Evidence for a Large-Scale Brain System Supporting Allostasis and Interoception in Humans

Abstract

Large-scale intrinsic brain systems have been identified for exteroceptive senses (e.g., sight, hearing, touch). We introduce an analogous system for representing sensations from within the body, called interoception, and demonstrate its relation to regulating peripheral systems in the body, called allostasis. Employing the recently introduced Embodied Predictive Interoception Coding (EPIC) model, we used tract-tracing studies of macaque monkeys, followed by two intrinsic functional magnetic resonance imaging samples (N = 280 and N = 270) to evaluate the existence of an intrinsic allostatic/interoceptive system in the human brain. Another sample (N = 41) allowed us to evaluate the convergent validity of the hypothesized allostatic/interoceptive system by showing that individuals with stronger connectivity between system hubs performed better on an implicit index of interoceptive ability related to autonomic fluctuations. Implications include insights for the brain’s functional architecture, dissolving the artificial boundary between mind and body, and unifying mental and physical illness.

The brain contains intrinsic systems for processing exteroceptive sensory inputs from the world, such as vision, audition, and proprioception/touch (e.g., ¹). Accumulating evidence indicates that these systems work via the principles of predictive coding (e.g., ^2–7), where sensations are anticipated and then corrected by sensory inputs from the world. The brain, as a generative system, models the world by predicting, rather than reacting to, sensory inputs. Predictions guide action and perception by continually constructing possible representations of the immediate future based on their prior probabilities relative to the present context^8,9. We and others have recently begun studying the hypothesis that ascending sensory inputs from the organs and systems within the body’s internal milieu are similarly anticipated and represented (i.e., autonomic visceral and vascular function, neuroendocrine fluctuations, and neuroimmune function)^10–16. These sensations are referred to as interoception^17–19. Engineering studies of neural design²⁰, along with physiological evidence²¹, indicate that the brain continually anticipates the body’s energy needs in an efficient manner and prepares to meet those needs before they arise (e.g., movements to cool the body’s temperature before it gets too hot). This process is called allostasis^20–22. Allostasis is not a condition or state of the body – it is the process by which the brain efficiently maintains energy regulation in the body. Allostasis is defined in terms of prediction, and recent theories propose that the prediction of interoceptive signals is necessary for successful allostasis (e.g., ^{10,15,23–25}). Thus, in addition to the ascending pathways and brain regions important for interoception (e.g., ^17,18,26,27), recent theoretical discussions (e.g., ¹¹) have proposed the existence of a distributed intrinsic allostatic/interoceptive system in the brain (analogous to the exteroceptive systems). A full investigation of the predictive nature of an allostatic/interoceptive brain system requires multiple studies under various conditions. Here, we identify the anatomical and functional substrates for a unified allostatic/interoceptive system in the human brain and reporting an association between connectivity within this system and individual differences in interoceptive-related behavior during allostatically relevant events.

In this paper, we first review tract-tracing studies of non-human animals that provide the anatomical substrate for our hypothesis that the brain contains a unified, intrinsic system for allostasis and interoception. Next, we present evidence of this hypothesized system in humans using functional connectivity analyses on three samples of task-independent (i.e., “resting state”) functional magnetic resonance imaging (fMRI) data (also called “intrinsic” connectivity). We then present brain-behavior evidence to validate the hypothesized allostatic/interoceptive system by using an implicit measure of interoception during an allostatically challenging task. Finally, we summarize empirical evidence to show that this allostatic/interoceptive system is a domain-general system that supports a wide range of psychological functions including interoception, emotion, memory, reward, cognitive control, etc.^28,29. That is, whatever else this system might be doing – remembering, directing attention, etc., – it is also predictively regulating the body’s physiological systems in the service of allostasis to achieve those functions²³.

Our work synthesizes anatomical and functional brain studies that together evidence a single brain system – comprised of the salience and default mode networks – that supports not just allostasis but a wide range of psychological functions (emotion, pain, memory, decision-making, etc.) that can all be explained by their reliance on allostasis. To our knowledge, this evidence and our simple yet powerful explanation has not been presented despite the fact that many functional imaging studies show that the salience and default mode networks support a wide range of psychological functions (i.e., they are domain general; e.g., ³⁰; for review, see ^28,29). Our paper provides the groundwork for a theoretical and empirical framework for making sense of these findings in an anatomically principled way. Our key hypotheses and results are summarized in Table 1.

Table 1.

Summary of this study’s hypotheses, predictions or questions, and results.

Embodied Predictive Interoception Coding (EPIC) Hypothesis	Experimental Prediction	Result in the Current Study
Interoception and visceromotor control are part of a unified brain system that supports allostasis (Fig 1)	Primary interoceptive cortex (e.g., dorsal mid/posterior insula) is anatomically and functionally connected to agranular and dysgranular visceromotor hubs of the cortex (e.g., sgACC, pACC, aMCC)	The interoceptive and visceromotor hubs are anatomically connected in monkeys (Table 2) The interoception and visceromotor hubs are functionally connected in humans (Fig. 2, Supplementary Table 1) Coordinates for human hubs are show in Table 3
	The allostatic/interoceptive system also includes subcortical and brainstem visceromotor regions.	Previously established subcortical and brainstem visceromotor regions (e.g., hypothalamus, periaqueductal gray) are part of the unified system for allostasis/interoception (Fig. 4, Supplementary Figure 6)
	The allostatic/interoceptive brain system contains limbic cortices.	The allostatic/interoceptive system comprises two established large scale brain networks that contain the majority of limbic cortices: the salience network and the default mode networks (Fig. 3, Supplementary Figure 3)
	Connectivity in the allostatic/interoceptive system is related to an implicit performance measure of interoception in humans	The correspondence between sympathetic arousal (electrodermal activity) and experienced arousal during an allostatically challenging task is related to functional connectivity within the allostatic/interoceptive system in humans (Supplementary Figure 8)
The allostatic/interoceptive system is domain-general.	The allostatic/interoceptive system sits at the core of the brain’s computational architecture.	Many hubs of the allostatic/interoceptive system have been previously identified as members of the “rich club,” which are the most densely connected within the brain and therefore help constitute the brain’s “neural backbone” for coordinating neural synchrony (Fig. 3, Supplementary Table 4)
	Brain activity and connectivity in the allostatic/interoceptive system is associated with a variety of psychological functions	Both the default mode network and the salience network support a variety of mental phenomena across major psychological domains (e.g., cognition, emotion, perception, and action; Fig. 5)

Other hypotheses, such as the computational dynamics of the proposed allostatic/interoceptive network are beyond the scope of this study. ACC = anterior cingulate cortex; aMCC = anterior midcingulate cortex; dmIns = dorsal mid insula; dpIns = dorsal posterior insula; pACC = pregenual anterior cingulate cortex; sgACC = subgenual anterior cingulate cortex.

