Substance use disorders are characterised by increased voxel-wise intrinsic measures in sensorimotor cortices: An ALE meta-analysis

. Subsequent analyses showed their involvement in action execution, somesthesis, finger tapping and vibrotactile monitoring/discrimination. Their numerous clinical correlates across included studies highlight the under-discussed role of sensorimotor cortices in SUD, urging a more attentive exploration of their clinical significance.


Introduction
Substance use disorders (SUDs) describe the harmful use of diverse psychoactive substances, which can vary in, e.g., pattern, severity, and psychosocial costs for the individual user.Despite many decades of research with improved methods and study designs, they remain among the most difficult psychiatric disorders to treat, even after participation in long-term treatment programmes (e.g., Beaulieu et al., 2021).The financial burden associated with SUDs continues to place a significant strain on health systems (Castelpietra et al., 2022;Patel et al., 2016), not least because of their relapsing nature and tendency to show chronic trajectories (Fleury et al., 2016;Gooding et al., 2022).There is therefore a fundamental need to identify key levers in the development and maintenance of these disorders, and how they can be prevented and treated to help people achieve sustained abstinence and regain control of their use.
Simultaneously, resting-state functional magnetic resonance imaging (rs-fMRI) maintains its enduring popularity in the field of addiction neuroscience.This is owed to its potential to foster the conception of innovative and effective treatments for individuals affected by SUD.Rigorous assessments of intrinsic resting-state functional connectivity have been widely used, as they can be performed quickly and conveniently without the need to apply a specific task-based paradigm or other stimulations.Originally misunderstood as noise rather than signal (e.g., Purdon and Weisskoff, 1998), fluctuations in the low-frequency range (<0.1 Hz) are now one of the most interesting aspects in trying to explain the topological architecture of intrinsic brain connectivity (Menon, 2011;Uddin et al., 2019).The large-scale organised brain networks thus discovered have been validated to explain metrics of personality, task-based performances, and even psychiatric symptoms (Nostro et al., 2018;Sadaghiani and Kleinschmidt, 2013;Zhang and Raichle, 2010).Recent reviews and meta-analyses conclude that several of these large-scale resting-state brain networks are altered across SUDs, e.g.limbic, fronto-parietal, default mode and salience network as well as in cortico-striatal circuit functioning (Taebi et al., 2022;Tolomeo and Yu, 2022).However, these large-scoped reviews and meta-analyses considered studies based on an a-priori placing of single seed regions in the brain from which the BOLD time series are then extracted.These analyses hold some comfortable statistical advantages regarding multiple comparison corrections.Conversely, they may be limited concerning missing other true effects.Other approaches such as component-wise assessments of functional connectivity networks (e.g., independent-component analysis), which apply automated classifiers to decompose the resting-state data into ideal and separate components can face similar limitations and rarely report local maxima of voxel-wise group comparisons (Nieto-Castanon, 2022).Together, they currently dominate the rs-fMRI literature.Both metrics are not the best candidates for being compiled in coordinate-based meta-analyses, as these generally require a priori whole-brain assessments and equal probabilities for each voxel to be activated to meet statistical assumptions (e.g., Müller et al., 2018).
Whole-brain and voxel-wise metrics, which have emerged over the last few decades, can complement the classic and widely used measures of resting-state functional connectivity to identify brain nodes of aberrant functioning.In contrast to the aforementioned methods that assess functional connectivity by also using intrinsic signals, these metrics do not rely on pre-selection procedures and use a data-driven approach to the functional topology of the human brain.Thus, voxel-wise methods may provide an adjuvant perspective by capturing different aspects of the intrinsic functional organisation of the brain in disease and health, beyond assessments of large-scale brain networks and seed-specific connectivity and therefore may minimise other potential biases associated with pre-selection.At the same time, their voxel-wise assessment makes them well-suited for coordinate-based meta-analyses.Four voxelwise intrinsic measures (VIMs) of brain activity and connectivity have been established in the literature and are increasingly popular and widely applied: Regional Homogeneity (ReHo), (Fractional) Amplitude of Low-Frequency Fluctuations ([f]ALFF), Voxel Mirrored Homotopy Connectivity (VMHC) and Degree Centrality (DC).In short, ReHo assesses the connectivity of a given voxel to its directly adjacent neighbouring voxels, therefore serving as a measure of regional synchronicity (Zang et al., 2004).It is estimated by computing Kendall's coefficient of concordance of the ranked correlation coefficients in the time series of each adjacent voxel's time points.ALFF measures the amplitude of spontaneous activity of a voxel over a defined time course, capturing areas of the brain that show spontaneous fluctuations within the BOLD signal at rest (Zang et al., 2007).Adapted parameters, such as fractional ALFF (fALFF), have also been developed, which are more resistant to physiological noise (Zou et al., 2008).VMHC can be used to determine how consistently the intrinsic time series of any given voxel correlates with its contralateral counterpart in the other hemisphere, hence measuring voxel-wise interhemispheric connectivity (Zuo et al., 2010).Differential interhemispheric connectivity between mirrored voxels is a key feature of lateralisation and supports the organisation of hemispheric dominance when required.DC is a measure that originated in graph theory and is understood by the raw count of non-zero edges (thresholded for a typically rather low Pearson's correlation coefficient) connecting a node to all other nodes in the brain for an unweighted binarised graph (Zuo et al., 2012).Apart from possible pre-definitions of the nodes and edges of interest in a classic graph-theoretical approach, DC can be assessed voxel-wise.It serves as an index of how consistently a voxel is integrated in not only local (i.e., ReHo) but also remote connectivity to all other brain voxels.On the face of it, these four VIMs cover strikingly different facets of what can be discovered about the functional organisation of the human brain.However, it has been shown that the whole-brain concordance rate between them is remarkably high at both the between-subject and within-subject levels, suggesting that the processes underlying these measures might be fundamentally related (Aiello et al., 2015;Yan et al., 2016).As their indices show this considerable degree of parallelism, we hypothesised that their alterations in SUDs might also reveal a pattern of convergence, and perhaps aid in the identification of critical brain nodes affected in SUD.
