Evidence for domain-general arousal from semantic and neuroimaging meta-analyses reconciles opposing views on arousal

Significance

Arousal is a central concept linking brain and body, often put forward to explain motivated behavior. Although the term is widely used, its definition remains elusive, with diverging views in classical textbooks. Is arousal an abstraction of the mind reflecting a myriad of distinct biological processes, or is there some common neuronal feature? We used large-scale text mining methods and neuroimaging meta-analyses to reveal the existence of domain-general arousal, a cortical process shared by situations as different as a cognitive task, an emotional context, the transition to wakefulness or sexual behavior, involving notably a specific area of the anterior insula in humans. Our results pave the way for refined taxonomies of arousal, where concepts and neural data are well aligned.

Abstract

Arousal refers to changes in brain-body state underpinning motivated behavior but lacks a proper definition and taxonomy. Neuroscience and psychology textbooks offer surprisingly different views on what arousal is, from a global brain-wide modulation of neuronal activity to a multidimensional construct, with specific brain-body patterns tuned to a given situation. The huge number of scientific articles mentioning arousal (~50,000) highlights the importance of the concept but also explains why such a vast literature has never been systematically reviewed so far. Here, we leverage the tools of natural language processing to probe the nature of arousal in a data-driven, comprehensive manner. We show that arousal comes in seven varieties: cognitive, emotional, physiological, sexual, related to stress disorders, to sleep, or to sleep disorders. We then ask whether domain-general arousal exists at the cortical level, and run meta-analyses of the brain imaging literature to reveal that all varieties of arousal, except arousal in sleep disorders for lack of data, converge onto a cortical network composed of the presupplementary motor area and the left and right dorsal anterior insula. More precisely, we find that activity in dysgranular insular area 7 (Jülich atlas), the region with the highest convergence across varieties of arousal, is also specifically associated with arousal. The domain-general arousal network might trigger the reorganization of large-scale brain networks—a global mechanism—resulting in a context-specific configuration—in line with the multidimensional view. Future taxonomies of arousal refining the alignment between concepts and data should include domain-general arousal as a central component.

Arousal is a central concept in neuroscience, referring to fluctuations in brain-body states during wakefulness, underpinning motivated behavior and influencing performance. Arousal is very often put forward as an explanatory variable for neuronal activity, bodily physiological parameters, and behavior in basic, translational, and clinical research, in both humans and animals: Almost 50,000 scientific articles mentioning arousal have been published. Given the importance of the concept, there have been surprisingly few attempts at establishing a taxonomy of arousal. Textbooks (1, 2) agree that arousal refers to brain-body states related to wakefulness and fluctuations within the wake state, that arousal modifies both central and autonomic nervous system activity, and that it is expressed in a myriad of bodily parameters. However, textbooks disagree on whether arousal is a unitary concept. On the one hand, in Principles of Neural Science (2021, p. 1025), arousal is defined as a global state: “[…] which ranges from excitement and vigilance to drowsiness and stupor.” Arousal can indeed appear as a global factor organizing behavior. The Yerkes–Dodson law (3, 4) states that both low and high arousal drive performance down, while intermediate arousal corresponds to optimal performance. The prevalence of the inverted-U-shape relationship between arousal and behavioral performance across many different sensory and cognitive tasks in both humans and animals suggests some underlying common neural principle (5). In the global view, arousal appears as an indiscriminate driving force acting on all cortices, whether sensory, associative, emotional, or motor. On the other hand, as advocated in the Handbook of Psychophysiology (2017, p. 415), the notion of global arousal should be rejected: “data […] were indeed incongruent with the idea that arousal is a unitary process,” and “arousal likewise became understood as a multidimensional process or construct.” Indeed, arousal arises from different ascending neuromodulatory systems (e.g., cholinergic, noradrenergic, etc.); it translates into different physiological changes (e.g., pupil diameter, heart-rate, skin conductance, etc.) that sometimes correlate only weakly with each other, or with subjective arousal ratings (6). In the multidimensional construct view, arousal is split into distinct varieties, such as emotional arousal vs. cognitive arousal, that represent distinct neurophysiological entities. Reconciling such divergent views calls for refining definitions and concepts to establish a taxonomy of arousal that aligns with underlying biological mechanisms.

Articulating brain systems and mechanisms with behavior and mental functions is the golden aim of cognitive neuroscience, and whether to proceed from brain to functions or the other way round is subject to epistemological debate (7, 8). A more pragmatic stance recently appeared (9, 10) where the automatic processing of the wealth of data accumulated over the years allows to examine how well terms and concepts align with neural data. We adopt this stance to ask whether different varieties of arousal share a common biological substrate at the cortical level. The motivation for this study was thus to lay the initial grounds for a data-driven taxonomy of arousal.

This approach requires determining how many varieties of arousal exist in the literature. A few attempts at synthetizing the literature pioneered the field (11, 12), but given the number of articles on arousal, their scope was necessarily limited by a priori selections of both articles and varieties of arousal. Here, we leverage the tools of natural language processing (13) and graph theory to comprehensively analyze this large and heterogeneous literature. We identify in the literature seven different varieties of arousal in a purely data-driven manner. We then conduct meta-analyses (14, 15) of the brain imaging literature in humans and find that different varieties of arousal share common cortical resources.

Results

Identifying Domains of the Arousal-Related Literature Using Semantic Analysis.