Anatomical evidence supporting the proposed allostatic/interoceptive system

Over three decades of tract-tracing studies of the macaque monkey brain clearly demonstrate an anatomical substrate for the proposed flow of the brain’s prediction and prediction error signals. Specifically, anatomical studies indicate a flow of information within the laminar gradients of these cortical regions according to the structural model of corticocortical connections developed by Barbas colleagues (³¹; for a review, see ³²). In addition, the structural model of corticocortical connections has been seamlessly integrated with a predictive coding framework^11,12. Unlike other models of information flow that work in specific regions of cortex, the structural model successfully predicts information flow in frontal, temporal, parietal, and occipital cortices^33–37. Accordingly, prediction signals flow from regions with less laminar development (e.g., agranular regions) to regions with greater laminar development (e.g., granular regions), whereas prediction error signals flow in the other direction. In our recently developed theory of interoception, called the Embodied Predictive Interoception Coding (EPIC) model¹¹, we integrated Friston’s active inference approach to predictive coding^38–40 with Barbas’s structural model to hypothesize that less-differentiated agranular and dysgranular visceromotor cortices in the cingulate cortex and anterior insula initiate visceromotor predictions through their cascading connections to the hypothalamus, the periaqueductal gray (PAG), and other brainstem nuclei known to control the body’s internal milieu^41–44 (also see ³²; red pathways in Fig 1); simultaneously, the cingulate cortex and anterior insula send the anticipated sensory consequences of those visceromotor actions (i.e., interoceptive predictions) to the more granular primary interoceptive cortex in the dorsal mid to posterior insula (dmIns/dpIns^18,45,46 ; blue solid pathways; Fig 1). Using this logic, we identified a key set of cortical regions with visceromotor connections that should form the basis of our unified system for interoception and allostasis (we also included one subcortical region, the dorsal amygdala (dAmy), in this analysis due to the role of the central nucleus in visceromotor regulation; for details, see endnote 1). This evidence is summarized in Table 2. As predicted by our EPIC model, most of the key visceromotor regions in the proposed interoceptive system do, in fact, have monosynaptic, bidirectional connections to primary interoceptive cortex, reinforcing the hypothesis that they directly exchange interoceptive prediction and prediction error signals. We also confirmed that these visceromotor cortical regions indeed monosynaptically project to the subcortical and brainstem regions that control the internal milieu (i.e., the autonomic nervous system, immune system, and neuroendocrine system), such as the hypothalamus, PAG, parabrachial nucleus (PBN), ventral striatum, and nucleus of the solitary tract (NTS) (Table 2, right column).

Fig 1.

We identified key visceromotor cortical regions (in red) that provide cortical control the body’s internal milieu, including the anterior mid cingulate cortex (aMCC; also called dorsal anterior cingulate cortex (dACC), e.g., ^41,42), pregenual anterior cingulate cortex (pACC), subgenual anterior cingulate cortex (sgACC; for a review of the cingulate, see ¹⁷⁶), and the ventral anterior insula (vaIns; also called agranular insula^43,183 or posterior orbitofrontal cortex¹⁹⁴); these regions have a less-developed laminar structure (i.e., they are agranular or dysgranular^32,176). We also included the dorsal amygdala because it contains the central nucleus which is also involved in visceromotor control (for a review, see ¹⁵³). Primary interoceptive cortex spans the dorsal mid insula (dmIns) to the dorsal posterior insula (dpIns)¹⁷ along a dysgranular to granular¹⁹⁵ gradient (green regions). Barrett & Simmons (2015) summarized preliminary tract-tracing evidence, supporting the EPIC model¹¹, that allostasis and interoception are maintained within an integrated system involving limbic cortices (in red) that initiate visceromotor directions to the hypothalamus and brainstem nuclei (e.g., periaqueductal gray, parabrachial nucleus, nucleus of the solitary tract; citations in Table 2) to regulate the autonomic, neuroendocrine, and immune systems (red paths). These visceromotor control regions (less developed laminar organization) also send anticipated sensory consequences of visceromotor changes (as interoceptive prediction signals) to primary interoceptive cortex (more-developed laminar organization; solid blue paths). The incoming sensory inputs from the internal milieu of the body are carried along the vagus nerve and small diameter C and Aδ fibers (dashed green paths) to primary interoceptive cortex in the dorsal sector of the mid to posterior insula (for a review, see ¹⁷); comparisons between prediction signals and ascending sensory input results in interoceptive prediction error. Current interoceptive predictions can be updated by passing prediction error signals to visceromotor regions (dashed blue paths); prediction errors are learning signals and also adjust subsequent predictions. (For simplicity, ascending feedback to visceromotor regions is not shown). aMCC = anterior midcingulate cortex; dAmy = dorsal amygdala; dmIns = dorsal mid insula; dpIns = dorsal posterior insula; pACC = pregenual anterior cingulate cortex; sgACC = subgenual anterior cingulate cortex; vaIns = ventral anterior insula.

Table 2.

Summary of tract-tracing study results in non-human animals, demonstrating anatomical connections between cortical visceromotor and primary interoceptive sensory regions, as well as between cortical and non-cortical visceromotor regions.

	Primary Interoceptive Cortex	Visceromotor Regions				Subcortical and Brainstem Visceromotor Structures
	To dpIns/dmIns	To vaIns	To sgACC (BA 25)	To pACC (BA 24, 32)	To aMCC (BA 24)	To Amygdala	To other subcortical and brainstem regionsa
From dpIns/dmIns	-	Case A, Fig 1155	Not evidentb	Case 1, Fig 5157	Case B, Fig 3158	Case 2, Fig 3154 Case BB-B, Fig 160	Hypothalamus (rat)159 PAG: not observed160 PBN (rat)161,162 V. Striatum 163 NTS (rat)162
From vaInsc	Case C, Fig 4155 Case A, Fig 1158	-	Case OM20, Fig 8164	Case 1, Fig 5157	Case 2, Fig 6157 Case A, Fig 1158	Case A, Fig 1158 Case 103, Fig 3165 Fig 2, Table 2166	Hypothalamus43 PAG160 PBN (rat)161 V. striatum167 NTS (rat)162
From sgACC (BA 25)	Not evidentd	Case M707168	-	Case 1, Fig 5157 Fig 2A169	Case 3, Fig 7157 Fig 3A169	Case 103, Fig 3165 Fig 5156	Hypothalamus154,170,171 PAG160,171 PBN 171 Striatum 171 NTS (rat)172,173
From pACC (BA 24, 32)	Not evidentd	Case M776168	Fig 1169	-	Case 3, Fig 7157 Fig 3A169	Case 103, Fig 3165 Fig 5156	Hypothalamus43, PAG160 PBN (cat)174 V. striatum (cat)174 NTS (rat)173
From aMCC (BA 24)	Case C, Fig 4155	Case A, Fig 1155	Case 3, Fig 4175	Case 1, Fig 5157 Fig 2A169	-	Case 103, Fig 3165 Fig 5156	Hypothalamus43 PAG160 PBN: not present176 V. striatum177 NTS (rat)172
From Amygdala	Case C, Fig 4155 Lateral basal nucleus; Case 5, Fig 6154	Case A, Fig 1155 Case 4, Fig 5154	Fig 6156	Fig 13169	Fig 6156	-	Hypothalamus43, PAG160 PBN178 V. striatum179 NTS178

Note. Connectivity evidence is in monkeys unless otherwise indicated (e.g., rats, cats). Some connections from dpIns/dmIns to the NTS are unclear due to ambiguity in how Saper (1982)¹⁶² reported subregions of the insula.