The number of studies investigating VIMs in SUDs has steadily increased over the last few years, keeping pace with the advances regarding other mental disorders.In contrast to studies on SUDs, however, VIMs have already been meta-analytically compiled in numerous reviews on other major neuropsychiatric disorders, such as schizophrenia (Qiu et al., 2022), major depressive disorder (Yuan et al., 2022), and post-traumatic stress disorder (Disner et al., 2018).These reviews only considered ReHo and ALFF simultaneously, as these are probably the ones most commonly applied in the literature to date.To the best of our knowledge, our paper is the first coordinate-based meta-analysis that integrated VIMs (ReHo, ALFF, fALFF, VMHC and DC) across different SUDs.
Accordingly, in this paper, we surveyed the current literature on VIMs in SUD populations to address the issue that individual fMRI studies often face the problem of low power and limited replicability.For this reason, we used a coordinate-based meta-analytic algorithm, namely activation likelihood estimation (ALE), which accounts for low power issues by calculating reliable meta-analytic estimators where VIM alterations in SUDs converge and replicate across the brain.ALE models the spatial uncertainty of single peak-voxels by fitting a 3D-Gaussian function of activation likelihood around that maximum, given that the spatial precision of a peak likely depends on between-study variance sources, i.e., different sample-, acquisition-and analysis-related choices.Statistical maps can then be plotted on the brain where fMRI studies have reported above-chance convergence of altered VIMs in SUDs.To extend our research question beyond considering all VIMs together, we also conducted sensitivity analyses of contrasts with adequate power.To augment ALE-derived results, we subjected significant convergence clusters to a behavioural domain analysis (BDA) and a paradigm class analysis (PCA) to decode their behavioural functions.To assess the robustness of our results against publication bias, we performed an ALEtailored Fail-Safe-N (FSN) analysis to counterbalance one-sided confirmation tendencies.We then computed meta-analytic connectivity modelling (MACM) for emerging clusters of above-chance convergence to gain deeper insight into how these clusters are organised in larger brain networks and performed conjunction and subtraction analyses on these MACM maps to assess similarities and differences in meta-analytic connectivity profiles.

Systematic literature search and data inclusion criteria
We pre-registered our systematic review and meta-analysis on PROSPERO (https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42023454093).The databases searched for relevant studies were EBSCOhost databases (https://search.ebscohost.com/)PsycINFO, PsycARTICLES, Medline Complete, CINAHL Complete and Psychology and Behavioral Sciences Collection as well as PubMed.MF assigned studies for eligibility and double-checked all data and coordinates.Researcher MM was consulted when uncertainties about study inclusion remained.The state-of-the-art guidelines applicable to the field of neuroimaging meta-analyses as well as generally accepted standards for systematic literature reviews (PRISMA-statement) were followed in the preparation of the paper (Müller et al., 2018;Page et al., 2021).For purposes of self-monitoring and transparency, we have documented the items for both checklists regarding PRISMA guidelines and neuroimaging meta-analyses in supplementary Table S1 and Table S2, respectively.The neuroimaging guidelines suggest some very central inclusion and exclusion criteria.These have been developed in response to methodological inconsistencies and questionable analysis protocols that have been observed over time in the application of ALE and other meta-analytic algorithms.Taking these into account allows for the standardisation of the methodological approach and increases the validity and robustness of the results.Inclusion and exclusion criteria are as follows: (1) According to these guidelines, only data whose peak voxels were reported in standard stereotaxic spaces, in particular the Montreal Neurological Institute (MNI) or the Talairach (TAL) space, were eligible.
(2) All fMRI measurements had to obtain whole-brain coverage fMRI images, which would otherwise violate important statistical assumptions of ALE, that each voxel in the brain has the same statistical probability of being activated.Thus, results from a-priori ROIs were not taken into account, as they limit the scope to the brain area of interest.This procedure not only satisfies a statistical assumption of ALE but also ensures that other true effects are not overlooked.(3) Similarly, only studies that were able to demonstrate a significant result can be considered, as null results without coordinates cannot be included in ALE.This also applies to studies that do not report the significant peak voxels in their original papers or do not make them available through personal correspondence.The limitation resulting from the noninclusion of negative results is discussed in more detail in paragraph 2.4 in the Method section.We extend these criteria with other methodological considerations that may support the robustness and consistency of the results: (4) Studies could only be included if they were conducted on a sample with SUD that met commonly accepted clinical criteria for the disorder (e.g., DSM criteria).This was done to ensure that the results of the samples analysed could be extrapolated to clinical populations that clinicians usually encounter in practice.(5) All subjects had to be adults (>18 years) to minimise developmental differences in brain morphology and function.(6) Severe psychiatric (e.g.psychotic disorders) or somatic comorbidities led to exclusion, as they could exert potentially confounding effects of no interest.Similarly, reports on controls must not suggest that they meet the criteria for any psychiatric disorder.(7) SUD populations must have been compared with an agematched control sample in a between-subjects contrast to be considered eligible.Within-subject designs were not considered.(8) Following the evidence on the replicability of meta-analyses and recommendations for the inclusion of original studies (Müller et al., 2017), we only considered contrasts that underwent a multiple comparison correction or survived their thresholds.This was done to minimise the number of false-positive results in the meta-analytic integration and thus avoid excessive noise in the data.
