Recruitment of the motor system during music listening: An ALE meta-analysis of fMRI data

Several neuroimaging studies have shown that listening to music activates brain regions that reside in the motor system, even when there is no overt movement. However, many of these studies report the activation of varying motor system areas that include the primary motor cortex, supplementary motor area, dorsal and ventral pre-motor areas and parietal regions. In order to examine what specific roles are played by various motor regions during music perception, we used activation likelihood estimation (ALE) to conduct a meta-analysis of neuroimaging literature on passive music listening. After extensive search of the literature, 42 studies were analyzed resulting in a total of 386 unique subjects contributing 694 activation foci in total. As suspected, auditory activations were found in the bilateral superior temporal gyrus, transverse temporal gyrus, insula, pyramis, bilateral precentral gyrus, and bilateral medial frontal gyrus. We also saw the widespread activation of motor networks including left and right lateral premotor cortex, right primary motor cortex, and the left cerebellum. These results suggest a central role of the motor system in music and rhythm perception. We discuss these findings in the context of the Action Simulation for Auditory Prediction (ASAP) model and other predictive coding accounts of brain function.


Introduction
In the case of (most) music, we do not merely passively receive temporal patterns, but actively engage with the sound stream by discerning an underlying periodicity. This profound shaping of temporal perception is central to understanding and participation in music, dance and even speech/conversation. In recent years, neuroimaging studies have shown that passively listening to music activates brain regions that reside in the motor system proper. The same neural correlates underlying the creation of music and moving to music appear to be involved even when one is only listening to a musical piece [1][2][3][4][5][6][7].
The motor system has received increasing attention in non-purely-motor domains [8][9][10][11][12]. Activity in motor regions during perception of human actions and language is ubiquitous. In early theories of cognitive processing, motor processes and perceptual processes were a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 The above theories all posit that cortical motor areas play a role in music listening. Another emerging theme is that of the motor system having a predictive role in perceptual processes. These accounts are primarily in agreement in terms of which sub-areas in the motor system are involved. Schubotz's framework directly proposes involvement of both lateral PMC and pre-SMA/SMA, while ASAP and Rauschecker's theory both set the dorsal auditory stream (which includes dPMC) as the primary substrate. However, activated regions within the motor system measured by neuroimaging methods tend to vary between research studies. For instance, numerous music listening experiments report motor activity in both supplemental motor area (SMA) and dorsal premotor cortex (dPMC) [1][2][3][4][5][6]. Among these studies, a few show neural activations in cerebellum [2,4,5] or primary motor cortex (M1) [7] during a music listening task. Said differently, most music-listening studies do not show activation in every region of the motor system, nor do they show uniform activation in any one part of the motor system. In order to gain insight into the functional contribution of the motor system to passive music perception, one necessary step is to determine which motor regions are consistently contributing across music listening instances.
There are many factors likely to contribute to differences across studies, as each individual experiment has its own musical stimuli that vary in terms of particular characteristics, such as rhythmicity, familiarity, and valence of the music, for instance. Stimuli consisting of highly regular rhythmic structure might engage brain regions important for timing and sequential structure (i.e., supplementary and pre-supplementary motor areas and the cerebellum), while others might not. Experiments also vary in terms of what a participant is directed to focus on in these paradigms, ranging from complete passive listening (not attending) to judging beat or other characteristics of the stimuli. Such task demands are also likely to influence which regions are active, as directing attention to a stimulus may encourage focusing on particular aspects of the music, such as its beat or rhythmicity. In the present study, we are interested in discovering what motor regions are engaged during all music perception-those activated during passive listening. We define passive listening as attentive listening while remaining still (i.e., not tapping along to the music).
Identifying which regions are active consistently across all music listening tasks would help gain insight into the underlying processes and hone existing theories. Many theories outlining the functional contribution of individual motor areas exist, which can be used to determine what particular function is being carried out in a task utilizing that motor region. If one critical component is the dorsal auditory stream, which has a proposed role in motor planning and mapping auditory information onto potential motor acts, we should observe observation in dPMC [22,23]. If activation is found in vPMC, the underlying mechanism might be similar to that proposed in the action observation network, which is responsible for mirror system activity for observed and produced actions [24,25,26]. Many studies that involve music with beat manipulation report activity in SMA and pre-SMA regions, which are presumed to be important for sequential processing of action-related stimuli and for inhibition of movements, respectively [27,28]. Thus, SMA activity might indicate processing of sequential aspects of the music, and pre-SMA the inhibition of the natural tendency to move or sway to the music. We also might observe activation of structures in the basal ganglia, which appear to be involved in beat perception [29]. The basal ganglia are important for movement timing and sequential movement execution [30,31]. M1 activation corresponds to particular motor commands that are carried out by specific muscle groups [32] and has also been found active during observation of actions [33,34,35]. Finally, the cerebellum is known for its crucial role in motor timing and coordination. Research on sensorimotor adaptation has long focused on the role of the cerebellum in predicting sensory consequences of movement and adapting to errors in these predictions (for a more interactive view see [36,37,38,39]). Furthermore, cerebellar activation in conjunction with hippocampal activity is thought to underlie spatiotemporal prediction of movements [40]. This implication in predictive processes of motor control might extend to imagined and simulated motor computations, e.g. the cerebellum might be active in musical prediction even when no direct motor control is required.
In order to determine which of these regions show reliable and consistent activation during music perception, we employed a meta-analysis of all neuroimaging experiments consisting of music listening using an activation likelihood estimation (ALE) [41]. We predict that from this meta-analysis will emerge a pattern of activation that will enlighten and instruct future theories aiming to explain the motor-specific contributions to passive music perception. Activation of any of the motor regions will provide conclusive evidence for the involvement of the regions of the brain typically considered "action areas", in the perceptual domain of passive music listening. This will inform theories about what roles are played by the traditional motor system.