^aWe did not assess for projections from subcortical and brainstem regions to cortical regions because we only wanted to determine if the cortical regions support visceromotor control.

^bConnection from dpIns/dmIns to sgACC not evident in several monkey studies that have the potential to show them (e.g., ^{158,169,180–182}).

^cThe medial portion of the vaIns exhibits connectivity with subcortical and brainstem regions, but not the lateral portion of the vaIns^43,183.

^dConnection from sgACC to dpIns/dmIns and from pACC to dpIns/dmIns not evident in several monkey studies that have the potential to show them (e.g., ^{155,168,180,181}), although weak, direct connectivity is evident in a recent tractography study in humans (Ghaziri, et al., 2015¹⁸⁴, Figure 5). Moreover, connections between sgACC, pACC, and dpIns have been observed in intrinsic functional connectivity analyses in humans (e.g., Fig. 6 of ¹⁸⁵). The discrepancy between human findings and the tract tracing studies in monkeys failing to show connectivity might reflect an expansion of Brodmann area (BA) 24 anterior and ventral to the corpus callosum in humans relative to monkeys and/or the presence of connections between BAs 25/32 and the posterior insula in humans that do not exist in monkeys (Evrard, H. personal communication, December 27, 2015).

BA = Brodmann area; aMCC = anterior midcingulate cortex; dmIns = dorsal mid insula; dpIns = dorsal posterior insula; NTS = nucleus of the solitary tract; PAG = periaqueductal gray; PBN = parabrachial nucleus; pACC = pregenual anterior cingulate cortex; sgACC = subgenual anterior cingulate cortex; V. striatum = ventral striatum.

Next, we tested for evidence of these connections in functional data from human brains. Axonal connections between neurons, both direct (monosynaptic) and indirect (e.g., disynaptic) connections, are closely reflected in intrinsic brain systems (for a review, see ^47,48). As such, we tested for evidence of these connections in functional connectivity analyses on two samples of low-frequency Blood Oxygenation Level Dependent (BOLD) signals during task-independent (i.e., “resting state”) fMRI scans collected on human participants (discovery sample, N = 280, 174 female, mean age = 19.3 years, SD = 1.4 years; replication sample, N = 270,142 female, mean age = 22.3 years, SD = 2.1 years). We then examined the validity of these connections in a third independent sample of participants (N = 41, 19 female, mean age = 33.5 years, SD = 14.1 years), following which we situated these findings in the larger literature on network function.

Results

Cortical and amygdalar intrinsic connectivity supporting a unified allostatic/interoceptive system in humans

Our seed-based approach estimated the functional connectivity between a set of voxels of interest (i.e., the seed) and the voxels in the rest of the brain as the correlation between the low-frequency portion of their BOLD signals over time, producing a discovery map for each seed region. Starting with the anatomical regions of interest specified by the EPIC model, and verified in the anatomical literature, we selected seed regions guided by previously published functional studies. We selected two groupings of voxels in primary interoceptive cortex (dpIns and dmIns) that consistently showed increased activity during task-dependent fMRI studies of interoception (Table 3, first and second rows). We selected seed regions for cortical visceromotor regions and the dAmy using related studies (Table 3, remaining rows). As predicted, the voxels in primary interoceptive cortex and visceromotor cortices showed statistically significant intrinsic connectivity (Fig. 2; replication sample Supplementary Figure 1). The dpIns was intrinsically connected to all visceromotor areas of interest (seven two-tailed, one-sample t-tests were each significant at p < 10⁻⁷; Supplementary Table 1), and dmIns was intrinsically connected to most of them (Supplementary Table 1). The discovery and replication samples demonstrated high reliability for connectivity profiles of all seeds (η² mean = 0.99, SD = 0.004).

Table 3.

Seeds used for intrinsic connectivity analyses.

Seed	Type of region predicted by EPIC model	Cortical Lamination	MNI Coordinates
dpIns	Primary interoceptive cortex	Granular	36, −32, 16186
dmIns	Primary interoceptive cortex	Dysgranular	41, 2, 3187
sgACC	Visceromotor control	Agranular	2, 14, −6188
pACC	Visceromotor control	Agranular	13, 44, 0186
aMCC	Visceromotor control	Agranular	9, 22, 33189
mvaIns	Visceromotor control	Agranular	30, 16, −14190
lvaIns	Sensory integration	Agranular	44, 6, −15189
dAmy	Visceromotor control	N/A	27, 3, −12191

Note: All seeds are in the right hemisphere. Evidence for cortical lamination comes from Vogt (2005)^42,192,193.

Each anatomical region of interest was represented by one 4-mm-radius seed except for the ventral anterior insula (vaIns), which required a medial and a lateral seed (mvaIns and lvaIns, respectively) to capture the previously-established functional distinction between the medial visceromotor network (containing mvaIns) and the orbital sensory integration network (containing lvaIns) in the orbitofrontal cortex¹⁸³.

aMCC = anterior midcingulate cortex; dAmy = dorsal amygdala; dmIns = dorsal mid insula; dpIns = dorsal posterior insula; lvaIns = lateral ventral anterior insula; mvaIns = medial ventral anterior insula; pACC = pregenual anterior cingulate cortex; sgACC = subgenual anterior cingulate cortex.

Fig. 2.

Eight regions (“seeds”) used to estimate the unified allostasis/interoceptive system connecting the cortical and amygdalar visceromotor regions and primary interoceptive regions. The left column shows the “seed” region for each discovery map on a human brain template. The middle column summarizes the anatomical connectivity derived from anterograde and/or retrograde tracers injected in macaque brains at a location homologous to the human seed (asterisks with blue arrows). The right column shows the human intrinsic connectivity discovery maps depicting all voxels whose time course is correlated with the seed’s (ranging from p < 10⁻⁵ in red to p < 10⁻⁴⁰ in yellow, uncorrected, N = 280). To avoid Type I and Type II errors, which are enhanced with the use of stringent statistical thresholds¹⁹⁶, we opted to separate signal from random noise using replication, according to the mathematics of classical measurement theory¹⁴⁸. These results replicated in a second sample, N = 270 participants, indicating that they are reliable and cannot be attributed to random error (Supplementary Figure 1). Functional connectivity to the entire amygdala and other subcortical regions are shown in Fig. 4. Tract tracing figures were adapted with permission as follows: subgenual anterior cingulate cortex (sgACC) via retrograde tracers in Fig 1 of Vogt & Pandya (1987)¹⁶⁹, pregenual ACC (pACC) via retrograde tracers in Fig 5 of Morecraft, et al. (2012)¹⁵⁷, anterior midcingulate cortex (aMCC) via retrograde tracers in Fig 7 of Morecraft, et al. (2012)¹⁵⁷, dorsal amygdala (dAmy) via retrograde tracers in Fig 3 of Aggleton, et al. (1980)¹⁶⁵, medial ventral anterior insula (mvaIns) and lateral ventral anterior insula (lvaIns) via anterograde tracers in Fig 1 of Mesulam & Mufson (1982)¹⁵⁸, dorsal mid insula (dmIns) and dorsal posterior insula (dpIns) via anterograde tracers in Fig 3 of Mesulam & Mufson (1982)¹⁵⁸. The monkey anatomical connectivity figures were colored red to visualize results and some were mirrored to match the orientation of the human brain maps. The figures from Morecraft, et al. (2012)¹⁵⁷ were adapted to show the insula in its lateral view.