We had not made a pre-selection regarding the inclusion of different developments of these VIMs, that emerged over recent years.For ALFF, these are fractional ALFF (fALFF) and percent amplitude fluctuations (PerAF) (see brief overview for ALFF-related offshoots in Jia et al., 2020).In addition, there are dynamic variants for ALFF and ReHo, considering a time-variant change on the within-subject level over a certain time window (Fu et al., 2018;Leonardi and Van De Ville, 2015), intuitively named dynamic ReHo (dReHo) and dynamic ALFF (dALFF).Accordingly, these were included in our analyses in addition to their classic representatives that we described in the introduction.We also include all ranges of frequency bandwidths from 0.01 to 0.198 Hz, although the frequency range from 0.01 to 0.08 Hz was the most commonly used, as again we did not make an empirically guided pre-selection for eligibility.
In the first place, we made sample overlap assessments to avoid statistical dependency of data on the between-study level (see paragraph 3.1 in the Results section for details) by checking converging image acquisition parameters, same author appearances, equivalence of research questions across papers, and reports of where measurements took place.In the subsequent preparation of eligible data, we identified misreported coordinates by inspecting a studies' figures, which usually display T1-weighted structural images with respective statistical t-maps or z-maps and compared them with the coordinates given in the corresponding tables.We checked all coordinates and their labels for plausibility using an interactive MNI or TAL atlas template in parallel.In doing so, we have taken into account that there are differences between the brain templates frequently used.We corrected misreported coordinates when necessary and then extracted foci from the included studies into a text file as described in the ALE manual (https://brainmap.org/ale).Lastly, we transformed coordinates that were reported in TAL into the MNI standard reference space using the Lancaster transform (Lancaster et al., 2007).We checked the remaining mask outliers for plausibility in the final step.

Activation Likelihood Estimation (ALE) and data preparation
We applied ALE as the meta-analytic algorithm of choice utilizing the GingerALE software v3.0.2 (https://brainmap.org)(Eickhoff et al., 2012(Eickhoff et al., , 2009;;Laird, 2009;Turkeltaub et al., 2012).An important aspect of approaching the replication of fMRI data is considering that even for a true result, it is unlikely that discrete brain voxels can be replicated in their exactness, due to natural measurement errors and variability attributable to confounders.This spatial uncertainty about where 'true effects' can be assigned to a specific brain site is likely the consequence of different sources of measurement variance, e.g., variation in sample-, preprocessing-and correction choices across included studies.Therefore, ALE models a 3D Gaussian probability function of activation likelihood around a statistically peaking voxel, taking this spatial uncertainty into account and estimating it as a function of sample size by creating more spatially precise probability maps of activation likelihood for larger samples' observations, while modelling less dense probability function gaussian kernels for those of smaller studies (Eickhoff et al., 2009;Laird et al., 2011).This is on the rationale that larger samples should produce more spatially precise estimates than smaller studies do.Probability maps created this way can be translated into p-values and are then tested against the null hypothesis of a random spatial distribution of statistical maxima.Consequently, brain regions that show above-chance convergence in alterations of VIMs comparing SUDs and controls can be derived from the input coordinate-based data.For all analyses of the above-chance converging data, we tested the resulting ALE probability maps against 1000 permutations of random spatial distributions.By the state-of-the-art recommended statistical corrections for multiple comparisons, we applied a voxel-wise threshold of p<0.001 and a cluster-forming threshold of family-wise-error correction (cFWE) p<0.05 as these represent the current gold standard for statistical adjustment for multiple comparisons of meta-analytic fMRI data (Eickhoff et al., 2016;Flandin and Friston, 2019).
Aside from the main analysis of VIM alterations comparing SUDs and controls, we performed sensitivity analyses, given they were feasible to conduct with a minimum power of 17 contrasts per synthesis (Eickhoff et al., 2016).As described in the registration protocol, we planned sensitivity analyses for (1) the main effect of single VIMs, i.e., any alteration of ReHo, ALFF, VMHC, or DC, (2) separate analyses of all increases and then of all decreases across VIMs in SUDs, and in a final step (3) main effects of single SUDs (e.g., alcohol, tobacco, cocaine, etc.).However, the only sensitivity analyses that fulfilled this minimum power criterion aside from the main analysis were: (1) effect of increased VIMs (SUD>controls), (2) effect of decreased VIMs (SUD<controls), (3) main effect of ReHo alterations, (4) main effect of ALFF alterations, (5) effect of increased ALFF (SUD>controls), (6) decreased ALFF (SUD<controls).We tested these accordingly and reported them in paragraph 3.2 in the Results section.The number of studies examining VMHC and DC was insufficient to perform single sensitivity analyses.This was also the case for superordinate substance classes regarding, e.g., alcohol, tobacco, stimulants, and opioids after considering possible sample overlaps.

