI signal, the disease or other intrinsic state studied, and selected observations. Note that several of these studies included a task in addition to resting state, see the individual studies for specifics.

Citation	Method	Disease	Selected observations
Sakoglu et al., 2010	Windowed correlation (64 s)	Schizophrenia	More differences seen in dynamic vs. static connectivity, task modulation different between groups
Jones et al., 2012	Windowed correlation (6–297 s)	Alzheimer's disease	Less time spent in default mode network modules in Alzheimer's disease
Damaraju et al., 2014	Windowed correlation (44 s) clustered into states, low-frequency power within states	Schizophrenia	Increases or decreases based on region, not all observed in static connectivity, low frequency power altered in subcortical regions
Ma et al., 2014	Windowed independent vector analysis (75 s)	Schizophrenia	Greater spatial fluctuations in schizophrenia group
Rashid et al., 2014	Windowed correlation (33 s) clustered into states	Schizophrenia, bipolar disorder	Differences between states and transitions particular to each disease, not captured by static rsfMRI
Shen et al., 2014	Windowed correlation (40 s) and low frequency power	Schizophrenia	Low frequency power in connected networks classifies disease versus health
Price et al., 2014	Windowed correlation (20–240 s)	Autism	(Conference paper) Dynamic information increases predictive accuracy of disease
Ou et al., 2014	Bayesian detection of change points (~160 s segments)	Attention deficit hyperactivity disorder	Some interaction states only present in ADHD, these located abnormal networks which distinguished ADHD
Li et al., 2014b	Windowed correlation (28 s)	Post-traumatic stress disorder	Some states appear only in PTSD, can be diagnosed with high selectivity
Laufs et al., 2014	Windowed correlation (30 s)	Epilepsy	Greater variance of connectivity in epilepsy group
Liao et al., 2014	Windowed correlation (100 s)	Epilepsy	Connectivity changes prior to and after seizures
Yu et al., 2015	Windowed correlation (40 s) clustered into states	Schizophrenia	Graph theoretical metrics are lower and there is less variance in schizophrenia
Madhyastha et al., 2015	Windowed correlation (41 s)	Parkinson's disease	No static rsfMRI differences, performance differences in disease (related to dynamic rsfMRI)
Douw et al., 2015	Windowed correlation (102 s)	Epilepsy	Dynamic rsfMRI, but not static rsfMRI or connectivity during task linked to memory disturbance
Morgan et al., 2015	Windowed correlation (60 s)	Epilepsy	Greater variance and greater covariance with seizure network connectivity as disease progresses
Nedic et al., 2015	Entropic analysis of BOLD amplitude autocorrelation	Epilepsy	Less chaotic dynamics in patients
Cassidy et al., 2016	Linear regression between network amplitudes	Schizophrenia	Altered connectivity in schizophrenia, linked to dopamine in schizophrenia only
Du et al., 2016	Windowed correlation (40 s) clustered into states	Schizophrenia	In default mode network, lower connectivity and graph theoretical measures in schizophrenia, duration of states different
Miller et al., 2016	Sums of windowed correlation (44 s) clustered into states	Schizophrenia	Less dynamism in schizophrenia, more pronounced with high levels of hallucinatory behavior
Kaiser et al., 2016	Windowed correlation (36 s)	Depression	Increases or decreases in dynamism based on region
Wee et al., 2016	Windowed correlation (270 s)	Mild cognitive impairment	Altered graph theoretical network properties in mild cognitive impairment
Falahpour, et al. 2016	Windowed correlation (30 s)	Autism	Dynamic rsfMRI indicates connectivity not reduced, but more variable, in autism, whereas static rsfMRI indicates reduction
Wang et al., 2017	Windowed correlation (30–120 s), Wavelet coherence	Chronic headache	Greater wavelet coherence and less dynamism in chronic headache
Jin et al., 2017	Windowed correlation with change points detected with Augmented Dickey-Fuller test (non-fixed length, 20–100 s)	Post-traumatic stress disorder	Better classification with dynamic than static analysis, better classification using varying window length
Liu et al., 2017	Windowed correlation (20–150 s) clustered into states	Epilepsy	Characteristics of states vary more from control as disease duration or seizure frequency increases
Ridley et al., 2017	Windowed nonlinear covariance (90 s)	Epilepsy	Networks involved in generating seizures and spikes have increased static but decreased dynamic rsfMRI versus spike-only networks
		Intrinsic difference
Qin et al., 2015	Windowed correlation (36 s)	Age	Increased variance in connectivity (weighted by salience) as age increases
Yaesoubi et al., 2015	Wavelet coherence	Males vs. females	State occupancy rates differ in males vs. females
Shen et al., 2016	Windowed correlation (56 s) clustered into states	Taxi drivers vs. non drivers	Dynamic rsfMRI, but not static rsfMRI higher in taxi drivers in vigilance network, dwell time in states altered