The arousal literature is organized in seven semantic communities corresponding to known fields of research.

We retrieved the 49 525 scientific references returned by the exact query “arousal” from Web of Science and PubMed and show that those articles are organized in seven meaningful semantic communities (Fig. 1). More precisely, from the 49 525 abstracts we automatically extracted terms, i.e., associations of two to three words, and selected the most frequent ones. We then excluded nonspecific terms (e.g., “significant result” or “previous research”; see SI Appendix, Table S1), as well as all brain-related terms (e.g., “brain imaging” or “prefrontal cortex”), and finally retained the remaining 1,200 most frequent terms to build a graph reflecting the semantic content of the arousal literature. In this graph, terms are node, connected by edges that reflect how often terms co-occur in the references. Using the elbow method (16), we found that seven semantic clusters optimally describe the graph. A post hoc silhouette analysis (17, 18) confirmed this number and further suggests a more refined partition in twelve clusters that was not further explored (SI Appendix, Fig. S1; for details, see the Materials & Methods section).

Fig. 1.

The semantic graph of arousal-related literature. (A) Schematic description of the semantic analysis workflow. From a dataset of 49 525 references containing the word arousal, we extracted the terms (association of two or three words) that appeared most frequently and computed the co-occurrence of these terms. Subsequently, the co-occurrence matrix is embedded into a semantic graph. In the semantic graph, terms that tend to co-occur, and to co-occur with the same other terms, are closer, forming a semantic community. (B) The semantic graph of the arousal literature constructed on the 1,200 most frequent terms. Each dot corresponds to a term, each edge corresponds to semantic similarity. Nodes that cluster together are assigned the same color. The four most frequent terms are listed for each community in gray font and were used to assign a label to each community (color, bold font). More details can be found in SI Appendix, Tables S2 and S3. (C) Physiological profiles of the semantic communities, showing the percentage of articles within each semantic community using terms related to eye-tracking/pupillary (label: pupil), cardiac (heart), respiratory (respiration), blood pressure (blood), or skin conductance (EDA) measures of arousal. Abbreviations: EDA - electrodermal activity.

The semantic graph of the arousal literature reveals seven communities (Fig. 1B) that we associate with seven varieties of arousal: arousal in cognition, emotion, physiology, stress disorders, sex, sleep, and sleep disorders. This association is based on the most frequent terms (Fig. 1B and SI Appendix, Table S2) as well as on the terms with the highest number of connections within each semantic community (SI Appendix, Table S3). We further extracted terms describing the type of paradigm or task used (SI Appendix, Table S2). Such terms were relatively frequent in some communities (e.g., cognition) but totally absent in others (e.g., sexual arousal, sleep, or sleep disorders). While we illustrate the results obtained with the analysis of the 1,200 most frequent terms, the seven communities and their respective locations in the graph are stable across a wide range of number of terms included in the analysis (from 300 to 2,285, SI Appendix, Fig. S2). The physiological arousal community is located at the center of the overall semantic space, the two arousal in sleep communities are close together, arousal in stress disorders sits next to emotional arousal, and sexual arousal is far away from all other semantic varieties of arousal.

Each semantic community has a specific physiological measure profile.

To cross-validate the existence of seven different varieties of arousal as identified by the semantic graph, we reasoned that if the semantic communities indeed refer to distinct fields of research, then they should differ in the measures used to operationalize arousal. We thus quantified the representation of physiological measures in the articles contained in each semantic community. We first assigned each article to semantic communities: If a term from a given community was indexed in the abstract, the article was considered belonging to this community. An article might thus project onto more than one semantic community. We inspected the occurrence of the words related to five common physiological measurements used as arousal proxies: cardiac activity, eye-tracking, respiration, skin conductance, and blood pressure (dictionary of terms in SI Appendix, Table S4).

The resulting “physiological measure profiles” (Fig. 1C) constitute distinctive signatures of how each semantic community operationalizes arousal. Measures of respiration dominate in the two sleep communities, cardiac measures are dominant in the stress and physiology communities, while pupillometry is almost exclusively used for the study of cognitive arousal. Altogether, each arousal community uses a specific combination of bodily measures of arousal. Interestingly, although pupillometry is an increasingly popular measure of arousal especially in animal studies (19), eye-tracking is not a dominant one in any of the communities. It is worth underlining that we obtain this result even though an article and its associated physiological measures could be assigned to more than one semantic community. In other words, across fields of research, arousal is associated with distinct physiological measure profiles. This result cross-validates our approach of categorizing arousal literature, as it retrieves field-specific experimental preferences.

We further analyzed whether some communities studied more phasic or tonic arousal. The terms phasic or tonic seldom appear in the Abstracts (in about 2% of the articles) and all communities seem to equally focus on either phasic or tonic arousal (SI Appendix, Table S5).

Neuroimaging Meta-Analysis: Identifying Brain Regions Associated with Arousal across Fields of Research.

The arousal cortical network: Three cortical regions where different varieties of arousal converge.

The semantic analysis reveals that arousal is a term employed in fields that differ by their object of study (i.e., sleep vs. emotion) and in how they measure arousal (i.e., pupil diameter vs. respiration). Is there nevertheless a shared neuronal substrate for arousal, common to the seven varieties of arousal we identified, or are the varieties of arousal so different that they share no common neuronal substrate?