Behavioural domain and paradigm class analysis
Behavioural domain (BDA) and paradigm class analyses (PCA) were utilized to characterize convergence clusters about their function profiles and their role in supporting higher-level behaviours across different behavioural domains and paradigm classes (Lancaster et al., 2012).The BrainMap database provides a wealth of metadata with which it is possible to compare significant ALE clusters to several task-dependent activations across implemented studies.We applied the respective tools provided within the software Mango v4.1 (http://ric.uthscsa.edu/mango/; behavioural domain v3.1; paradigm analysis v1.6).To do so, we masked emerging ALE clusters as ROIs, then transformed these ROI masks into TAL reference space as this is mandatory for BDA and PCA in Mango and tested how consistently activation sights in the database regarding circumscribed behaviours or experimental tasks match with our ROI masks.Behavioural domains comprise the superordinate categories of action, perception, cognition, emotion, and interoception which can be further partialized into 59 subcategories.Reference data for paradigm classes are organized in 111 different experimental tasks.Both of these analyses necessitate rather conservative statistical corrections weighting information of the raw activation counts in the database against the ROI masks' volume spanned.BDA associations are considered significant if they exceed a threshold of z= 3.0 and PDA associations if they exceed a threshold of z= 3.3.These z-thresholds imply a conservative correction for total ROI size and correction for multiple comparisons across the number of behavioural domains or paradigm classes based on Bonferroni.For further information see http://ric.uthscsa.edu/mango/versionhistory.html#v401.

Fail-Safe-N analysis
Fail-Safe-N (FSN) is a method we applied to address the one-sided confirmation bias that results from ALE being insensitive to null studies and studies missing due to potential publication bias.Accordingly, ALE can only integrate results that report significant peaks.Null results cannot be taken into account, which increases the likelihood that ALE results are biased towards positive findings.Similarly, it is not known how many studies remain unpublished because of their nonsignificant results and therefore are not considered in meta-analytic papers.This common problem is referred to in the literature as the file drawer problem.Both of these issues can make ALE results improperly resilient to non-replication.This could lead to an overestimation of ALEderived clusters' robustness and the perceived homogeneity observed in the literature.Samartsidis et al. (2020) addressed this issue by conducting exhaustive simulations, estimating the proportion of neuroimaging findings that remain in the file drawer to be 6-30%.As a practical solution, Acar and colleagues adapted the FSN method to ALE to estimate the exact number of possible null findings to which a given ALE cluster is robust (Acar et al., 2018).Ideally, the obtained FSN for a given cluster exceeds a defined critical lower boundary, indicating sufficient robustness against a certain number of null findings, but without exceeding a defined critical upper boundary, which would indicate that the results are driven by a few dominant studies.Considering that the most conservative but empirically based choice for a critical lower boundary for a given analysis is 30%, we have applied this for all our FSN calculations.For the upper boundary, Acar and colleagues pointed out that if a meta-analytically derived cluster remains significant even though it is driven by only 5-10% of the included studies, this is an indication that this cluster results from a few but disproportionately influential studies.We set the upper boundary at 5% of studies that must still contribute to the effect of a cluster.The resulting estimate for a critical upper boundary is calculated using the formula "((number of studies contributing to a cluster)/0.05)-(number of studies included in the ALE meta-analysis))".The specific boundaries for individual contrasts are reported in paragraph 3.4 in the Results section, as these can only be derived from the actual resulting number of studies per contrast.
To calculate the FSN, we generated noise studies in R statistics v4.1.0(R Core Team, 2021), which contain the number of foci and sample sizes estimated based on the ALE data we input.These noise studies serve as hypothetical null studies that are assumed to remain in the file drawer.The starting points are the upper and lower critical boundaries.Noise studies are then iteratively added to or removed from the original dataset and ALE analyses are constantly repeated to examine the exact FSN at which a given ALE cluster remains significant.We performed these FSN-related ALE analyses with the same statistical thresholds of voxel-wise p <0.001 and cFWE p <0.05 for the cluster threshold as we did for the main analyses.We adhered to the parsimonious protocol and the code for this FSN-calculation procedure provided by Acar and colleagues (Acar et al., 2018).The code can be obtained via https://github.com/NeuroStat/FailSafeN.

Meta-analytic connectivity modelling (MACM), conjunction and subtraction
We performed MACM analyses based on data provided by the BrainMap database to investigate how significant each convergence cluster is embedded in larger organised brain networks (Eickhoff et al., 2011;Fox et al., 2014;Langner et al., 2014).The MACM method yields comparable results to those of functional connectivity studies, whereby derived meta-analytically.To conduct them, we masked significant clusters as ROIs in the TAL reference space analogue to the BDA and PCA in mango v4.1.These ROIs were then individually input into Sleuth (htt ps://www.brainmap.org/sleuth/)and searched for which studies reported activations that fell into the respective ROI.Those studies that report foci of co-activation were integrated into the MACM analysis and analysed for an above-chance convergence pattern.A few filters were applied in Sleuth.We used "Activations: Activations Only", "Diagnosis: Normals", "Imaging Modality: fMRI" and "Context: Normal Mapping".Extracted studies and their foci were piled in a text file and subjected to ALE to compute the respective MACMs using the same voxel-wise threshold p<0.001 and cluster-wise threshold FWE<0.05 as in the main results.Thus, maps of above-chance convergence in co-activated brain areas for a given cluster were created.