Next, we computed η² for all pairs of maps to determine their spatial similarity⁴⁹ (mean = 0.56, SD = 0.17), and then performed K-means clustering of the η² similarity matrix to determine the configuration of the system. Results indicated that the allostatic/interoceptive system is composed of two intrinsic networks connected in a set of overlapping regions (Fig. 3; replication sample, Supplementary Figure 2). The spatial topography of one network resembled an intrinsic network commonly known as the default mode network (Supplementary Figure 3 and Supplementary Figure 4; for a review, see ⁵⁰). The second network resembled an intrinsic network commonly known as the salience network (Supplementary Figure 3 and Supplementary Figure 4; e.g., ^51,52), the cingulo-opercular network⁵³, or the ventral attention network⁵⁴. Resemblance was confirmed quantitatively by comparing the percent overlap in our observed networks to reconstructions of the default mode and salience networks reported in Yeo, et al.⁵⁵ (Supplementary Table 2). Other cortical regions within the interoceptive system shown in Fig. 3 (e.g., dorsomedial prefrontal cortex, middle frontal gyrus), not listed in Table 2, support visceromotor control via direct anatomical projections to the hypothalamus and PAG (Supplementary Table 3), supporting our hypothesis that this system plays a fundamental role in visceromotor control and allostasis.

Fig. 3.

The unified allostatic/interoceptive system is composed of two large-scale intrinsic networks (shown in red and blue) that share several hubs (shown in purple; for coordinates, see Supplementary Table 4). Hubs belonging to the “rich club” are shown in yellow. Rich club hubs figure adapted with permission from van den Heuval & Sporns (2013)⁸⁵. All maps result from the sample of 280 participants binarized at p < 10⁻⁵ uncorrected from a one-sample two-tailed t-test. These results replicated in a second sample, N = 270 participants, indicating that they are reliable and cannot be attributed to random error (Supplementary Figure 2). aMCC = anterior midcingulate cortex; dAmy = dorsal amygdala; dpIns = dorsal posterior insula; dmIns = dorsal mid insula; IFG = inferior frontal gyrus; ITG = inferior temporal gyrus; lvaIns = lateral ventral anterior insula; MCC = midcingulate cortex; mvaIns = medial ventral anterior insula; pACC = pregenual anterior cingulate cortex; PHG = parahippocampal gyrus; pMCC = posterior midcingulate cortex; PostCG = postcentral gyrus; sgACC = subgenual anterior cingulate cortex; STS = superior temporal sulcus.

Subcortical, hippocampal, brainstem, and cerebellar connectivity supporting a unified allostatic/interoceptive system in humans

Using a similar analysis strategy, we assessed the intrinsic connectivity between the cortical and dorsal amygdalar seeds of interest and the thalamus, hypothalamus, cerebellum, the entire amygdala, hippocampus, ventral striatum, PAG, PBN, and NTS. The observed functional connections with these cortical and amygdalar seeds, which regulate energy balance, strongly suggest that the proposed allostatic/interoceptive system itself also regulates energy balance (see Supplementary Discussion for details). All results replicated in our independent sample (N = 270; Supplementary Figure 5, η² mean = 0.98, SD = 0.008). Fig. 4 illustrates the connectivity between default mode and salience networks and the non-cortical targets in the discovery sample. Supplementary Figure 6 shows connectivity between the individual cortical and amygdalar seed regions listed in Table 2. We also observed specificity in the proposed allostasis/interoception system: non-visceromotor brain regions that are unimportant to interoception and allostasis, such as the superior parietal lobule (Supplementary Figure 7), did not show functional connectivity to the subcortical regions of interest.

Fig. 4.

Subcortical connectivity of the two integrated intrinsic networks within the allostatic/interoceptive system (N = 280; p < 0.05 uncorrected). These results replicated in a second sample of N = 270 (Supplementary Figure 5). PAG = periaqueductal gray; PBN = parabrachial nucleus; V. Striatum = ventral striatum; NTS = nucleus of the solitary tract.

The cortical hubs of the allostatic/interoceptive system also overlapped in their connectivity to non-cortical regions involved in allostasis (purple in Fig. 4), including the dAmy, the hypothalamus, the PBN, and two thalamic nuclei – the VMpo and both the medial and lateral sectors of the mediodorsal nucleus (MD, which shares strong reciprocal connections with medial and orbital sectors of the frontal cortex, the lateral sector of the amygdala, and other parts of the basal forebrain; for a review, see ⁵⁶). Additionally, the connector hubs also shared projections in the cerebellum and hippocampus (see Fig. 4).

Taken together, our intrinsic connectivity analyses failed to confirm only five monosynaptic connections (8%) that were predicted from non-human tract-tracing studies: hypothalamus-dAmy, hypothalamus-dpIns, PAG-dAmy, PAG-medial ventral anterior insula (mvaIns), and NTS-subgenual anterior cingulate cortex (sgACC). This is approximately what we would expect by chance; however, there are several factors that might account for why these predicted connections did not materialize in our discovery and replication samples. First, all discrepancies involved the sgACC, PAG, or hypothalamus, whose BOLD data exhibit poor signal to noise ratio due to their small size and their proximity to white matter or pulsating ventricles and arteries⁵⁷. Second, individual differences in anatomical structure can make inter-subject alignment challenging, particularly in 3-T imaging of the brainstem where clear landmarks are not always available. Of the connections that did not replicate, one involved the anterior insula; there is some disagreement in the macaque anatomical literature as to the exact location of the anterior insula (e.g., ^45,58–60), which might help explain any lack of correspondence between intrinsic and tract-tracing findings that we observed.

Validating the functions of the allostatic/interoceptive system in humans

The allostatic/interoceptive system reported in Fig. 3 replicated in the validation sample (η² mean = 0.84, SD = 0.05 compared with discovery sample cortical maps; η² mean = 0.76, SD = 0.07 compared with discovery sample subcortical maps). These η² values are respectable and demonstrate adequate reliability of the system according to conventional psychometric theory, although the lower η² values are likely due to the smaller sample size which magnifies the effects of poor signal-to-noise ratio in subcortical regions. Convergent validity for the proposed allostatic/interoceptive system was demonstrated in that individuals with stronger functional connectivity within the system also reported greater arousal while viewing images that evoked greater sympathetic nervous system activity. Participants viewed ninety evocative photos known to induce a range of autonomic nervous system changes and corresponding feelings of arousal⁶¹, as well as changes in BOLD activity within these regions^62,63. We predicted, and found, that individuals showing stronger intrinsic connectivity within the allostatic/interoceptive system (specifically, connectivity between dpIns and anterior midcingulate cortex (aMCC)) also demonstrated a stronger concordance between objective and subjective measures of bodily arousal while viewing allostatically relevant images (p = 0.003; see Supplementary Figure 8; see Supplementary Discussion for details).