To answer this question, we first ran an Activation-Likelihood Estimation (ALE) (14, 15) analysis of the existing human brain imaging literature, for each field of research separately. Here, the rationale was to give equal weight to each variety of arousal, even if the number of brain imaging studies in some fields (e.g., sleep or sexual arousal) is likely to be much smaller than in others (e.g., emotional or cognitive arousal).

From the 415 neuroimaging studies available in the open-source NeuroQuery (20) database, we excluded studies using a priori regions of interest, since regions of interest constitute a known biasing factor in meta-analysis (21, 22). Out of the remaining 228 studies, 221 (corresponding to an estimate of 6,647 participants) could be reliably assigned to at least one semantic community when using 1,200 terms to generate the semantic graph. Except for sleep disorders, where the number of available studies (n = 6) is very low, we can identify a network associated with arousal in each field of research, as shown in Fig. 2. Some arousal-related regions are found selectively in some fields of research but not others—for instance, the left lateral occipital cortex is consistently found in emotional and sexual arousal but is not observed for other varieties of arousal. Still, there is convergence of four or five varieties of arousal in several regions.

Fig. 2.

Meta-analysis of arousal in the seven varieties or arousal and their overlap. Results of the ALE meta-analysis for each semantic community (semantic analysis with 1,200 terms), projected for visualization on inflated surfaces (voxel-level cluster forming P < 0.001 uncorrected; cluster level P < 0.05 corrected for multiple comparisons using Monte-Carlo permutations). The “n” indicates the number of brain imaging studies included in each meta-analysis. Note that because an article could be assigned to more than one semantic community, the sum of the number of imaging studies included in each meta-analysis is larger than 221. The number of available imaging studies for the sleep disorders community (n = 6) was too low to obtain significant results. The Bottom row shows the overlap across activations obtained for each variety of arousal.

Our results reveal the existence of a cortical arousal network, composed of three regions where four or five different varieties of arousal converge (Fig. 2, Bottom row): the right and left anterior insula, and presupplementary motor areas (preSMA). As detailed in Table 1, cognitive, emotional, and physiological arousal converge on all three nodes of the cortical arousal network, but left and right anterior insula differ when it comes to other types of arousal: Sexual arousal and arousal in sleep appear in the right anterior insula, and arousal in stress disorders in the left anterior insula. Sexual arousal is absent from preSMA, and arousal in stress disorders and arousal in sleep overlap only marginally within this node, arousal in sleep being more anterior than arousal in stress disorders in preSMA. These results hold true even when considering unthresholded meta-analytical results (SI Appendix, Fig. S3). In other words, all six varieties of arousal for which enough data exist load, sometimes with different weights, on the three nodes of the cortical arousal network.

Table 1.

Relationship between the seven different varieties of arousal and the three nodes of the arousal cortical network, in right and left anterior insula (AI) and preSMA

	right AI	preSMA	left AI
MNI (x,y,z) mm	36, 21, 2	−3,12, 48	−35, 21, 4
cognitive	+	+	+
emotional	+	+	+
physiological	+	+	+
in stress disorders	−	+	+
sexual	+	−	−
in sleep	+	(+)	−
in sleep disorders	not enough data

Coordinates correspond to center of mass. The +symbol indicates the presence of significant activation and the—symbol the absence of significant activation (voxel-level cluster forming P < 0.001 uncorrected; cluster level P < 0.05 corrected for multiple comparisons using Monte-Carlo permutations). The presence of activation related to arousal in sleep in preSMA is in parenthesis because this result was not stable when varying the granularity of the semantic analysis.

The organization of data in semantic space and of data in neuronal space shares some interesting similarities. The three varieties of arousal (cognitive, emotional, and physiological) that are present in all the three nodes of the cortical arousal network are also close in semantic space. Arousal in sleep and arousal in stress disorders are far away from each other in semantic space, and share only the preSMA node in neuronal space. Sexual arousal, which is far from all other varieties of arousal in semantic space, differs most from the other forms of arousal in cortical space as well, since it shows up in right dorsal anterior insula only.

In the results presented above, we excluded results coming from a priori regions of interest. When including these results in the meta-analyses, results are overall similar (SI Appendix, Fig. S4), but with two noticeable differences. First, there are now more studies available for the field of arousal in sleep disorders (n = 16), which reveals an activation in left anterior insula. Second, a significant overlap across six types of arousal (all but arousal in sleep disorders) is now found in the bilateral anterior insula as well as the amygdala, covering all amygdala nuclei and some neighboring areas such as the hippocampus or entorhinal cortex (SI Appendix, Fig. S4).

The left and right anterior insula nodes of the cortical arousal network are both robust to the parameters of the semantic analysis and specific to arousal.

We then tested whether our results were robust against changes in the parameters of the semantic analysis. The results presented above were obtained when including 1,200 terms in the semantic analysis. We already showed that when varying the number of terms (or granularity) used for semantic analysis, we retrieve the same number of semantic communities with similar overall spatial organization (SI Appendix, Fig. S2). However, increasing granularity induces two changes. First, terms which are not central to a given community but rather at the border might be reassigned to a different community. Second, with more terms in the semantic analysis, a given experimental study is likely to match more terms. Because those terms might belong to different semantic communities, a given experimental study might be assigned to more semantic communities when the granularity of the analysis increases (SI Appendix, Fig. S2, last column). To examine these effects, we repeated the analysis described above and varied the number of terms used (1,200, 1,500, 2,000, or 2,285). We then computed in which regions there was an overlap of brain activation across four or five types of arousal, across all semantic granularity levels (Fig. 3A and SI Appendix, Fig. S5). We found that the association between left and right insula with different types of arousal was highly consistent across different granularity levels of the semantic analysis (Fig. 3B). In preSMA, there was a consistent overlap of cognitive, emotional, and physiological arousal as well as of arousal in stress disorders; however, the link between arousal in sleep and preSMA was not stable (Table 1).