Results from static rsfMRI indicated that reliability was improved when multiple scans taken at different times from the same subject were used together, rather than a single long scan of equivalent length (Noble et al., 2017a; Noble et al., 2017b). This suggests the state may be shifting over longer time periods, which a single scan did not capture.

2.2.5 Other considerations

All of these studies provide indirect evidence, however. It could be possible that that anesthesia and task studies are not comparable to fully resting state studies as they alter the subjects’ brain states through expectations of future behavior (Laumann et al., 2016) or through unstable physiology under anesthesia (Low et al., 2016). Also, the increased discrimination of disease vs. health seen with dynamic approaches could be due to noise and subsampling altering what is measured, but still measuring different parts of an otherwise static process (Laumann et al., 2016).

As simple perturbations such as eyes open versus eyes closed can alter baseline brain metabolism and its fMRI correlates by 10% or more (Thompson et al., 2016), it is not excessively speculative to suggest that metabolic baseline variations may be occurring naturally during resting state scans. The bottom of Table 1 lists effects of measured spontaneous state changes on dynamic rsfMRI, and some of these state changes certainly occur in other resting state scans without being explicitly induced or monitored. Many of these states are linked to arousal and vigilance (sleep, eyes open or closed, etc.), which are also affected by most of the exogenous agents studied (Table 2). Arousal and vigilance are also influenced by the autonomic nervous system, which has been implicated in changes in spontaneous BOLD fluctuations (Fan et al., 2012). If arousal and vigilance fluctuations do alter dynamic rsfMRI measurements, it would indicate that even stable wakeful rest or uninterrupted sleep would show a change in the state over time.

Even if we assume changes in state are happening during rsfMRI, it is also difficult to determine if the changes measured with rsfMRI are reflecting underlying neural activity. Care needs to be taken to ensure that neural perturbations relevant to the study are being measured rather than changes in baseline physiology. E.g. breath rate, breath rate variability (Birn et al., 2008), and heart rate variability (Chang et al., 2013b; Shmueli et al., 2007) impact the local BOLD signal both through physiological artifacts (such as aliased breathing motion) and also through feeding back to the neural activity which plays a role in cognitive and emotional regulation at rest (Holzman and Bridgett, 2017). This makes separation of neural metabolism from other physiology difficult when state is shifting (Hutchison et al., 2013a; Keilholz et al., 2016). Different states, especially anesthesia and age, also impact subject motion which can create spurious correlations (Power et al., 2012).

These issues make a deeper biological understanding of dynamic rsfMRI critical, which will be discussed in the following sections.

2.3 Dynamic rsfMRI correlates with neural and metabolic activity

Between-group differences observed in dynamic rsfMRI (section 2.2) suggest it reflects underlying neurometabolic differences, and also indicate possible applications for these techniques, e.g. in disease diagnosis. However, this is indirect evidence, and to further apply such techniques (e.g. to develop a drug to treat a condition where dynamic rsfMRI is altered) the underlying mechanism must be understood. Complicating this, multiple underlying mechanisms may alter the dynamic rsfMRI signal, so studies to understand the neurometabolic differences are crucial.