There were three reasons for demonstrating the convergent validity of the proposed allostatic/interoceptive system using this task. First, there is a decades-old body of research indicating that interoception enables the subjective experience of arousal (⁶⁴; e.g., ^65,66). Thus, the amount of joint information shared by an objective, psychophysiological measure of visceromotor change (skin conductance) and the subjective experience of arousal (self-report ratings) is an implicit, behavioral measure of interoceptive ability. Indeed, individuals with more accurate interoceptive ability exhibit a stronger correspondence between subjective arousal and physiological arousal in response to similar evocative photos⁶⁷. Second, explicit reports of interoceptive performance on heartbeat detection tasks (e.g., ^68–70) are complex to interpret neutrally because they require synthesizing and comparing information from other systems (somatosensory system⁷¹, frontoparietal control systems, and, for heartbeat detection, the auditory system); in addition, these tasks are sometimes too hard (yielding floor effects) or have questionable validity⁷⁰.

At this juncture, it is tempting to ask if the unified allostatic/interoceptive system is specific to allostasis and interoception. From our perspective, this is the wrong question to be asking. The last two decades of neuroscience research have brought us to the brink of a paradigm shift in understanding the workings of the brain, setting the stage to revolutionize brain: mind mapping. Neuroscience research is increasingly acknowledging that brain networks have a one (network) to many (function) mappings^{28–30,72–74}. Our findings contribute to this discussion: a brain system that is fundamental to allostasis and interoception is not unique to those functions, but instead is also important for a wide range of psychological phenomena that span cognitive, emotional, and perceptual domains (Fig. 5.). This finding is not a failure of reverse inference. It suggests a functional feature of how the brain works.

Fig. 5.

The default mode and salience networks each support a wide array of psychological functions, as evidenced by a literature review of psychological or other states that are sensitive to functional or structural features of these networks. These results are consistent with the idea that the default mode and salience networks are domain-general networks that support interoception and allostasis, which we propose are key processes that contribute to all psychological functions. Each sub-figure shows a set of results from an independent study, with citations as follows. Default mode network: Social fear¹⁹⁷, Physical fear¹⁹⁷, Atypical emotions¹⁹⁸, Emotion¹⁹⁹, Emotion concepts²⁰⁰, Subjective value²⁰¹, Social affiliation²⁰², Chronic pain²⁰³, Trait judgments²⁰⁴, Empathy²⁰⁵, Moral judgments²⁰⁶, Theory of mind²⁰⁴, Reward²⁰⁷, Smoking addiction²⁰⁸, Memory²⁰⁹, Prospection²⁰⁹, Association²¹⁰, and Concepts²¹¹. Salience network: Atypical emotion¹⁹⁸, Affect²¹², Effortful recall²¹³, Executive attention²¹⁴, Atrophy and stress (chronic yellow, current red)²¹⁵, Atrophy and mental illness¹²³, Interoception²¹⁶, Recognition memory²¹⁷, Bilingualism²¹⁸, Multimodal integration¹, Thermal pain²¹⁹, Alcohol craving²²⁰, Empathy²²¹, Decision making²²², Errors²²³, Word form (yellow)²²⁴, Propranolol during aversion²²⁵, and Hot spots²²⁶.

Discussion

The integrated allostatic/interoceptive brain system is a complex cortical and subcortical system consisting of connected intrinsic networks. Our work demonstrates a single brain system that supports not just allostasis but also a wide range of psychological phenomena (emotions, memory, decision-making, pain) that can all be explained by their reliance on allostasis. Other studies have already shown that regions controlling physiology are also regions that control emotion. In fact, this was Papez’s original logic for assuming that the “limbic system” was functionally for emotion. This paper goes beyond this observation. Regions controlling and mapping of inner body physiology lie in networks that also social affiliation, pain, judgments, empathy, reward, addiction, memory, stress, craving, decision making, etc. (Fig. 5). More and more, functional imaging studies are finding that the salience and default mode networks are domain-general (e.g., ³⁰; for review, see ^28,29). Our paper provides the groundwork for a theoretical and empirical framework for making sense of these findings in an anatomically principled way.

Our investigation was strengthened by our theoretical framework (the EPIC model¹¹), the converging evidence from structural studies of the brain (i.e., tract-tracing studies in monkeys plus the well-validated structural model of information flow), our use of multiple methods (intrinsic connectivity in humans, as well as brain-behavior relationships), and our ability to replicate the system in three separate samples totaling over 600 human participants. Our results are consistent with prior anatomical and functional studies that have investigated portions of this system at cortical and subcortical levels (e.g., ^{17,18,26,27,75–78}), including evidence that limbic cortical regions control the brainstem circuitry involved with allostatic functions such as cardiovascular control, respiratory control, and thermoregulatory control⁷⁹, as well as prior investigations that focused on the intrinsic connectivity of individual regions such as the insula (e.g., ⁸⁰), the cingulate cortex (e.g., ⁸¹), the amygdala (e.g., ⁸²), and the ventromedial prefrontal cortex (e.g., ⁸³); importantly, our results go beyond these prior studies in several ways. First, we observed an often-overlooked finding when interpreting the functional significance of certain brain regions: the dorsomedial prefrontal cortex, the ventrolateral prefrontal cortex, the hippocampus, and several other regions have both a structural and functional pattern of connectivity that indicates their role in visceromotor control. A second often-overlooked finding is that relatively weaker connectivity patterns (e.g., between the visceromotor sgACC and primary interoceptive cortex) are reliable, and future studies may find that they are of functional significance. Third, we demonstrated behavioral relevance of connectivity within this network, something that prior studies of large-scale autonomic control networks have yet to test (e.g., ^75–77). Taken together, our results strongly support the EPIC model’s hypothesis that visceromotor control and interoceptive inputs are integrated within one unified system¹¹, as opposed to the traditional view that the cerebral cortical regions sending visceromotor signals and those that receive interoceptive signals are organized as two segregated systems, similar to the corticospinal skeletomotor efferent system and the primary somatosensory afferent system.

Perhaps most importantly, the allostatic/interoceptive system has been shown to play a role in a wide range of psychological phenomena, suggesting that allostasis and interoception are fundamental features of the nervous system. Anatomical, physiological, and signal processing evidence suggests that a brain did not evolve for rationality, happiness, or accurate perception; rather, all brains accomplish the same core task²⁰: to efficiently ensure resources for physiological systems within an animal’s body (i.e., its internal milieu) so that an animal can grow, survive, thrive, and reproduce. That is, the brain evolved to regulate allostasis²¹. All psychological functions performed in the service of growing, surviving, thriving, and reproducing (such as remembering, emoting, paying attention, deciding, etc.) require the efficient regulation of metabolic and other biological resources.

Our findings add an important dimension to the existing observations that the default mode and salience networks serve as a high-capacity backbone for integrating information across the entire brain⁸⁴. Diffusion tensor imaging studies indicate, for example, that these two networks contain the highest proportion of hubs belonging to the brain’s “rich club,” defined as the most densely interconnected regions in the cortex^73,85 (several of which are connector hubs within the allostatic/interoceptive system; see Fig. 3, Supplementary Table 4). All other sensory and motor networks communicate with the default mode and salience networks, and potentially with one another, through these hubs^1,85. The agranular hubs within the two networks, which are also visceromotor control regions, are the most powerful predictors in the brain^11,32. Indeed, hub regions in these networks display a pattern of connectivity that positions them to easily send prediction signals to every other sensory system in the brain^12,32.