For the conjunction and subtraction of the calculated MACM-maps, there are 3 datasets required which underly these analyses (MACMmap of C1, MACM-map of C2, and a pooled MACM-map for both clusters co-activations together) with the same applied statistical thresholds.First of all, we had to calculate a MACM for each of the two clusters to be compared, as described previously.Then the merge function implemented in GingerALE was used to merge all foci of the individual MACMs for C1 and C2 into one text file.The conjunction map was created using the voxel-wise minimum value of the MACM maps of C1 and C2 and the pooled MACM map.These 3 datasets were already highly thresholded, which makes it possible to set more liberal thresholds for the conjunction analyses per se.
For subtraction analyses, it is necessary to consider different sample sizes between the individual MACM datasets.The GingerALE software operates in such a way that it merges the two foci datasets of individual cluster MACMs, then randomly divides them into two groups of equal sample size (corresponding to the original data) and subtracts them from each other counterbalancing uneven sample sizes and hence uneven weights between MACMs (Eickhoff et al., 2011).For each unique dataset generated in this way, a MACM map is created with ALE values that have been subtracted from one another and compared with the original true data.This is done in a series of simulations, which we set to 10,000 permutations.Voxel-wise p-value thresholded maps were then generated that were transformed into z-maps for easier interpretation.It is evident, that again all statistical thresholds of these statistical subtraction maps are already conservatively and adequately corrected for multiple comparisons on their own.Accordingly, it is not a requirement to set the voxel-wise threshold or the cluster extent for the subtraction too conservatively.For this reason, the voxel-wise threshold for the conjunction and subtractions was set to p<0.01 and a cluster extent of at least 50 mm 3 was defined.All procedures are described step-by-step in the ALE manual for GingerALE 2.3 (https://brainmap.org/ale).

Systematic literature search results and sample overlap
As a result of the systematic literature search, we identified and included 51 eligible papers examining a minimum of 1439 SUD subjects (when sample overlap was considered) across 40 samples, which we assessed as independent.Overall, VIMs were investigated in a large variety of substance classes comprising 11 studies for alcohol, 15 for tobacco, 8 for heroin, 6 for methamphetamine, 2 for betel-quid, 4 for codeine, 3 for cocaine, 2 for ketamine, and 1 for cannabis.Demographic data and study-related characteristics are obtainable through Table 1.The entire process of the search including exclusion of individual studies is documented in Fig. 1 in a PRISMA chart.The reasons for individual study exclusions at the screening stage are documented in Table S3.
However, in the overall assessment of the eligible studies, we found what seemed to be considerable sample overlap.We identified some studies to investigate the same samples by comparing image acquisition data.This procedure, however, did not always prove sufficient.In some cases, we suspected study overlap, although image acquisition parameters diverged between studies.These appeared to us as coming from the same participant pools, which were continuously expanded over time by additional participants.A divergence in the acquisition parameters might be explained by the fact that new subjects could have been assessed multicentrically over the years by using a different scanner, for example.However, these seem to have been analysed at intervals over the years about another VIM.This initially disguises the fact that these samples appear to overlap, sometimes completely and sometimes only partially.Nonetheless, because there would be at least partial overlap, the assumption of independence of measurements between these samples would be violated.We carefully evaluated these cases and pooled those where we suspected overlap into one experiment in the final analyses (Müller et al., 2018).To keep with conservative testing and to not inflate ALE values by overestimating single foci weights, we constantly assigned the smallest sample size for a pooled contrast in the ALE analysis.Cases in which we suspected sample overlap were assigned to the same sample number in the respective column in Table 1.
Furthermore, we adjusted 9 coordinates that we suspected to be incorrect.Adjusted coordinates were then included in the ALE analysis.However, most errors were related to typing or sign errors so they could be straightforwardly reconstructed.An overview of how we have adjusted single coordinates, and from which original studies these coordinates originate can be found in Table S4 in the supplement.

ALE meta-analytic results
The main analysis, which considered all eligible papers regardless of the type of VIM, effect size polarity, and substance consumed did not show a significant cluster of above-chance convergence.We therefore proceeded with the sensitivity analyses that we described earlier in the method section.
The sensitivity analysis (1) showed a significant convergence cluster (C1) of increased VIMs (SUD>controls) with a statistical peak in the left precentral gyrus (PrCG).The cluster convergence map of C1 contains 56.2% voxels that can be assigned to the left postcentral gyrus (PoCG) and 43.8% left PrCG.This result was driven by the VIMs ReHo (80%) and ALFF (20%) with different SUDs contributing to it (i.e., alcohol (Hong et al., 2018), ketamine (Liao et al., 2012), betel quid (Liu et al., 2016), methamphetamine (Nie et al., 2022), and cannabis (Wolf et al., 2023)).We found the same cluster (C1x) with the same statistical peak voxel for the comparison between SUDs and controls in the main contrast of ReHo alterations (sensitivity-analysis [2]), highlighting C1 to be driven by higher values of ReHo in SUDs within this cluster.However, C1x shrank in cluster size and ALE-value of the peak when the ReHo contrast was analysed in isolation due to being reduced by one contributing study that measured ALFF (Hong et al., 2018).Similar to C1 the cluster C1x is assignable to 52.6% of the PoCG and 47.4% of the PrCG.As the clusters C1 and C1x only differ slightly in size and the volume spanned, we only reported consecutive results for C1 in the main manuscript for the sake of brevity and clarity and refer to where the corresponding results for C1x can be found in the supplemental material.Additionally, by calculating the sensitivity analysis (3), we observed another cluster of significant convergence (C2) slightly more inferior than C1 that emerged from the contrast of increased ALFF (SUD>controls).Cluster C2 showed two peak voxels in the left postcentral gyrus (PoCG) and consisted of an area that could be assigned to 82.9% of the left PoCG and 14.6% of the left PrCG, with the remaining 2.4% not assignable.Similar to C1, C2 exhibited contributions from different SUD populations contributing to it (2x methamphetamine (Gong et al., 2022;Liang et al., 2023), codeine (Qiu et al., 2016), and betel quid (Liu et al., 2016)).Since we subsumed different measures for the low-frequency fluctuations under the umbrella term ALFF that differ in their specificity for the low-frequency range, we provide a video file displaying the unthresholded probability maps for ALFF and fALFF plotted on an MNI152 template in the supplemental materials (Video S1).This was done to investigate whether they show a distinction in underlying brain correlates.Aside from fALFF studies not reaching an adequate number (k= 10) to be considered as another sensitivity analysis, we did not observe a pattern of a convincing distinction between brain areas detected by either method variant.