As the resting state is, by definition, not entrained to a task or stimuli, it can be difficult to link resting state fMRI measurements to other measurements of neural and physiological processes. Simultaneous, multimodal, recording provides one possible method of making this connection, since for simultaneous measurements the organized processes observed in resting state fMRI and their underlying neurometabolic sources would have a correlation at a set phase (Shmuel and Leopold, 2008). Several such studies are reviewed in this section. This type of study provides the current best evidence that short time scale measurements of rsfMRI are meaningful, though such studies are limited in how far they can go towards determining cause and effect.

2.3.1 Correlating fMRI and neural activity

As neural action/local field potentials form the brain’s common language, they have been frequently used to validate functional connectivity. Initial work compared noninvasive EEG (Goncalves et al., 2006; He et al., 2008; Laufs et al., 2003; Mantini et al., 2007), electrocorticography in patients (Keller et al., 2013), invasive EEG or local field potentials (LFP) in animals (Lu et al., 2007; Pan et al., 2011; Pan et al., 2013; Shmuel and Leopold, 2008; Wang et al., 2012a), and voltage sensitive dyes/optical imaging (Li et al., 2014a; Vazquez et al., 2014) to static rsfMRI and found much correspondence. Other studies found correspondence between static rsfMRI networks and large scale correlation structures in EEG and magnetoencephalography (MEG) (Deligianni et al., 2014; Hipp and Siegel, 2015). Evidence that these correlations may represent transient states came from comparing the spatial extent of EEG microstates to that of static rsfMRI networks (Britz et al., 2010).

Calculation of dynamic rsfMRI indicated a relationship between correlation coefficients in a sliding window and alpha-, theta- and gamma-band EEG in humans (Chang et al., 2013a; Tagliazucchi et al., 2012b) and LFP in rats (Thompson et al., 2013b), suggesting that these neural correlates were preserved down to comparatively short time scales (10s to 100s). Similarly, dynamic rsfMRI correlated with dynamics in invasive, intracranial EEG in epilepsy patients in alpha, beta and gamma bands, though only in non-seizure regions (Ridley et al., 2017). Higher correlation between fMRI and EEG, or fMRI and near-infrared spectroscopy (NIRS) blood-oxygen measures, was observed when two minute versus ten minute windows were used (Korhonen et al., 2014). This may also translate to brain states as described in the previous section, since the differences between eyes open and eyes closed in dynamic rsfMRI appear linked to distinct dynamic electrophysiological states (Allen et al., 2017). Scholvinck et al. observed large fluctuations over time when correlating LFP power to rsfMRI amplitude using a sliding window, attributing this to dynamic neurovascular coupling (Scholvinck et al., 2010). They also observed stronger correlation between LFP power and rsfMRI amplitude for closed versus open eyes (Scholvinck et al., 2010), which suggests that their results were potentially related to fluctuations in vigilance state, linked to dynamic rsfMRI (Chang et al., 2013a; Thompson et al., 2013a). Likely related to either fluctuating neurovascular coupling or shifting connectivity states, the relationship between infraslow scalp potentials and BOLD fluctuates and can be dynamically measured and clustered into networks which resemble static rsfMRI networks (Grooms et al., 2017).

These results suggest that the matched spatial patterns of correlation seen between different modalities weren’t emerging only from anatomical similarity (for example, (Chen et al., 2008; He et al., 2007)), but rather that the fluctuations which create the correlations in fMRI have a substrate in the brain’s electrical potentials and metabolic fluctuations.

The large scale waves observed using dynamic rsfMRI were also linked to electrical potentials in LFP in rats, but of a much lower frequency than usually used in EEG or LFP: around 0.03–0.4Hz (Thompson et al., 2014b), and in humans at 0.01–0.08 Hz (Grooms et al., 2017). In rats, higher frequencies of electrical activity had a much weaker relationship to the large scale waves, and their amplitudes did not couple to phases of infraslow MRI fluctuations (Thompson et al., 2014a; Thompson et al., 2015). In humans, conversely, a similar spatial range and central frequency of fluctuations was observed in gamma-band power in EEG (Ko et al., 2011) to what was observed using fMRI (Majeed et al., 2011) Also, in MEG, transient formation of similar networks was observed in theta-, alpha- and beta-bands, though these networks were limited to a single hemisphere (de Pasquale et al., 2010). Some of these differences may be due to EEG and MEG being recorded non-invasively through the skull whereas LFP will usually be an inherently more local measure. However, much more work is required to understand the large-scale waves, and if and how they relate to higher frequency neuronal signaling.