Fig. 3.

Robustness and specificity of the cortical arousal network. (A) Overlap of the different communities (as last row in Fig. 2) but for increasing levels of granularity of semantic analysis. (B) Surface representation of the network robust to parameters of semantic analysis. The green color denotes areas where we found an overlap of four, and yellow an overlap of five, semantic communities across all levels of granularity of the semantic analysis. (C) Specificity of each node of the robust network to arousal. For each node: left AI, right AI and preSMA, we tested how likely an activation in the region was in the nonarousal literature (gray histogram, chance level distribution) as compared to the arousal literature (colored bar). Abbreviations: AI (anterior insula), preSMA (presupplementary motor area).

Finally, we tested whether the activation of the three nodes of the arousal cortical network was specifically associated with arousal. Indeed, both the anterior insula (23–25) and the preSMA (26) (often labeled dorsal anterior cingulate cortex or dorso-medial prefrontal cortex) are notoriously often activated in any paradigm, including in paradigms not particularly designed to target arousal. We thus computed the level of meta-analytic activation in those three nodes when randomly selecting brain imaging studies not using the term arousal, and compared it to the level of activation in brain imaging studies with the term arousal (Fig. 3C). In practice for each node, we computed a statistic reflecting activation likelihood across the 415 arousal related studies (Materials & Methods). We then computed the same statistic in 415 studies not related to arousal randomly drawn from the NeuroQuery database, and repeated the process 10,000 times to obtain a chance level distribution. The comparison between the chance level distribution and the original value in arousal data provides the probability that the activation likelihood in the original arousal data could be obtained by chance. Because this analysis includes studies using a region of interest (ROI) approach, results might be slightly distorted, and P-values are only indicative. Still, there is a marked difference between preSMA, where the arousal-related results correspond to what could be expected when selecting studies randomly (Fig. 3C), and the left and right anterior insula arousal-related nodes, which appear to be significantly (left AI, Monte-Carlo P = 0.0065) or marginally [right AI(5), P = 0.082, right AI(4), P = 0.05971] specific to arousal. In other words, activation in the preSMA node is not specific to arousal: A similar level of activation could be obtained for other sets of human brain imaging data not mentioning arousal. However, the activation of the left and right anterior dorsal insula, at the location where varieties of arousal converge, is unlikely to be produced by brain activation reported in articles not mentioning arousal, i.e., we find a specific association between arousal and left and right dorsal anterior insula.

We conclude that the most robust and specific association with arousal is found in the left and right anterior insula. The arousal-related nodes in the anterior insula were anatomically symmetrical (Fig. 4 A and B), although they correspond to the convergence of different types of arousal (Table 1). We further characterized the insular arousal hubs. Anatomically, the insular arousal hubs were confined to the anterior short gyrus (27) (Fig. 4C). Taking advantage of the detailed anatomical parcellation of the insula in the Jülich atlas (28–30), we find the insular arousal hub to be located in a dysgranular dorsal anterior territory, centered in area Id7 and extending dorsally in Id6 as well as marginally ventrally in Id8 (Fig. 4B). In terms of resting state networks, the dorsal insula arousal hubs belong to the Ventral Attention Network (31) (Fig. 4D). They overlap more specifically with the typical seed used for the cingulo-opercular (COP) network (32), but are more dorsal than the seed region of the homeostatic salience (HS) network (33). The dorsal location of the insula arousal hubs is confirmed when comparing our results with the tripartite organization (dorsal cognitive, ventral emotional, and posterior sensorimotor) of the insula based on both resting state functional connectivity (34) and meta-analytic task-related activations (35) (Fig. 4E). Finally, the arousal hubs in dorsal anterior insula are included in a larger region related to predominantly parasympathetic regulation of the autonomic nervous system during tasks, and are more dorsal than the predominantly sympathetic regulation (36) (Fig. 4F).

Fig. 4.

Dysgranular area 7 in dorsal anterior insula is a key hub for arousal. (A) Volume representation of the robust and specific association with arousal in left and right anterior insula. (B) Surface representation of the right and left anterior insula with description of the cytoarchitecture defined in the Julich-Brain (28–30). (C) Location of the most robust node in the right anterior insular cortex, with respect to the anatomical parcellations of Faillenot et al. (27). (D) Resting state connectivity networks as defined in ref. 31, and seeds used to define COP network in ref. 32, and the HS network in ref. 33, as well as (E) functional parcellation from ref. 34 with functional description from Kurth et al. (35), and (F) contours of sympathetic (EDA) and parasympathetic (HRV) task-evoked changes as described in the meta-analysis of Beissner et al. (36). Abbreviations: Id (dysgranular insula); ASG (Anterior Short Gyrus), MSG (Middle Short Gyrus), PSG (Posterior Short Gyrus), AIC (Anterior Inferior Cortex), ALG (Anterior Long Gyrus), PLG (Posterior Long Gyrus), dAI (dorsal Anterior Insula), PI (Posterior Insula), vAI (ventral Anterior Insula), COP (cingulo-opercular), HS (homeostatic salience), HF-HRV (high frequency heart rate variability), EDA (electrodermal activity).