Moreover, in the studies contributing to both clusters, the co-use of drugs was a frequently observed phenomenon, which seemed to be particularly true for the co-use of tobacco (Gong et al., 2022;Hong et al., 2018;Qiu et al., 2016;Wolf et al., 2023).This also included a range of other substances beyond the index SUD and the use of tobacco, such as cocaine, MDMA, and ketamine (Hong et al., 2018;Nie et al., 2022).One sample even showed a pattern that could be considered polytoxicomanic with the co-use of ecstasy, (meth-)amphetamines, marijuana, benzodiazepines and heroin (Liao et al., 2012).For a detailed overview of co-use patterns across contributing studies see Table S9 in the supplemental materials.No other feasible sensitivity contrasts could show a significant difference between SUDs and controls.Visualisations for significant clusters can be seen in Fig. 2 and coordinates are displayed in Table 2.

Cluster characterisation with BDA and PDA
As depicted in Fig. 3 both convergence clusters show significant associations with behavioural domains and paradigm classes.Notably, C1 only showed a significant association with the behavioural domain Action: Execution (unspecified) with no significant associations to any single paradigm class.The same results applied to C1x accordingly.Moreover, C2 showed significant associations with the behavioural domains Execution (unspecified), Somesthesis (unspecified), and for the paradigm classes an association with finger tapping/button press.When C1 and C2 were masked and analysed together, the associations mostly conform with those observed in only C2.However, they were extended by the paradigm class Vibrotactile monitor/discrimination and the overall zvalues for other associations increased moderately.

FSN cluster robustness assessment
We have determined the upper and lower limits for the FSN analysis as described in the method section.These resulted in the following ranges for C1 of 10<FSN C1 <64, for C1x of 7<FSN C1x <59 and for C2 of 8<FSN C2 <56, respectively.Within these ranges, the results can be considered stable, yet not driven by single overly influential studies.The analysis showed that the FSNs of clusters C1 and C1x are within the predefined upper and lower boundary with FSN= 20 (56%) and FSN= 17 (71%), respectively and thus showed solid and satisfactory robustness towards publication bias within their respective contrast.For C2, however, the FSN= 4 (17%) fell below the critical lower boundary and therefore showed susceptibility towards publication bias.

Meta-analytic connectivity modelling (MACM), conjunction, and subtraction results
After completing all preparations by masking the particular ROIs, we computed MACMs for C1 (C1-MACM), C1x (C1x-MACM) and C2 (C2-MACM) respectively as detailed in the methods section.For clusters C1, C1x, and C2, we observed remarkably similar MACM maps, a logical outcome given their high spatial proximity.Connectivity patterns of cortical areas predominantly manifested as ipsilateral for C1-MACM and bilateral for C2-MACM, encompassing regions such as PrCG, PoCG, inferior parietal lobule, and precuneus.Additionally, both maps revealed connectivity to subcortical regions solely within the left hemisphere, extending to numerous thalamic nuclei, adjacent pulvinar, and the posterior insula.The left putamen was covered by both maps, with C1-MACM showing connectivity extending toward the globus pallidus and C2-MACM extending more towards the left claustrum and mid-insula.Furthermore, both exhibited bilaterally organized connectivity encompassing the medial frontal gyrus, the cingulate gyrus, and the paracentral lobule.Lastly, C1-MACM and C2-MACM demonstrated right-lateralized co-activations with the culmen of the cerebellum (see Fig. 4).Detailed coordinates for individual MACMs, including C1x-MACM, are reported in supplemental Tables S5-S7.
Despite MACM analyses being part of the pre-registered analysis protocol, we decided to perform a conjunction and subtraction analysis of the C1-MACM and C2-MACM to confirm similarities and examine differences in the coactivation patterns.These were post-hoc analysis decisions and were not initially declared in the registration protocol.Abbreviations.ALFF amplitude of low frequency fluctuations, cFWE family-wise error correction, cFDR false-discovery rate correction, dALFF dynamic amplitude of low frequency fluctuations, DC degree centrality, GRF gaussian random field correction, MDD major depressive disorder, PerAF percent amplitude fluctuations, ReHo regional homogeneity, VIM voxel-wise intrinsic measure, VMHC voxel-mirrored homotopy connectivity a Voxel-mirrored-homotopy-connectivity always generates two coordinates of contralateral equivalent foci sights of altered intrinsic brain activity.The conjunction and subtraction were based on 227 contrasts examining 3063 participants.The conjunction confirmed their engagement in a cortico-thalamo-limbic-cerebellar circuit comprising areas named in the previous paragraph (Fig. 4).