2.3.2 Correlating fMRI and brain metabolism

The fluctuations in the static rsfMRI BOLD signal are closely linked to changes in brain metabolism (Tomasi et al., 2013). Some of the changes in metabolic baseline may include confounds which do not directly relate to the underlying neural activity of interest, e.g. heart-rate variability (Chang et al., 2013b). These can be particularly difficult to separate as shifting vigilance levels likely shift both the neural state and also autonomic activity (Fan et al., 2012; Thompson et al., 2013a). However, many results are highly structured spatially and thus likely related to functional connectivity.

Calculations of the CBF and CMRO₂ show networks similar to those found using BOLD (Chen et al., 2015b; Wu et al., 2009), which is a requirement if dynamic fluctuations reflect neural activity. CBF has been dynamically measured in the resting state, and fluctuates in a similar manner to BOLD (Chen et al., 2015b; Tak et al., 2014), shows stronger coupling to BOLD in networks (using static rsfMRI methods) (Liang et al., 2013; Tak et al., 2014), and coupling between CBF and functional connectivity strength may be modulated by tasks similar to dynamic rsfMRI (as shown in the top section of Table 1) (Tak et al., 2015). Measuring CBF and CMRO₂ using NIRS, functional dynamics can be observed during tasks (Dalmis and Akin, 2015; Harrivel et al., 2016) and at rest using a sliding window approach similar to dynamic rsfMRI (Li et al., 2015b). Dynamic measurements of CBV also show large-scale waves similar to BOLD, though at a different peak frequency (Magnuson et al., 2010). CO₂ blood content (Nikolaou et al., 2016) also modulates dynamic rsfMRI. While glucose utilization measured using positron emission tomography (PET) cannot be measured with the temporal resolution of fMRI, it strongly correlates with static functional connectivity measures (Tomasi et al., 2013) which have been shown to vary dynamically in sliding windows (Tomasi et al., 2016).

2.3.3 Understanding correlations

The results from these correlation studies, summarized in Table 4, implicate a wide variety of temporal scales in the source neural activity which leads to the observed resting state fMRI changes. For example, noninvasive EEG in humans indicates anti-correlations between BOLD sliding-window results and certain frequency bands of electrical activity, whereas invasive LFP in rats indicates positive correlations between them (Tagliazucchi and Laufs, 2015). It is also worth noting that the usual rsfMRI signal measured is the <0.1Hz “infraslow” signal which has been studied much less using electrophysiology (Pan et al., 2013; Vanhatalo et al., 2005).

Table 4.

Selected multimodal studies related to dynamic rsfMRI. Columns include citation, the species studied, the method used for calculating changes in the fMRI signal, the other modality, and selected observations. Note that several of these studies either included a task in addition to resting state, or used anesthesia in animal models, see the individual studies for specifics. See also the table in (Keilholz, 2014). Acronyms used in table: EEG electroencephalography LFP local field potentials CBF cerebral blood flow CBV cerebral blood volume NIRS near infrared spectroscopy QPP quasi-periodic patterns (a type of large-scale wave).