The fact that default mode and salience networks are concurrently regulating and representing the internal milieu, while they are routinely engaged during a wide range of tasks spanning cognitive, perceptual, and emotion domains, all of which involve value-based decision-making and action⁸⁶ (e.g., ^{87–90; 30;} for a review, see ⁸⁸), suggest a provocative hypothesis for future research: whatever other psychological functions the default mode and salience networks are performing during any given brain state, they are simultaneously maintaining or attempting to restore allostasis and are integrating sensory representations of the internal milieu with the rest of the brain. Therefore, our results, when situated in the published literature, suggest that the default mode and salience networks create a highly connected functional ensemble for integrating information across the brain, with interoceptive and allostatic information at its core, even though it may not be apparent much of the time.

When understood in this framework, our current findings do more than just pile on more functions to the ever-growing list attributed to the default mode and salience networks (which currently spans cognition, attention, emotion, perception, stress, and action; see ^28,30). Our results offer an anatomically plausible computational hypothesis for a set of brain networks that have long been observed but whose functions have not been fully understood. The observation that allostasis (regulating the internal milieu) and interoception (representing the internal milieu) are at the anatomical and functional core of the nervous system^18,20 further offer a generative avenue for further behavioral hypotheses. For example, it has recently been observed that many of the visceromotor regions within the unified allostatic/interoceptive system contribute to the ability of SuperAgers to perform memory and executive function tasks like young⁹¹.

Furthermore, our findings also help to shed light on two psychological concepts that are constantly confused in the psychological and neuroscience literatures: affect and emotion. If, whatever else your brain is doing—thinking, feeling, perceiving, moving—it is also regulating your autonomic nervous system, your immune system, and your endocrine system, then it is also continually representing the interoceptive consequences of those physical changes. Interoceptive sensations are usually experienced as lower-dimensional feelings of affect^92,93. As such, the properties of affect—valence and arousal^94,95—can be thought of as basic features of consciousness^96–102 that, importantly, are not unique to instances of emotion.

Perhaps the most valuable aspect of our findings is their value for moving beyond traditional domain-specific or “modular” views of brain structure/function relationships¹⁰³, which assume a significant degree of specificity in the functions of various brain systems. A growing body of evidence requires that these traditional modular views be abandoned^28,104,105 in favor of models that acknowledge that neural populations are domain-general or multi-use. The idea of domain-generality even applies to primary sensory networks, as evidenced by the fact that multisensory processing occurs in brain regions that are traditionally considered unimodal (e.g., auditory cortex responding to visual stimulation^106,107). The absence of specificity in brain structure/function relationships is not a measurement error or some biological dysfunction, but rather it is a useful feature that reflects core principles of biological degeneracy that are also evident in the genome, the immune system, and every other biological system shaped by natural selection¹⁰⁸.

No study is without limitations. First, there are potential issues identifying homologous regions between monkey and human brains⁴⁷; nonetheless, we still found evidence for the majority of the monosynaptic connections predicted by the EPIC model. Second, we used an indirect measure of brain connectivity in humans (functional connectivity analyses of low-frequency BOLD data acquired at rest) that reflects both direct and indirect connections and can, in principle, inflate the extent of an intrinsic network⁴⁷. Moreover, low frequency BOLD correlations may reflect vascular rather than neural effects in brain¹⁰⁹. Nonetheless, our results exhibit specificity: the integrated allostatic/interoceptive system conforms to well-established salience and default mode networks and is remarkably consistent with both cortical and subcortical connections repeatedly observed in tract-tracing studies of non-human animals. Third, although our fMRI procedures were not optimized to identify subcortical and brainstem structures and study their connectivity (e.g., ^57,75,76,110), we nonetheless observed 92% of the predicted connectivity results. Finally, many studies find that activity in the default mode and salience networks have an inverse or negative relationship (sometimes referred to as “anti-correlated”), meaning that as one network increases its neural activity relative to baseline, the other decreases. Such findings and interpretations have recently been challenged on both statistical and theoretical grounds (e.g., ¹¹¹; see Supplementary Discussion). In fact, when global signal is not removed in pre-processing, the two networks can show a pattern of positive connectivity (e.g., ¹¹²). Fourth, our demonstration of a brain/behavior relationship (using the evocative pictures) was merely a preliminary evaluation of how individual differences in the function of this system are related to individual differences in behavior. Additionally, our use of electrodermal activity as a measure of sympathetic nervous system activity is arguably too specific because different components of the sympathetic nervous system react differently¹¹³, and peripheral sensations associated with changes in electrodermal activity might not be processed by the interoceptive brain circuitry that we are studying here, thus complicating the interpretation of our results. However, we did not intend to assess any particular path carrying information about electrodermal activity specifically, and we believe that – despite their limitations – our results are still useful and hypothesis-generating. Future work will be needed to understand this and other brain/behavior relationships involving this system more thoroughly.

This work one in a series of future studies to precisely test the EPIC model, including its predictive coding features (not just the anatomical and functional correlates as shown here). Future research must focus on the ongoing dynamics by which the default mode and salience networks support allostasis and interoception, including the predictions they issue to other sensory and motor systems. It is possible, for example, that both networks use past experience in a generative way to issue prediction signals, but that the default mode network generates an internal model of the world via multisensory predictions (consistent with ^114–116), whereas the salience network issues predictions, as precision signals, to tune this model with prediction error (consistent with the salience network’s role in attention regulation and executive control; e.g., ^51,117,118). Unexpected sensory inputs that are anticipated to have allostatic implications (i.e., likely to impact survival, offering reward or threat) will be encoded as “signal” and learned to better support allostasis in the future, with all other prediction error is treated as “noise” and safely ignored (¹¹⁹; for discussion, see ¹²⁰). These and other hypotheses regarding the flow of predictions and prediction errors in the brain (e.g., incorporating the cerebellum, ventral striatum, and thalamus²⁴ can be tested using new methods such laminar MRI scanning at high (7 T) magnetic field strengths (e.g., ¹²¹).

Future research that provides a more mechanistic understanding of how the default mode and salience networks support interoception and allostasis will also reveal insights into the mind-body connections at the root of mental and physical illness and their comorbidities. For example, in illness, the neural representations of the world that underlie action and experience may be directed more by predicted allostatic relevance of information than by the need for accuracy and completeness in representing the environment. Indeed, atrophy or dysfunction within parts of the interoceptive system are considered common neurobiological substrates for mental and physical illness^122–124, including depression¹²⁵, anxiety¹²⁶, addiction¹²⁷, chronic pain¹²⁸, obesity¹²⁹, and chronic stress^130,131. By contrast, increased cortical thickness in MCC is linked to the preserved memory of SuperAgers relative to their more typically performing elderly peers^132,133, suggesting a potential mechanism for how exercise (via the sustained visceromotor regulation it requires) benefits cognitive function in aging¹³⁴ and why certain activities, such as mindfulness or contemplative practice, can be beneficial (e.g., ^135,136). Ultimately, a better understanding of how the mind is linked to the physical state of the body through allostasis and interoception may help to resolve some of the most critical health problems of our time, such as the comorbidities among mental and physical disorders related to metabolic syndrome (e.g., depression and heart disease¹³⁷), or how chronic stress speeds cancer progression¹³⁸, as well as offer key insights into how an opioid crisis¹³⁹ and recorded numbers of suicides¹⁴⁰ emerge.