Methods
Meta-analyses provide a formal, statistical integration to combine the results of several studies that address a set of related research questions. There are several methods available for the metaanalysis of neuroimaging data and careful consideration was given as to which was most appropriate for this study. First, our study aims were to synthesize neuroimaging data of studies comparing rest and passive listening. More specifically, we wanted to identify regions of consistent activation across studies. Activation likelihood estimation (ALE) meta-analysis [41] addresses this by treating the spatial relationship between within study foci as fixed effects and between study relationships as random effects. Secondly, we considered the characteristics of our dataset. Unlike some other methods (e.g., KDA and MKDA), ALE uses a Gaussian kernel. When several distinct foci are located within the same general area, the Gaussian kernel is most likely to recover the separate foci. And, in general, if the spatial error on peak locations is approximately Gaussian (a reasonable assumption), then the Gaussian kernel will likely yield the most sensitive results. To investigate our research questions, we conducted ALE meta-analysis. Imaging studies commonly report brain locations of task-induced activations as coordinates in 3D space (x,y, and z). ALE meta-analysis techniques can be used to identify reliable activation patterns in 3D space across studies. ALE is a coordinate-based approach to a meta-analysis, allowing researchers to integrate imaging data. Studies are collected, coded and interpreted using analytical methods to assess likelihood of activation through agreement or overlap in activation patterns.
To perform the ALE meta-analysis, we began by first locating relevant studies. Relevant studies were those that utilized functional brain imaging of healthy subjects listening tasks. We conducted literature searches in Medline and the BrainMap database [42] using a combination of the following: (1) a functional brain imaging modality, including positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) and (2) relevant adjectives related to auditory stimuli. For example, a single search consisted of "Imaging" AND "passive listening" OR "fMRI or functional magnetic resonance imaging" AND "auditory". The literature search of Medline was performed February 2016 and returned 132,294 papers. The literature search of BrainMap was performed September 14, 2016 and returned 244 studies. To ensure our ability to investigate the specified research questions a subsequent study selection process was done by applying the following inclusion criteria to the studies: (1) subjects were healthy adult participants; (2) The analyses include contrasts against rest or a suitable low-level control condition; (3) peak coordinates of group-level activations were reported; (4) foci activation were available in the form of standardized stereotaxic coordinates in either Talairach or Montreal Neurological Institute (MNI) space; (5) that results from the entire scanned volume were reported; and (6) data were available as of September 2016. An effort was made to obtain unreported coordinates from selected studies meeting all other criteria, however, this effort did not return any results. The subsequent review process was performed in two phases. First, an automated review of study titles was done using the R environment (R Development Core Team, 2008) to remove studies that were not in healthy human subject populations. The automated review removed 8144 papers from the database. Next, reviewers read the abstract and/or methods sections of remaining studies to assess appropriateness using the above inclusion criteria. Fig 1 illustrates the full review process for the meta-analysis. The process yielded 42 experiments that met the criteria for inclusion. A full list of experiments included can be found in Table 1. Experiments included a total of 386 unique subjects, approximately 195 male and 171 female.
Coordinates (X, Y, Z) for selected studies were recorded and, where necessary, transformed to Talairach space. Coordinates from individual studies were transferred to a text file formatted for analysis in GingerALE 2.3.6 (http://www.brainmap.org/ale/; Research Imaging Center, University of Texas, San Antonio, TX). These were transferred either using brainmap's Sleuth software (if the studies were located in the brainmap database), which outputs coordinates in the correct format for GingerALE, or were transferred individually by hand. The ALE metaanalysis was carried out in GingerALE. The ALE procedure was as follows: (1) model of singlestudy activation foci as peaks of three-dimensional Gaussian probability densities with subject-based full-width at half-maximum values [43]; (2) summation of probability densities to produce a statistical map estimating the likelihood of activation at each voxel; (3) thresholding of this ALE map based on the null hypothesis of a uniform distribution of foci; (4) correcting for multiple comparisons by family-wise error thresholding. Resulting statistical maps show clusters where convergence between foci is greater than would be expected by chance. Statistical maps were thresholded using cluster-level family-wise error correction P<0.05 (clusterforming threshold voxel-level P<0.001).  We split the data into separate studies that used either musicians only or nonmusicians only, with the intention of performing a contrast analysis between the two groups. Unfortunately, there were too few studies in these groups individually (14 experiments in each group), so we were unable to complete this contrast. Fig 2 shows the activations during passive listening, demonstrating the common brain network underlying music perception. Talairach coordinates for these ALE foci are presented in Table 2. Activations were seen in the bilateral superior temporal gyrus, transverse temporal gyrus, insula, pyramis, bilateral precentral gyrus, and bilateral medial frontal gyrus. As shown in Fig 2, there was activation in the left and right premotor cortex (BA 6), right primary motor cortex (BA 4), and the left cerebellum.