Citation	Species	rsfMRI method	Other modality	Selected observations
Scholvinck et al., 2010	Monkeys	Amplitude in 260 s window	LFP power (260 s window)	Variable neurovascular coupling, stronger with eyes closed
Magnuson et al., 2010	Rats	QPP	CBV-weighting (iron oxide nanoparticles)	Similar spatiotemporal structure, waves appear faster in CBF versus BOLD
Tagliazucchi et al., 2012b	Humans	Windowed correlation (125 s), graph theoretical measures of it	EEG power (125 s windows)	Positive correlations with gamma power, negative with alpha and beta power
Magri et al., 2012	Monkeys	BOLD amplitude	LFP power (0.5 s window)	BOLD amplitude scales with gamma power, complimentary information given by alpha and beta power
Chang et al., 2013a	Humans	Windowed correlation (40 s)	EEG power (2 s windows), mean in 40 s windows	Negative correlations with alpha power, rsfMRI network anticorrelation spatial extent linked to alpha power
Thompson et al., 2013b	Rats	Windowed correlation (10–100 s)	LFP power (0.5 s window)	Positive correlation with theta, high beta, and gamma power correlation in matched window
Korhonen et al., 2014	Humans	Windowed correlaiton (120–600 s)	NIRS, EEG infraslow amplitude < 0.55 Hz (120–600 s window)	Higher absolute correlation between modalities in shorter window
Thompson et al., 2014b	Rats	QPP	LFP infraslow amplitude (0.03–0.4 Hz)	Spatial extent of correlation with LFP matches QPP, QPP strength directly correlates
Thompson et al., 2015	Rats	Windowed correlation (50 s), QPP	LFP power (hilbert transform), LFP infraslow amplitude (0.04–0.3 Hz)	Infraslow wave linked to QPP, while higher frequency power linked to windowed correlation
Tak et al., 2015	Humans	Number of functional connections per voxel	CBF (Arterial spin labeling)	As task load increases, CBF/functional connectivity more spatially correlated, dependent upon network
Allen et al., 2017	Humans	Windowed correlation (60 s) clustered into states	EEG power (2 s windows)	States have distinct EEG spectra, one state that only occurs with eyes closed was linked to reduced alpha and increased delta and theta power
Ridley et al., 2017	Humans	Windowed nonlinear covariance (90 s)	Intracranial EEG power (5 s windows, epilepsy patients)	Dynamic rsfMRI correlates to dynamic connectivity in alpha, beta, and gamma bands only in non-seizure regions
Grooms et al., 2017	Humans	Windowed correlation (50 s), QPP	EEG infraslow amplitude (0.01–0.08 Hz)	Dynamic EEG-fMRI coupling produces patterns similar to static rsfMRI networks, QPP strength directly correlates

Infraslow phases coupling to changes in power in higher frequencies, or “phase-amplitude coupling”, is a likely source of some of the translation between temporal scales (Hutchison et al., 2013a; Raichle, 2011). However, thus far it has been difficult to link higher frequency band-limited power to the infraslow wave observed in electrophysiology and rsfMRI. While phase-amplitude coupling has linked relatively slow frequencies >1Hz (Sotero et al., 2015), attempts thus far to connect the infraslow <0.1Hz fluctuations observed in rsfMRI to 1Hz or greater frequencies have not shown this relationship, other than burst-properties due to anesthesia (Nakhnikian et al., 2016; Thompson et al., 2014a). A follow-up study, comparing infraslow and higher frequency neural activity to dynamic rsfMRI found that the high frequencies most strongly correlated with a sliding window method of measuring dynamic rsfMRI, while the infraslow frequencies most strongly correlated with the strength of large scale waves (Thompson et al., 2015). This suggests that there may be multiple underlying neurometabolic processes combining to create the rsfMRI signal, and using dynamic rsfMRI methods may help separate them. Furthermore, static correlations in the fMRI signal exist within higher frequencies measured using MR encephalography (Lee et al., 2013), so the infraslow band’s importance may be partially due to the methods that are used.

Interestingly, the alpha wave, which is strongly linked to vigilance and also the eyes open/eyes closed state, both directly correlates and modulates dynamic rsfMRI measurements (Chang et al., 2013a; Magri et al., 2012; Tagliazucchi et al., 2012b). Considering the effect of intrinsic state on dynamic rsfMRI (Table 1 bottom), this may indicate that shifting vigilance levels (which alter neural activity) may be causing some of the state shifts observed in dynamic rsfMRI.