Methods

Participants

Discovery and replication samples

We randomly selected 660 participants (365 female, 55%, 18–30 years) from 1,000 healthy participants described in Yeo, et al.^55,141. The 1,000 participants were native English-speaking adults, 18–35 years, with normal or corrected-to-normal vision, and reported no history of neurological or psychiatric conditions. We removed 79 participants (11%) due to head motion and outlying voxel intensities; we removed 31 more participants (4.7%) due to lack of signal in superior and lateral parts of the brain (see Analysis section). Our final dataset of 550 participants was randomly divided into a discovery sample of N=280 (174 female, 62%, mean=19.3 years, SD=1.4 years) and a replication sample of N=270 (142 female, 53%, mean=22.3 years, SD=2.1 years).

We also randomly selected 150 participants (75 female, 50%, mean=22.5, SD=2.0 years) from the N=1,000 to generate maps of the established default mode and salience networks.

Validity sample

We selected all 66 young and middle aged participants (33 female, 18–60 years, mean=34.8 years, SD=13.8 years) from an existing dataset of 111 participants (56 female, 18–81 years, mean=46.6 years, SD=18.9 years) recruited from the Boston area during 2012–2014 for a study examining age-related changes in how affect supports memory¹⁴². Only 41 participants (14 female, 47%, 20–60 years, mean=33.8 years, SD=14.1 years) had both high-quality fMRI BOLD data and sufficient electrodermal activity changes according to previously established procedures (see Analysis section). Specifically, 12 participants exhibited excessive head motion and outlying voxel intensities, and 16 participants lacked electrodermal responses. Participants were right-handed, native English speakers and had normal or corrected-to-normal vision. None reported any history of neurologic or psychiatric condition, learning disability or serious head trauma. Participants did not smoke and did not ingest substances that interfere with autonomic responsiveness (e.g., beta-blockers, anti-cholinergic medications).

Sample size

No pre-specified effect size was known, so we used a large portion of a third-party dataset (N=660) and the maximum size of a second dataset collected in our lab with young and middle-aged adults (N=66).

Procedure

Discovery and replication samples

Participants provided written informed consent in accordance with the guidelines set by the institutional review boards of Harvard University or Partners Healthcare. Participants completed MRI structural and resting state scans and other tasks unrelated to the current analysis. MRI data were acquired at Harvard and the Massachusetts General Hospital across a series of matched 3T Tim Trio scanners (Siemens, Erlangen, Germany) using a 12-channel phased-array head coil. Structural data included a high-resolution multiecho T1-weighted magnetization-prepared gradient-echo image (multiecho MP-RAGE). Parameters for the structural scan were as follows: repetition time (TR)=2,200 ms, inversion time (TI)=1,100 ms, echo time (TE)=1.54 ms for image 1 to 7.01 ms for image 4, flip angle (FA)=7°, 1.2×1.2×1.2-mm voxels, and field of view (FOV)=230 mm. The functional resting state scan lasted 6.2 min (124 time points). The echo planar imaging (EPI) parameters for functional connectivity analyses were as follows: TR=3,000 ms, TE=30 ms, FA=85°, 3×3×3-mm voxels, FOV=216 mm, and 47 axial slices collected with interleaved acquisition and no gap between slices.

Validity sample

Participants were consented in accordance with the institutional review board. Data were acquired on separate sessions across several days. The first session consisted of a 6-min seated baseline assessment of peripheral physiology, the EXAMINER cognitive battery¹⁴³, a second 6-min seated baseline, the evocative images task, and other tasks. Only the evocative images task is relevant for this study. Electrodes were placed on the chest, hands, and face to record electrocardiogram, electrodermal activity, and facial electromyography, respectively. A belt with a piezoelectric sensor was secured on the chest to record respiration. Only the electrodermal activity data are reported here. Electrodermal activity was recorded using disposable electrodermal electrodes (containing isotonic paste) affixed to the thenar and hypothenar eminences of the left hand. Data were collected using BioLab v3.0.13 (Mindware Technologies, Gahanna, OH, USA). Participants sat upright in a comfortable chair in a dimly lit room. Ninety full-color photos were selected from the International Affective Picture System (IAPS) and used to induce affective experiences⁶¹. The pictures were selected based on normative ratings of pleasure/displeasure (valence) and arousal experienced when viewing them (i.e. unpleasant-high arousal, pleasant-high arousal, unpleasant-low arousal, pleasant-low arousal, neutral valence-low arousal; Supplementary Table 5). Participants viewed the photos sequentially on a 120×75 cm high-definition screen two meters away. Photos were grouped into three blocks of thirty each, with the order of the photos within each block fully randomized. For each trial, participants viewed an IAPS photo for six seconds, and then rated their experience for valence and arousal using the Self Assessment Manikin (SAM¹⁴⁴). Only the arousal ratings are relevant to this report and they ranged from 1 (“Very calm”) to 5 (“Very activated”). A variable inter-trial interval of 10–15 seconds followed the rating prior to presentation of the next picture. Before beginning the task, participants were familiarized with the SAM rating procedure and practiced by rating five pictures. The photos and rating scales were administered via E-Prime (Psychology Software Tools, Pittsburgh, PA).

The second laboratory testing session involved MRI scanning, consisting of a structural scan, resting state scan, and other tasks unrelated to the present report (presented elsewhere¹⁴²). MRI data were acquired using a 3T Tim Trio scanner (Siemens, Erlangen, Germany) using a 12-channel phased-array head coil. Structural data included a high-resolution T1-weighted MP-RAGE with TR=2,530 ms, TE=3.48 ms, FA=7°, and 1×1×1-mm isotropic voxels. The functional resting state scan lasted 6.40 min (76 time points). The EPI parameters were as follows: TR=5,000 ms, TE=30 ms, FA=90°, 2×2×2-mm voxels, and 55 axial slices collected with interleaved acquisition and no gap between slices. Participants were instructed to keep their eyes open without fixating and remain as still as possible.

Analysis of task-independent (“resting state”) fMRI data

Quality assessment

We applied established censoring protocols for head motion and outlying signal intensities using AFNI (https://afni.nimh.nih.gov/afni/) following Jo, et al. ¹⁴⁵ and described in the following three steps: First, we disqualified an fMRI volume if AFNI’s enorm motion derivative parameter (derived from afni_proc.py) was greater than 0.3 mm. Second, we disqualified an fMRI volume if the fraction of voxels with outlying signal intensity (AFNI’s 3dToutcount command) was greater than 0.05. Third, if a volume surpassed either criterion, we removed that volume, the prior volume, and the next two volumes. In a separate procedure, we disqualified discovery and replication participants who lost more than 10% of their 124 volumes due to either criterion (79 participants, 11%). Quality assessment for surface-based processing required removing 31 additional participants (4.7%) due to a lack of signal in the most superior and lateral parts of the brain, which would result in incomplete group connectivity maps; no participants were removed for this reason in the validity sample. In the validity sample, we removed participants who lost more than 40% of their 76 volumes, removing 12 participants (18%); we used a more lenient threshold due to the small sample size (N=66). The fraction of volumes censored per participant using the aforementioned approach by Jo, et al.¹⁴¹ yielded nearly identical results to another established censoring approach described in Power, et al. ¹⁴⁶ as implemented in AFNI’s afni_restproc.py script.