Discussion

We identify in a purely data-driven manner seven varieties of arousal: cognitive, emotional, physiological, and sexual arousal, as well as arousal in sleep, in sleep disorders, and in stress disorders. The seven varieties of arousal differ in semantic space, in the physiological variables used to measure arousal, and in specific cortical correlates. Yet, except for arousal in sleep disorders where data are scarce, all varieties of arousal converge, albeit with different weights, onto three nodes of a cortical arousal network composed of left and right anterior insula and preSMA. The amygdala is likely to be part of this network, but because it is most often isolated through a ROI analysis likely to distort meta-analyses (21, 22), we limit our conclusions to cortical regions. Our results show that even if the different varieties of arousal refer to markedly distinct domains, there is enough commonalities in the underlying phenomenon for neuronal data to converge on the same cortical arousal network. Among the nodes of the arousal network, the ones located in left and right dorsal anterior insula were not only robust against changes in the granularity of the semantic analysis, but they were also specific to arousal. These hubs of convergence of different arousal varieties were located in the anterior short gyrus of the anterior insula, specifically in dysgranular area Id7. While cognitive, emotional, and physiological arousal were present in both left and right dorsal anterior insula Id7, sexual arousal, and arousal in sleep were right lateralized, and arousal in stress disorders was left lateralized. Our results thus indicate that a common network underlies the notion of arousal across very different domains, and reveal the central position of dysgranular insula area Id7 (Jülich atlas) in domain-general arousal. Combining those results with anatomo-functional knowledge retrieved from the literature, as developed below, offers a framework to understand how the global and multidimensional construct views of arousal can be reconciled.

Situating the Domain-General Cortical Arousal Network in the Brain Network Literature.

Arousal refers to a brain-body state involving ascending neuromodulatory systems (37). It is beyond the scope of this article to review such systems in depth, especially since neuromodulatory subcortical structures, that are only beginning to be imaged in humans (see e.g., refs. 38–40), are absent from the current meta-analysis, which is limited to cortical structures. Still, recent results in humans (38) show that the three nodes of the cortical arousal network we identify here are functionally connected to ascending neuromodulatory systems. More precisely, the short gyrus of the anterior insula, that includes the robust and specific node of the domain-general arousal network, insular dysgranular area 7 (Jülich atlas), is functionally particularly well connected (38) with cholinergic (several reticular nuclei), noradrenergic (locus coeruleus, subcoeruleus), and serotoninergic (dorsal raphe) nuclei, and expresses receptors to all three neuromodulators (41). However, projections from subcortical neuromodulatory nuclei are spatially distributed in numerous cortical regions (38), way beyond anterior insula and preSMA. The domain-general cortical arousal network we reveal here thus represents only a small subset a much larger ensemble of cortical regions connected to neuromodulatory nuclei. The domain-general cortical arousal network also overlaps with regions covarying with bodily measures of autonomic activity during tasks (36), in line with the notion that arousal reflects a mobilization of brain and body resources. More precisely, the preSMA node covaries with measures of sympathetic activity and the dorsal insula nodes covary with high-frequency heart rate variability, a measure of parasympathetic activity. However, the cortical autonomic network (36) is much larger than the arousal network we reveal here. The domain-general cortical arousal network can thus be viewed as a subset of two much larger cortical networks: the targets of ascending neuromodulatory systems, and regions associated with autonomic regulations.

The nodes of the domain-general arousal network are part of a large resting-state network often referred to as the ventral attention network (31) or salience network (42). Some authors further subdivide this network in two, the cingulo-opercular (COP) network (32), and the homeostatic salience (HS) network (33, 43). Within the anterior insula, the HS network (33, 43, 44) is typically more ventral than the arousal hub we identified. Our results appear better aligned with the COP network, which is more dorsal in the anterior insula. The COP network has been related to task set maintenance (32), and to a state of tonic alertness, or “self-initiated (rather than externally driven) preparedness to process and to respond” (45). While the anatomical distinction between HS and COP networks in the anterior insula is still debated (42), both networks are part of the overall salience network and considered to play a general role in regulating the integration or segregation of other brain networks (46). Interestingly, the reconfiguration of brain networks triggered by the salience network has recently been put forward as an arousal-related mechanism at the cortical level (47–50), in addition to, or in coordination with, the activity of neuromodulatory subcortical nuclei. Our results allow to generalize this conclusion to all varieties of arousal.

The arousal network is composed of brain regions known to be very often activated irrespective of the specific nature of the task (23, 25, 26). In the present study, tasks are often reported for some varieties of arousal (i.e., cognition), but not in others (i.e., sleep), making it unlikely that the convergence we observe is merely due to the presence of a task. We also explicitly computed the probability that each of three nodes would appear in any randomly chosen subset of the brain imaging literature. We indeed find that the arousal node in preSMA is not specific to arousal. Such a finding might be interpreted in two different ways: One is that arousal, construed as a mobilization of brain and body resources, is an underlying factor in most if not all brain imaging study, even when arousal is not explicitly studied. Alternatively, it could be that the association between preSMA and arousal in our results is driven by another factor common to both arousal and nonarousal studies. The present results do not allow to disambiguate between these two interpretations in preSMA. However, in the case of the dorsal anterior insula nodes of the arousal network, the interpretation is much more straightforward: The dorsal anterior insula nodes were more active in studies explicitly mentioning arousal than in random sets of studies not using the term arousal, indicating a specific association between the dorsal anterior insula nodes and arousal.