2.3.4 Limitations of purely correlative methods

A limitation of comparing simultaneously recorded neural electrical activity and BOLD using correlation is that it cannot fully discriminate cause and effect. While the hemodynamic response delay to task or stimulus is generally assumed, it is difficult to know if this assumption translates to the resting state (though there is some evidence, e.g. Pan et al. 2011 (Pan et al., 2011)). Correlation coefficients are highly sensitive to SNR, which will differ between modalities, and thus can only be compared indirectly between them.

Standard correlation methods will produce the strongest statistics if underlying processes are linear (Pearson correlation) or rank (Spearman correlation) in terms of the ground truth underlying relationship. However, non-linearities have been demonstrated in terms of how individual dynamics relate to underlying neural activity (Magri et al., 2012) and influence activations (Wang et al., 2014). An understanding of the neural basis for such nonlinear relationships would require a better model of the underlying systems which correlation alone cannot provide. Use of transfer functions (Herman et al., 2013) or mutual information (Magri et al., 2012) may help overcome some of these limitations, though these methods are impacted by SNR just as correlation coefficients are.

Another important consideration when using correlation coefficients is that the mean (the baseline in a resting state study) of underlying signals will be removed. The baseline of metabolic activity is critical to accurate neural activity measurements (Hyder et al., 2016), so removing the mean through using sliding windows may skew results, as each window will have its own baseline which is removed. In addition, common BOLD pre-processing techniques such as global signal regression also appear to remove the neurometabolic baseline (Thompson et al., 2016). This is concerning considering many of the dynamic rsfMRI state changes may be due to shifting vigilance states, since the global signal is strongly linked to EEG vigilance measures (Wong et al., 2013). Thus, when comparing results described in this section to those described in the previous section, current methods may make linking dynamic rsfMRI between different brain states to correlates of dynamic rsfMRI in different modalities difficult, unless new methods are developed.

The studies described thus far provide good evidence that dynamic rsfMRI measurements do indeed reflect an actual underlying change in neurometabolic activity. However, finding the underlying source of this neurometabolic activity is more challenging.

2.4 Moving toward understanding mechanisms of dynamic rsfMRI

Correlating dynamic rsfMRI to other modalities (section 2.3) provides evidence it has a neurometabolic basis, and interestingly shows that multiple mechanisms may be involved. However, to ultimately understand the processes in the brain that create the organizations observed with dynamic rsfMRI measurements, we need to hypothesize on these processes, alter them, and observe the change in the dynamic rsfMRI. This understanding is necessary not only to apply dynamic rsfMRI results (e.g. to develop treatments for diseases where it’s altered), but also may help us separate which effects observed in the dynamic rsfMRI signal relate to which underlying processes.

Given the relative newness of dynamic rsfMRI, its primary study in humans where intervention is often not possible, and considering the multitude of possible sources for synchronization in resting state fMRI (Keilholz, 2014), little work has been done here thus far. Animal models where invasive studies are possible will need to play an important role (Hutchison and Everling, 2012; Pan et al., 2015).

2.4.1 Neural communication as a mechanism of dynamic rsfMRI

A study by Magnuson et al. investigated the large-scale waves moving across the brain in anesthetized rats prior to and following transection of the corpus callosum. Interestingly, after transection the waves went from being primarily bilateral to being either unilateral or not present (Fig. 3) (Magnuson et al., 2014a). This is evidence that callosal connections are needed to maintain the bilaterality of that particular dynamic rsfMRI result.

Figure 3.

Examples of quasi-periodic patterns (QPP) measured using BOLD, a large-scale wave that is observed to repeatedly propagate across the cortex during the resting state. Coronal images of the brains of rats anesthetized with dexmedetomidine are shown, see Fig. 1B for reference. (a) Example of bilateral wave as QPP. (b) Example of unilateral wave as QPP. (c) Example of where no clear wave was present in the QPP. While many runs showed no wave after any surgery (c), bilateral waves were only present after sham callosotomy surgery (a) and unilateral waves were more present after true callosotomy surgery (b). (Reprinted with permission from Magnuson et al. (Magnuson et al., 2014a). The publisher for this copyrighted material is Mary Ann Liebert, Inc. publishers.)