Although there were considerable similarities, subsequent subtractions showed that there were regions across the MACM maps that differed significantly between C1-MACM and C2-MACM.Both subtractions revealed clusters with higher meta-analytic connectivity to the volumes of their directly adjacent seed-ROI, which were comparably more superior for C1 and more inferior for C2, respectively.In comparison, C1-MACM further showed higher meta-analytical connectivity with contralateral PrCG counterparts in the right hemisphere, with superior and middle frontal gyri, volumes of the medial frontal gyrus, and with the bilateral paracentral lobule.On the other hand, C2-MACM  to y=-30) and the superior-inferior axis (z= 62 to z= 50) are displayed on the right and left sides, respectively.This image was created using Mango v4.1 (http://ric.uthscsa.edu/mango/)plotted on the ICBM MNI152 nonlinear asymmetric 1×1×1 mm template (Fonov et al., 2011(Fonov et al., , 2009)).
showed higher connectivity with bilateral mid-cingulate gyri, a peak within the right intra-parietal sulcus (no grey matter label assignable) and the right declive in the cerebellum (Fig. 4).Notably, co-activations surrounding the putaminal structures extending to the pallidum for C1-MACM and extending to the claustrum and mid-insula for C2-MACM were not found in either the conjunction or the subtraction analysis.Coordinates for the conjunction and subtraction analyses can be obtained via the supplemental Table S8.

Discussion
In our main analysis, we did not identify a converging pattern of altered VIMs across all SUDs and all metrics with our ALE analysis.However, the sensitivity analyses revealed two significant clusters of above-chance convergence.These clusters showed increased levels of predominantly ReHo (C1) and ALFF (C2) in the left sensorimotor cortices, specifically in the medial central sulcus between the left PrCG and PoCG across SUDs.Interestingly, both clusters were driven by different substance classes, indicating they are not the result of a particular substance.However, the contributing studies also revealed a high prevalence of drug co-use with other substances.While this could be a further indicator that the effect, we identified is not linked to specific substance classes, it should be noted that co-use of other substances may characteristically alter intrinsic brain signatures and therefore has the potential for unknown interactions (e.g., Zhang et al., 2024).
Cluster C1 (and C1x), which is primarily influenced by ReHo, exhibited robustness against studies remaining in the file drawer.However, cluster C2, which is driven by ALFF, showed susceptibility to potential publication bias, as suggested by FSN results.We argue that C2's robustness exceeds what FSN indicates due to its proximity to C1. GingerALE categorises a study as contributing to a cluster if one of its peaks falls within the cluster boundaries.However, it does not consider the migration of density functions across cluster boundaries.Some foci's probability maps of contributing studies from C1 and C2 overlap, adding   (Fonov et al., 2011(Fonov et al., , 2009)).
to the maximum ALE-value within both clusters.FSN calculations overlook this, compromising the conservative assessment of FSN results to some extent.Thus, we consider the effects in the PrCG and PoCG to be robust, as C1 itself is robust and covers PrCG and PoCG volumes almost equally.However, caution should be exercised when extending these effects to C2.At the same time, our data shows that reductions in VIMs across SUDs do not exhibit a convergent pattern, and the evidence for alterations in VMHC and DC still requires substantial replication to be approximately conclusive.
Upon reviewing the contributing studies, notable observations were made regarding the behavioural correlates of increased VIM-values.These studies found associations between VIMs and addiction-related markers, such as craving (Liao et al., 2012), a mediating role in response to naltrexone (Qiu et al., 2016), measures of addiction severity including the amount of use (Liao et al., 2012;Wolf et al., 2023), and evidence for an association with illness duration.Interestingly, the Fig. 4. MACM, conjunction and subtraction results.For the subtraction, the red-yellow colour indicates areas where the C1-MACM shows higher meta-analytic coactivation compared to C2-MACM and vice versa for the green colour.This image was created using Mango v4.1 (http://ric.uthscsa.edu/mango/)plotted on the ICBM MNI152 nonlinear asymmetric 1×1×1 mm template (Fonov et al., 2011(Fonov et al., , 2009)).
associations between increased VIMs and behaviour showing dominantly negative correlation coefficients were overall counterintuitive.This means that SUD participants who display higher VIM-values in sensorimotor cortices crave less, respond better to naltrexone treatment, and seem less severely affected (i.e.consume less).Furthermore, it is noteworthy that the effects were mainly observed in individuals with SUD who had relatively short durations of illness, around the 5-year mark or earlier (Gong et al., 2022;Liang et al., 2023;Liao et al., 2012;Nie et al., 2022;Qiu et al., 2016;Wolf et al., 2023), with the exception of the studies by Liu et al. (2016) and Hong et al. (2018), who present with a high variability in illness durations.Overall, these findings suggest that higher levels of ReHo and ALFF may be associated with an earlier stage and less severe pattern of SUD.This is consistent with other research in the SUD literature on VIMs, including studies that have found a negative correlation between addiction severity measures and other regions of the sensorimotor cortices (Luo et al., 2017;Weng et al., 2021).Data by Guo et al. (2022) suggest that increases in VIMs in sensorimotor cortices are dependent on consumption patterns.Only regular patterns of alcohol use beyond the amount of consumption were associated with increases in ReHo and ALFF in sensorimotor cortices with a simultaneous decline in grey matter in these areas.This finding is consistent with meta-analytic results of a grey matter decline in left sensorimotor vertices in alcohol use disorder, similar to the ones found in our work (Spindler et al., 2021).In gaming disorder, it has been demonstrated that ReHo-and ALFF-values in sensorimotor cortices are among the most reliable predictors of gaming disorder severity when evaluated with data-driven methods (multi-voxel pattern analysis; Ye et al., 2022).These associations highlight the clinical significance of altered VIMs for addiction-relevant measures.