The Robust and Specific Nodes of the Arousal Network: Id7 in the Dorsal Anterior Insula.

Dysgranular area Id7 (Jülich atlas) in the left and right dorsal anterior insula appears as a key hub for arousal for three reasons. First, it reflects the convergence of the highest number of varieties of arousal (cognitive, emotional, physiological, sexual, sleep related in the right hemisphere; cognitive, emotional, physiological, in stress disorders in the left hemisphere). Second, this convergence was robust to changes in the granularity of the semantic analysis. Last but not least, as detailed just above, activity in this region is specific to arousal—in other words, this region is more likely to be activated when arousal is mentioned in the article than when arousal is not mentioned. Area Id7 has only been recently anatomically identified in humans (27) and functional studies are scarce, but it constitutes a promising seed for large-scale brain analysis (44, 51) to refine the definition and mechanisms of domain-general arousal.

Anatomo-functional specification within the insula, while recognized as important (23, 52), has often been limited in the experimental literature in humans to a rough partition into three regions, typically posterior, dorsal anterior, and ventral anterior (34, 53). Recently however, anatomical studies in humans (29) and nonhuman primates (54) revealed a much finer-grained organization. Anatomical specification is all the more important that the anterior insula is disproportionately enlarged in humans compared to other species (55). To the best of our knowledge, whether the human Id7 has a homolog in nonhuman species is not yet known. Functionally, the electrical stimulation of the posterior insula in human patients can elicit various physiological changes and/or somatovisceral sensations (56, 57). In contrast, the stimulation of the anterior part of the insula (57), and specifically of the right anterior insula area Id7 (58), does not elicit any noticeable response in human patients at rest. This suggests that Id7 is neither a primary viscero-sensory nor viscero-motor node. However, Id7 becomes active in a simple visual oddball task, with infrequent nontarget visual scenes eliciting large neuronal responses (58), suggesting this area is able to keep track of task structure to detect deviant stimuli.

Hemispheric Lateralization of Id7 in Dorsal Anterior Insula.

We observe some hemispheric specialization for arousal in dorsal anterior insula: While cognitive, emotional, and physiological arousal are bilateral, arousal in stress disorders appears in left Id7 and sexual and sleep-related arousal in right Id7. While a functional asymmetry is often observed during tasks between the left and right dorsal anterior insula (see e.g., ref. 59), there seems to be no marked lateralization in either cytoarchitectonic organization (60) or cortical resting-state functional connectivity (25). However, both functional connectivity (61) and tractography (62) data suggest a lateralization of the connectivity between anterior insular cortices and basal ganglia.

Arousal in stress disorders is the only variety of arousal not present in the right insula. The fMRI studies of arousal in stress disorders included in the meta-analysis cover quite a broad range of disorders, including anxiety and depression but also bipolar disorder, borderline personality disorder, obsessive compulsive disorder, etc. Despite the heterogeneity of the disorders, all those studies converge in the preSMA and left insula, which shows that they have enough in common to activate two nodes of the cortical arousal network. However, they fail at providing any evidence, even at the subthreshold level, for an activation in the right anterior insula. One finding that seems recurrent in the stress disorder literature is that there is a partial decoupling between subjective reports and peripheral measures of arousal in patients, with for instance limited objective signs of autonomic hyperarousal at rest in chronic anxiety patients (63), and only a weak association between subjective reports of sleep disturbance and objective polysomnographic measures of sleep in posttraumatic stress disorder (64). A better understanding of the functional lateralization of arousal in dorsal anterior insula might thus require refining how arousal is measured, objectively or subjectively, with potential clinical implications.

Conclusion: Domain-General Arousal Is a Data-Driven Concept That Could Reconcile the Global State View with the Multidimensional Construct View.

We show that arousal in situations as different as a cognitive task, an emotional context, the transition to wakefulness, or sexual behavior, overlap in a shared domain-general arousal cortical network. Domain-general arousal is thus a concept aligned with neural data and hence should become a solid anchor point for any future taxonomy of arousal.

The most robust and specific nodes of the domain-general arousal network in the dorsal anterior insula Id7 (Jülich atlas) are known to trigger large-scale cortical network reconfiguration. This result reconciles the seemingly opposing views on arousal. The most likely function of the domain-general arousal network is to coordinate the overall configuration of other brain networks—in that sense, it is a global mechanism, that affects the entire cortex. But the pattern resulting from the reconfiguration is likely to be highly specific to a given experimental situation. Indeed we find that each variety of arousal has a distinctive cortical signature—in that sense, arousal is also specific and has multiple dimensions.

Materials & Methods

Semantic Analysis.

Data collection.