As the dynamic examined by Magnuson and colleagues was cortical, its requirement of callosal connectivity for bilaterality is a strong indicator that it reflects synchronized neural communication (cross-callosal tracts are typically implicated in bilateral neural communication). Similar bilateral dynamics have been observed in non-cortical regions unconnected by the corpus callosum (Majeed et al., 2011; Thompson et al., 2014b), however, suggesting that other systems must maintain bilaterality elsewhere. Further evidence for the hypothesis of an underlying system creating sporadic, synchronized oscillations comes from observations of sporadic, correlated fluctuations in local oxygen levels (Li et al., 2015a) whose central frequency is similar to observed dynamic waves in the rsfMRI signal (Majeed et al., 2009).

The general hypothesis of neural communication underlying dynamic rsfMRI is compelling as it would explain much of the indirect and correlative evidence discussed in prior sections. Computational models suggest that observed dynamics could emerge from neural communication (Deco et al., 2013; Havlicek et al., 2011). Further evidence is that deficits in the release of cortical dopamine (an important large scale neuromodulator) were related in unmedicated schizophrenia to poor working memory, whereas dynamic rsfMRI was a better predictor of performance than dopamine in healthy subjects (Cassidy et al., 2016). Neuromodulation has also been observed in differences in the task/rest relationship between electrical potentials and blood flow (Ardestani et al., 2016). If neural communication is the source of most of the useful dynamic rsfMRI measurements, many questions will need to be answered, including the origin point of dynamics and how the resultant neural potentials lead to stronger or weaker hemodynamic effects which ultimately lead to stronger or weaker correlation coefficients.

2.4.2 Other potential mechanisms of dynamic rsfMRI

While neural communication is a compelling explanation, evidence is emerging that the resting state fMRI signal is influenced by multiple processes which can create different types of functional connectivity. Models from studies of direct-current potentials from scalp direct-current EEG (DC-EEG) implicate the blood-brain barrier in creating volume conduction (Vanhatalo et al., 2003; Voipio et al., 2003). DC-EEG has been linked to resting state fMRI in humans (Hiltunen et al., 2014) and DC-LFP rats (Pan et al., 2013). The global signal in resting state fMRI has been shown to have strong effects on observed networks, even if it is regressed (Murphy et al., 2009), and it can dynamically alter correlation coefficients (Scheinost et al., 2016). Researchers will need to consider volume conduction and other global effects.

While not volume conduction, other factors that are unrelated to the neural communication being studied can influence macroscopic brain regions. Cardiac and respiratory effects may produce cerebrospinal fluid (CSF) pulsation, which may produce reproducible spatiotemporal patterns in the brain which are weaker but still present in gray matter (Kiviniemi et al., 2016). This likely has interplay with neural communication, and further study is needed.

Slow to infraslow oscillations in calcium concentration reflect neural activity and correlate with hemodynamic fluctuations that translate into the BOLD signal (Du et al., 2014), and can be initiated by relatively small groups of cortical neurons (Stroh et al., 2013). While these oscillations may be driven by neural communication, the slow time scale and large spatial spread (possibly including into white matter (Kiviniemi et al., 2016)) suggests other physiological mechanisms are likely in effect. Under a forepaw stimulation protocol, slower oscillations were linked to astrocytic networks and complicated prediction of BOLD activation patterns (Schulz et al., 2012). Astrocytic calcium activity has been found on a similar spatiotemporal scale to typical rsfMRI studies (Kuga et al., 2011). Conversely, electrical stimulation elicits a response in fMRI that appears to be independent of astrocytic calcium release due to the inositol 1,4,5-triphosphate-type-2 receptor (Jego et al., 2014). However, this may not account for rhythmic slow calcium activity at rest (DiNuzzo, 2014), and the observation that baseline activity complicates prediction of stimuli-response (Schulz et al., 2012) may indicate that resting state fMRI studies may be measuring the effects of multiple mechanisms (Thompson et al., 2015).