Preprocessing

We applied standard Freesurfer preprocessing steps to both samples of resting state data (http://surfer.nmr.mgh.harvard.edu). These included removal of the first four volumes, motion correction, slice timing correction, resampling to the MNI152 cortical surface (left and right hemispheres) and MNI305 subcortical volume (2 mm isotropic voxels), spatial smoothing (6 mm FWHM, surface and volume separately) and temporal filtering (0.01 Hz high-pass filter and 0.08 Hz low-pass filter). We did not use global signal regression as to prevent spurious negative correlations (“anti-correlated networks”), which can interfere with interpreting the connectivity results¹¹¹.

Functional connectivity analysis

We estimated cortical connectivity using surface-based analyses, affording more sensitive and reliable discovery maps and reducing artifacts around sulcal and opercular borders by registering each participant’s native space to MNI152 space via Freesurfer’s reconstruction of each participant’s cortical surfaces¹⁴⁷. The surface-based intrinsic analyses also allowed us to incorporate the selected subcortical seed (dAmy), but did not allow us to analyze connectivity to subcortical structures more broadly. We first created a 4-mm radius sphere centered on the MNI coordinates identified in Table 3 and found the vertex on the MNI152 pial surface that is closest to the spherical seed. We then smoothed this single vertex by 4 mm on the surface and mapped the resulting cortical label to each individual subject’s cortex. The individual cortical label was projected back into the subject’s native volumetric space to calculate the averaged time series within the seed. For the subcortical seed (dAmy), we directly projected the spherical seed into each subject’s native volumetric space and extracted its time course. On the subject level, we ran a voxel-wise regression on left and right hemispheres of MNI152 and subcortical volume of MNI305 to compute the partial correlation coefficient and correlation effect size of the seed time series, taking into account several nuisance variables: cerebrospinal fluid signal, white matter signal, motion correction parameters, and a 5^th order polynomial. On the group level, we concatenated the contrast effect size maps from all subjects and ran a general linear model analysis to test if the group mean differed from zero. This yielded final group maps that showed regions whose fluctuations significantly correlated with the seed’s BOLD time series.

To estimate cortical-subcortical connectivity, we used a more liberal statistical threshold compared to the analyses of corticocortical connectivity. The smaller size of subcortical regions, as well as their anatomical placement, renders their signal noisier and less reliable⁵⁷, yielding relatively smaller estimates of intrinsic connectivity. Thus, guided by classical measurement theory¹⁴⁸, we relied on replication to determine which connectivity values were meaningful.

K-means cluster analysis of discovery maps

First, we computed the 8x8 η² similarity matrix for each pair of maps⁴⁹. Based on visual inspection of the eight maps, we used K-means clustering with k=2 and k=3 using the kmeans function in MATLAB (Mathworks, Natick, MA). Our results confirmed that k=2 captured the default mode versus salience distinction across these maps, whereas k=3 further divided the ‘salience cluster’ into two sub-categories depending on whether or not somatosensory cortices are included. Because sub-categories within the salience network were not important to our study goals, we used the k=2 cluster solution.

Identification of the interoceptive system networks

We confirmed that Network 1 is the established default mode network (for a review, see ⁵⁰) and Network 2 is the established salience network^51,52. The reference maps were constructed using coordinates obtained from Yeo, et al.⁵⁵ as follows. Using a random sample of N=150, we created a mask of the default mode network by conjoining functional connectivity maps from two hubs in the default mode network⁵⁵: a 4-mm seed at the dorsomedial prefrontal cortex (MNI 0, 50, 24) and a 4-mm seed at the posterior cingulate cortex (MNI 0, −64, 40). We likewise created a mask of the salience network by conjoining functional connectivity maps from two bilateral hubs in the salience network (labeled as the ventral attention network in Yeo, et al.⁵⁵): 4-mm seeds at the left and right supramarginal gyrus (MNI ±60, −30, 28) and 4-mm seeds at the left and right anterior insula (MNI ±40, 12, −4). We thresholded our maps to p < 10⁻⁵ uncorrected (as in all our analyses) and we thresholded the default mode and salience networks to z(r) > 0.05 where z is the Fisher’s r-to-z transformation. We then calculated the percent of each established network (default mode or salience) that covered each of our networks (Network 1 or 2), and the complementary measure: the percentage of each of our networks (Network 1 or 2) that covered each established network (default mode or salience). These calculations used only the right hemisphere.

Reliability analyses

We used η² as an index of reliability because it shows similarity between maps while discounting scaling and offset effects⁴⁹. An η² value of 1 indicates spatially identical maps, while an η² value of 0.5 indicates statistically independent maps. For each of our eight cortical and amygdalar seeds, we calculated η² between the discovery and replication samples using the effect size (gamma) maps generated by the group-level general linear model analysis. Then we calculated the mean and SD of the eight η² values across all seeds to index overall similarity between samples. This was done separately for the cortical and subcortical maps. We repeated the same procedure to compare the reliability between the discovery and validation samples.

Analysis of the evocative images task

We analyzed electrodermal activity data using Electrodermal Activity Analysis v3.0.21 (Mindware). For each 6-second trial when the photo was visible, we measured the number of event-related skin conductance responses (SCRs) according to best practices¹⁴⁹. We considered an SCR to be event-related if both the response onset and peak occurred between 1 and 6 seconds after stimulus onset, with an amplitude ≥0.01 μS. It is commonly observed that a substantial proportion of healthy adults produce relatively few if any SCRs¹⁵⁰. We disqualified 16 of our 66 participants (24%) because they generated event-related SCRs during fewer than 5% of the evocative photo trials. We analyzed our data using the number of SCRs (as opposed to the amplitude of the SCRs) per prior work from our group (e.g., ¹⁵¹) and others (e.g., ¹⁵²).

Multilevel linear modeling to assess correspondence between objective physiological and subjective arousal during an allostatically relevant task

We used HLM v7.01 with robust parameter estimates (Scientific Software International; Skokie, IL). Level-1 of the model estimated the linear relationship (slope and intercept) between physiological arousal (number of event-related SCRs) and subjective arousal (1=“Very calm” to 5=“Very activated”) in response to each of ninety photos. Thus, the model was adjusted for mean individual reactivity. Level-2 estimated the extent to which intrinsic connectivity between viscerosensory and visceromotor regions (e.g., dpIns-aMCC) moderated the relationship between objective and subjective arousal (i.e., moderated the slope of the Level 1 model). All variables were unstandardized. Level-1 variables were group-mean centered (for each participant) and Level-2 variables were grand-mean centered (across participants).

Data availability

The data that support the findings of this study are available from the corresponding author upon request.

Code availability

The code to analyze data are available from the corresponding author upon request.