We collected the 122 590 records returned by the exact query arousal searched in all available fields from Web of Science (46 924, search on May 5th, 2023) and Pubmed [53 211, search on April 10th, 2023 using Pubmed Entrez Direct command line utility (https://ftp.ncbi.nlm.nih.gov/entrez/entrezdirect/)]. Note that by default the Pubmed API extends queries by semantic similarity (i.e., the query arousal extends to “wakefulness”), but here we enforced the exact query mode. We excluded records that were missing data in any of the following fields: “abstract,” “title,” “PubMed ID,” “language,” “document type,” or “publication date.” We then selected records corresponding to original research articles written in English. Finally, we removed duplicates based on PubMed ID. This resulted in a dataset composed of 49 525 unique records.

Semantic network.

To explore the semantic landscape of the arousal literature we used the bunkatech library (65) designed for Natural Language Processing applications. The bunkatech term extractor uses Textacy (https://textacy.readthedocs.io/en/latest/), a library built on top of Spacy Deep Learning Language model (https://spacy.io/models/en). The rationale for choosing parameter values in this analysis was to capture multiple levels of contextualization and nuances in the dataset, i.e., using large number of terms and large number of projection dimensions, and then verify that the coarser structure, at the level of clusters in semantic space, was stable across different granularity levels.

We first recovered the 3,500 terms most commonly occurring in the dataset. We found that beyond 3,500, extracted terms were related predominantly to scientific methods rather than the semantics of the research questions contained in the abstracts, and hence were not useful for the present study.

The term extractor was parameterized to extract the most frequent terms composed of 2 or 3 words, to include nouns, proper nouns, and adjectives. A number of terms were excluded from further analysis (see SI Appendix, Table S1 for exclusion criteria) because they were too general (i.e., “control group” or “significant results”). We also removed all brain-related terms like “brain imaging,” “prefrontal cortex” etc., by matching with NeuroNames (66) dictionary and further manual exclusion. Term exclusion was performed manually by authors MS and CTB separately, disagreements were discussed until resolved, to obtain a final list of 2,285 terms, sorted by occurrence. The final list of the 2,285 terms retained is available on the data repository of the article.

The following steps were completed for variable numbers of terms (300, 600, 900, 1,200, 1,500, 2,000, 2,285). We selected the number of terms for network projections from the maximal number of terms available (2,285, maximal semantic granularity) down to a somewhat arbitrary small number (300 terms only, limited semantic granularity to describe almost 50,000 articles).

We first created a co-occurrence matrix of the terms, with co-occurrence being defined within the same entry in the dataset (i.e., terms appearing in the same abstract). We then computed the semantic similarity (also called semantic distance) between the terms using cosine similarity (computed on the co-occurrence), and retained the 15 closest neighbors of each term. The result is an edge table consisting of nodes (terms) and edges (cosine distance) between the nodes. We embedded the edge table into 700 dimensions using Node2Vec. We chose a larger than usual (128) number of dimensions for the projection in order to capture semantic relationships even with a large number (2,285) of terms. The analysis presented in SI Appendix, Fig. S2 shows that indeed with this number of dimensions we obtained embeddings that captured semantic relationships between terms at granularity sufficient to compare the embeddings between networks of different sizes. The nodes were then clustered using K-means. The “number of clusters” parameter was optimized using the elbow method as implemented in the Kneedle algorithm (16) and validated post hoc using the silhouette method (17) as implemented in the scikit-learn (18) library (SI Appendix, Fig. S1). Finally, the dimensionality was reduced with Uniform Manifold Approximation and Projection for Dimension Reduction (67) for visualization.

Article-cluster assignment.

We associated scientific articles with one of the seven semantic clusters based on the terms detected in the abstracts. If all the terms detected in the abstract belonged to one semantic cluster, the corresponding article was associated with this cluster. If an abstract was indexed with two terms belonging to two different semantic clusters, i.e., “heart rate” and “stress disorder,” the corresponding article was associated with two clusters, in this example “physiology” and “stress disorders”.

Semantic Prevalence of Physiological Arousal Measures.

To calculate the prevalence of the physiological measure of arousal in each cluster, we searched for words from a dictionary of related terms we created (SI Appendix, Table S4). We matched the words using simple python string matching by checking whether the term in the dictionary is present in the dataset item. The final result is the proportion of articles in which the relevant terms were found present over all the articles in the semantic cluster.

Neuroimaging Meta-Analysis—Activation Likelihood Estimation.

Overall rationale.

Because we review a vast and heterogenous literature, we use semiautomated tools for meta-analysis. Compared to more classical manual meta-analysis, an automated approach has the advantage of limiting errors in coordinate extraction and of collating data over larger numbers of studies (68). Imprecisions due to differences in stereotactic coordinate spaces or confusion between activations and deactivations were found to be limited in practice (69). Automated approaches nevertheless come with drawbacks, notably the fact that all reported coordinates are taken into account, whatever the task and the contrast. We detail below the specific pipeline for data selection, including manual intervention to apply some inclusion/exclusion criteria, as well as the different statistical tests performed.

Data Selection.