As cerebrovascular reactivity has been shown to influence many aspects of static rsfMRI (Golestani et al., 2016; Liu et al., 2013), it likely influences dynamic rsfMRI measures. This may be a potential source of fluctuating neurovascular coupling (Grooms et al., 2017; Scholvinck et al., 2010).

Many potential mechanisms likely combine to create the dynamic rsfMRI signal, both in terms of what we observe and also in terms of how it relates to behavior, disease, etc. Figure 4 lists some of the discussed mechanisms including their spatial and temporal scales. Many of these mechanisms could affect behavioral variation: For example, vigilance/arousal fluctuations could be affected by the autonomic nervous system and conscious thought processes (neuronal communication), but on longer time scales affected by calcium waves or CSF pulsation (e.g. large-scale wave phase affects task performance (Abbas et al., 2016)), and on even longer time scales modified based on metabolic baseline shifting over minutes or hours (e.g. caffeine consumption alters CBF and CMRO₂ (Merola et al., 2017)).

Figure 4.

Potential mechanisms of dynamic rsfMRI at multiple spatiotemporal scales. Spatial scale is estimated roughly on the horizontal axis, temporal scale estimated roughly on the vertical axis. Spatially, the microscopic scale may include actual neural communication (which may be across macroscopic distances, however), local metabolism (which may be linked to global metabolism, however), and the anatomical structure on which these act. The macroscopic, but still not global, scale may include factors such as infraslow calcium waves and CSF pulsation. Some factors that influence the entire brain globally, such as volume-conduction and global metabolic changes (in particular between groups) likely also influence dynamics. While in general smaller structures correspond to faster changes, consider that neuronal communication (as fast as a few ms for an action potential) is constrained by axonal structure which may not change for years. Global brain metabolism may also fluctuate relatively quickly, but its baseline may be due to intrinsic brain health and anatomy. CSF pulsation, infraslow calcium waves, and volume conduction cluster at a temporal range close to that of dynamic rsfMRI and its correlated infraslow electrophysiology.

The studied systems, potentially unknown systems, and correlated noise such as motion (Power et al., 2012) and gray matter SNR variability (Logothetis et al., 2009), all combine to influence the rsfMRI signal. Separating the contributions of individual sources becomes even more critical as the time scale is reduced. However, the complexity of this issue means that it is a very rich field for future investigation. Even partially unraveling the causative mechanisms of short-time scales of correlation and individual fluctuations of rsfMRI has potential to expand the uses of rsfMRI substantially. Can dynamic rsfMRI be useful for psychiatric drug discovery, for example? This would require knowing the cell populations and receptor types to target. While preliminary, studies such as Magnuson et al. (Magnuson et al., 2014a) and Jego et al. (Jego et al., 2014) put this idea within reach.

3. Conclusion

Studies comparing dynamic rsfMRI to behavior or disease have shown that these techniques can often distinguish changing brain states better than static rsfMRI. This implicates a neurometabolic basis for these techniques, and multimodal studies comparing the BOLD signal to other methodologies have provided evidence this is the case. However, such multimodal studies implicate multiple mechanisms, and correlation-based experimental designs are limited in terms of their ability to separate cause and effect. Thus, to understand the multiple mechanisms underlying dynamic rsfMRI, work is needed to intervene in processes and understand the causation involved. While some evidence exists that dynamic rsfMRI measures may emerge from neural communication, other mechanisms including the blood-brain barrier, CSF flux, astrocytic calcium activity, or variable neurovascular coupling may interact and likely heavily influence the dynamics observed in BOLD as well. Currently, applications of dynamic rsfMRI are generally not considered for typical microscopic studies such as those of metabolism, individual neural processes such as receptor-ligand interactions, chemical engineering, and drug discovery. However, an understanding of dynamic rsfMRI’s neural basis is a promising way to make this connection. Understanding from where these dynamics emerge will greatly improve the utility of dynamic rsfMRI.