Results
An inspection of Table 3 reveals that Cluster 1 is centered over the right primary auditory cortex, and spans from BA 22 and BA 41/42 (primary and secondary auditory cortices) in the right hemisphere to BA 6 (right premotor cortex). Likewise, in the left hemisphere, cluster 2 is centered over the left primary auditory cortex, and spans from BA 22 and BA 41/42 (primary  and secondary auditory cortices) in the left hemisphere to BA 6 (left premotor cortex). Cluster 3 reveals motor system activation in the right hemisphere, centered in right premotor cortex and spanning from premotor to primary motor cortex. Finally, cluster 4 is located in the left cerebellum. Fig 2 depicts the activation patterns seen bilaterally for a range of z values.

Discussion
We found evidence for consistent activation of various regions of the brain during passive music listening. As expected, our results showed activation in the primary and secondary auditory areas bilaterally. This is consistent with the existing literature showing that these areas are the critical regions of cortex for processing incoming auditory information [44]. Other activated areas included both right primary motor cortex, right and left lateral premotor cortex, and left cerebellum. We discuss in turn the implications for each of these findings below.

Activation of premotor cortex
We were unable to pinpoint activation to any further subregions of lateral PMC (i.e., dorsal or ventral), as the activation pattern could be consistent with either dorsal or ventral PMC. The average coordinates for these regions overlap in such a way that neither can be ruled out. This means that premotor involvement could be via dorsal, ventral, or both. The activation of PMC in the present analysis is consistent with both ASAP and the HAPEM framework. However, because we do not know whether this activation is localized to ventral or dorsal PMC, it is unclear if this activity reflects involvement of the dorsal stream, or potentially the action observation network that recruits vPMC for action simulation. Also, given that these clusters only represent aggregate BOLD activation, we do not have insight into the temporal dynamics of this activity, which will be crucial for inference about its origin.

Activation of Primary Motor Cortex
M1 activity could reflect either an excitatory or inhibitory contribution, as the BOLD signal does not differentiate the two. Vigneswaran et al. [45] report that while many M1 neurons are active during action observation and thus classified as mirror neurons [46], M1 neurons directly connecting to spinal circuitry and thus contributing to observed action are suppressed during the observation of action, in order to prevent explicit action. Simulative properties of mirror fMRI meta-analysis of music listening fMRI meta-analysis of music listening neurons have also been confirmed in response to auditory sounds [47]. Thus, either excitatory, inhibitory, or both could give rise to activation of M1 during passive music listening. Some theories of mirror neuron activity [48] additionally claim that the mirror neuron network uses active inference during the perception of observed actions using predictive mechanisms.
Further examination of the studies that contributed to the right premotor/primary motor activation cluster reveals that a number of them used tasks that were not passive listening in the same way as passive background listening. For example, Chen et al. [5] required in some experimental conditions for participants to anticipate later tapping to the beat in subsequent trials, which may recruit motor planning regions during the listening task. Grahn and Brett [6] asked their participants to determine whether the rhythms of two stimuli were the same or different, which may recruit motor areas to assist with the detection task. Finally, Tsai et al. [49] asked participants in some of their tasks to covertly hum along with the music that they were hearing, which may recruit motor areas for subvocalization. Therefore, more work should be done to decisively conclude whether motor areas are recruiting during background listening, in addition to passively listening for properties of the music while remaining still.