For the neuroimaging meta-analysis, we used the open-access neuroimaging peak-coordinate dataset available from NeuroQuery (20). We selected eligible studies from our arousal literature dataset by matching PubMed ID with the NeuroQuery database. We included only original studies reporting coordinates of brain activations, i.e., from the 429 studies available in the NeuroQuery database, we included electroencephalogray, magnetoencephalography, functional magnetic resonance imaging studies but excluded seven morphometry studies, one diffusion tensor imaging study, five meta-analyses, and one article that did not report any coordinates. Because NeuroQuery relies on automatic extraction of coordinates from tables, we further manually verified which of those 415 studies reported activation coordinates resulting of a ROI analysis in a table. This is an important step to reduce the bias for oversampling a given region (e.g., amygdala in emotion studies). For each paper, we assessed Abstract, Materials and Methods, Tables, and when necessary Results to select studies where coordinates in tables corresponded to results of whole brain analysis, or used functionally defined ROI. We identified 187 articles where tables reported coordinates of results based on a priori defined ROIs, leaving a dataset of 228 whole-brain neuroimaging studies. Most of those studies contained at least one of the terms of the semantic network and could thus be associated to at least one semantic cluster. The exact number of studies that could be associated with semantic clusters depends on the number of terms used in the semantic analysis, for instance with 1,200 terms in the semantic network 221 neuroimaging studies contained at least one term used in the semantic network and could thus be assigned to a least one semantic cluster. In summary, inclusion criteria were presence of the term arousal, study included in the NeuroQuery database, and brain activation study; exclusion criteria were meta-analysis, morphometry study, and presence of a table reporting coordinates for an a priori defined ROI.

Sample size.

We used the online application Elicit to retrieve information on sample sizes for the neuroimaging papers. For the articles for which the language model was not successful, we noted the sample size by hand.

ALE meta-analysis.

Coordinate-based quantitative meta-analyses of neuroimaging results were performed using NiMARE v0.0.11 (14, 70) leveraging its integration with the NeuroQuery database (20). ALE is based on computing consistent activation foci by modeling the probability distribution of activation at given coordinates against the null distribution of summary-statistic values and their expected frequencies under the assumption of random spatial associations between studies, via a weighted convolution. We ran a separate ALE meta-analysis for each semantic field, with the following parameters: fixed full-width half-max of 10 for Gaussian kernel, cluster forming uncorrected significance level of 0.001, and cluster-level inference with correction for multiple comparisons using family-wise error correction with Monte-Carlo permutations to identify brain areas consistently activated across articles of interest. The results of the meta-analyses were represented in MNI152 coordinates, that were then transformed into freesurfer (http://surfer.nmr.mgh.harvard.edu/) inflated surface representation using pysurfer (v0.11.0) (https://pysurfer.github.io/index.html) vol2surf function for visualization. The maps were thresholded for significance at one-sided Monte-Carlo alpha level = 0.05.

Convergence between research fields.

To identify structures common across the results of meta-analyses in each cluster, we overlapped the significance maps across semantic clusters and identified brain regions where the highest convergence between semantic fields was observed. To assess the robustness of resulting overlap foci against variations in semantic analysis parameters, we ran the fMRI meta-analysis following a semantic analysis with a variable number of terms: 1,200, 1,500, 2,000, or 2,285. We created an overlap map for each number of terms. To determine the probability that an overlap of 4 (or 5) semantic communities occurs at a given location irrespective of the granularity of the semantic analysis (robustness maps), we combined the overlap maps obtained from different number of terms in the semantic analysis. In practice, we created a mask for 4 (or 5) communities in overlap at each granularity level, and created a conjunction across granularity levels.

We report MNI152 coordinates of the center of mass [computed using ndimage.center_of_mass function from SciPy library (v1.11.2) (71)]. For visualization, we projected resulting maximal overlap foci to volume using the freesurfer surf2vol function and reported the MNI152 coordinates. The identified foci were visually compared with existing parcellations in surface. Surface projections were available for Julich-Brain (28–30) (v2.9 https://search.kg.ebrains.eu/instances/61c09a8a-bbfe-49eb-ab45-354b05f9a600) and Yeo et al. (31). Seven network projection (included in the freesurfer package). The parcellations from Deen et al. (34), and Faillenot et al. (27) (www.brain-development.org/brain-atlases/adult-brain-atlases/), were available in volume, we projected them to surface using the freesurfer vol2surf function.

Control for specificity: Maximal statistic and null distribution.

For each of the three nodes of the cortical arousal network, we selected cluster mass as the maximal statistic. Cluster mass refers to the sum of summary statistic (in our case the z-transformed t test) across all voxels in the cluster. This statistic is often used for cluster inference—the cluster mass null distribution can be generated with permutation tests, and then the tested cluster mass significance is calculated based on that null distribution. In practice, we created overlap ROIs using the robustness maps—for each area where we find a consistent overlap of four (or five) semantic communities. For each overlap ROI, we computed the maximal statistic for all 415 arousal-related articles (including articles with an a priori ROI approach). We then established the null distribution by repeatedly selecting sets of 415 articles not explicitly mentioning arousal and computed cluster mass within the overlap ROI. The only criterion for an article to be included in the chance level estimation was that it should not contain the word arousal; the article could explicitly mention emotion, or sleep, or pupil diameter, etc… and could include an a priori ROI approach. We repeated this ALE analysis 10,000 times to generate the null distribution of activation likelihood when arousal is not mentioned. We then compared the value obtained in overlap ROIs for the 415 arousal-related articles to the null distribution to obtain a Monte-Carlo p.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The data and code to reproduce results in the manuscript can be found on OSF https://osf.io/rv7ja/?view_only=fe40ff535f474edf92fed66d8f0a9fcf (72).