Activation of cerebellum
Many studies contributed to the cluster indicating left cerebellum. The activation of PMC and cerebellum during music listening supports predictive theories of the motor system, such that the cerebellum might provide the predictive component in a forward model of the upcoming sensory consequences. The cerebellum may be providing an inverse model for mapping sensory input to the simulated movement that would give rise to that sensation. An investigation of the temporal dynamics of communication among these regions can again provide further insight into the mechanism.

Lack of activation in SMA/pre-SMA and basal ganglia structures
While quite a few studies did report activation of SMA and pre-SMA, we did not find corresponding activation in our meta-analysis. We also failed to find evidence of basal ganglia activation. One potential reason for this discrepancy may be that another process on top of passive fMRI meta-analysis of music listening listening is needed to engage these regions. The generally agreed upon roles of both SMA and the basal ganglia are of sequential learning and timing [31,50,51]. Because these are properties of the majority of music, this region is likely to be recruited in many musical contexts. However, without directly listening for these properties of the music, it appears that automatic SMA and basal ganglia activation is not prevalent. Looking more closely at those experiments that explicitly report SMA/pre-SMA activation, they do appear to have a musical beat component to them, which relies on the underlying sequential and timing properties of the music. Bengtsson et al [4] encouraged participants to focus on rhythmic properties of the music, as did Chen, Penhume and Zatorre [5] and Grahn and Brett [6]. Baumann et al. [1] required subjects to do a counting task during passive listening as a distractor, which may have resulted in SMA activation. Experiments showing basal ganglia activation also appear to involve a beat detection task [29]. It also may be the case that activation of SMA/pre-SMA is only prevalent in musicallytrained individuals, who are more likely to attend to and perceive the structural aspects of the music due to their background training. Baumann et al. [1] report increased activation of both pre-SMA and SMA in musicians compared to nonmusicians, as did Bangert [3] for SMA activation. Participants in Meister et at [2] showing SMA activation were all musicians. Thus, it appears SMA activation is likely due to either a trained musical background and/or a focus on the rhythmic properties of music.
These results show that the passive perception of music engages a large and complex network of brain regions. This includes activation of areas in the motor system proper. Activation of the cerebellum and primary and premotor cortices suggests that perceived music is partly processed in areas typically considered as important for action-relevant information only. Recruitment of premotor areas during music listening supports many theories of motor involvement during perceptual tasks [11,17,19]. The idea of shared neural resources for tasks with underlying computational similarities has gained recent theoretical and grounded neurobiological support [16]. Most current theories suggest that perceptual processing involves the same kinds of temporal prediction involved in action, making a shared circuit useful for action-based and perception-based processing. An alternative (or potentially compatible) hypothesis is that involvement of PMC reflects the process of simulation [35], where the motor system underlies simulation of the actions required to create the observed sensory information. Our findings are consistent with both of these theoretical frameworks, though it does not provide any insight for distinguishing which theory best fits the data, as this meta-analysis only tells us which areas are active at some point in the process. This work supports the currently merging conceptualizations of action and perception [14,48].
One limitation of the present meta-analysis was that we were unable to obtain data from contacted authors for studies that did not report all of the observed brain activations. It is possible that unpublished or unreported activations may have biased our results toward reporting motor and auditory areas, as studies that do not find activation in these areas of interest are less likely to be reported. This inability to obtain unavailable data also likely contributed to our inability to obtain enough studies for the musician/nonmusician contrast. Further exploration of differential activation in musicians relative to nonmusicians is important for advancing this work. Musicians exhibit plasticity-induced changes perceptual and motor abilities, as well as changes in structural and functional neuronal connectivity [52][53][54][55]. In particular, we believe that musicians passively listening to music should also recruit supplementary motor cortex, and might show greater activation of the cerebellum, which has a larger volume in musicians [54]. Another interesting avenue to pursue is to run more studies that directly compare different types of music listening tasks. For instance, we might compare background listening to listening in anticipation of some movement to listening for particular musical features, such as rhythm or grooviness. This will incorporate context-dependent music listening, which may reveal that different (but likely highly overlapping) networks are recruited in separate contexts. An additional limitation of this approach is that while we can identify which brain areas are active at some point during the music listening task, the BOLD signal cannot tell us anything about the temporal dynamics of the process. Complementary methods, such as EEG, should be used along with imaging data to investigate the functional connectivity among these music listening networks. This will also allow us to determine which of the existing theories fit best with the data.
In summary, this study adds support to the idea that motor planning activity serves not only to help us move but is recruited for music perception even in the absence of movement. Further exploration will elucidate the functional purpose of this recruitment, as well as why and how different music listening contexts seem to engage slightly different networks. An understanding of the auditory-motor interactions underlying music perception could explain a growing number of findings suggesting an important link between music perception and the action systems of the brain.