So far, sensorimotor areas as promising sites of SUD symptoms have received the most attention through the study of drug cue reactivity paradigms.Yalachkov et al. (2010) suggested that repetitive drug use leads to neural representations in sensorimotor cortices based on value and experience, which are later associated with cravings and promote relapse.Their remarks align with our review's observations regarding associations between these representations, craving, and SUD severity metrics.In these cases, cue-evoked activations correlate positively with craving and severity.Casartelli and Chiamulera (2016) propose a model of high-and low-order cue reactivity, with low-order reactivity involving the motor cortex potentially facilitating drug-seeking and drug-taking behaviour.Both models suggest the initial acquisition of neural representations in sensorimotor cortices, being crucial in later SUD stages.This aligns with our data if one assumes that increased ReHo and ALFF correlate with neuroplastic adaptations due to addictive behaviour in the acquisition stages that precede a then consolidated functional reorganisation of these areas in favour of drug-associated cues in later stages.
To broaden the view beyond what little is known about their role in SUD, there is evidence to suggest that sensorimotor cortex functions may go beyond crude involvement in motor domain-specific action execution.Recent studies by Gordon et al. (2023) and Jensen et al. (2023) have shown areas in the motor cortex, termed inter-effector regions and Rolandic motor area sites, respectively, to exhibit functions typically characteristic of cortices with higher associative functions.Task-based activations within these areas are less specific to particular motor movements than expected based on their spatial localisation in a model of strict somatotopy.Further, they demonstrate higher connectivity to, e.g., nodes assignable to the cingulo-opercular network than those sensorimotor sites more specific to a certain motor domain.Upon visual inspection, it appears that one inter-effector region may converge with the clusters we have identified, as it is also localized in a similar vertex in the left hemisphere that lies medially within the central sulcus.However, we cannot validate an overlap based on our data.These inter-effector regions are hypothesised to facilitate complex motor behaviours (Jensen et al., 2023), whole-body action processes, and sympathetic arousal (Gordon et al., 2023), suggesting our clusters may be functionally involved in higher somato-cognitive programmes in SUD, representing a trace of repetitive drug application/-seeking and craving-related arousal.Our consecutive functional characterisations using BDA and PCA imply that C1 and C2 are involved in sensorimotor processes such as unspecified action execution, unspecified somesthesis, finger tapping and vibrotactile monitoring/discrimination.These behavioural associations could fit this notion of a rather unspecific functional contribution to sensorimotor functions by referring to unspecific aspects of action execution and somesthesis rather than concrete motor skills.Further research is required to determine how alterations in these areas could aid in contextualising behavioural findings related to deficits in simple and complex motor performance, coordination, and sensory integration observable in SUDs (Dervaux et al., 2013;Güçlü et al., 2023;Roizenblatt et al., 2021;Sánchez-Camarero et al., 2019;Umut et al., 2020;Wolf et al., 2023).
On the network level, subsequent MACMs and conjunction analyses demonstrated their involvement in a cortico-thalamo-limbic-cerebellar circuit.Therefore, it is likely that they supply various rudimentary as well as complex motor execution, coordination, and monitoring functions, e.g. in sensorimotor and serial reaction time tasks, facilitating conscious and unconscious motor learning processes.(Bernard and Seidler, 2013;Hardwick et al., 2013;Turesky et al., 2016).The subtraction analyses support this picture, as C1 has stronger meta-analytic connectivity to other motor areas in the contralateral and same hemisphere, such as PrCG and paracentral lobule, while C2 exhibits more robust connectivity to areas involved in salience mapping (i.e., cingulate gyrus) and hand-eye movement coordination (i.e., declive of the cerebellum; Park et al., 2018).Furthermore, aberrant subcortical-cortical signalling of serotonin and dopamine within motor-related circuits projecting from the basal ganglia and thalamus to cortical sensorimotor regions has been proposed to facilitate psychomotor functions across different psychiatric diagnoses (Northoff et al., 2021).In line with this, recent evidence has also shown that higher thalamo-cortical and pallido-cortical rsFC with sensorimotor regions contribute to psychomotor retardation and agitation in depression, respectively (Wüthrich et al., 2023).Therefore, the connectivity network identified in our MACM analyses that in part co-activates with these subcortical structures could play a role in SUD by representing withdrawal or craving-associated psychomotor agitation and retardation.
In summary, our findings suggest associations between symptoms of substance use disorder (SUD) and elevated intrinsic signatures within sensorimotor cortices.Regarding their apparent clinical relevance for SUD-affected individuals, they align with neural representations originating from task-based drug cue-reactivity studies.However, the exact contribution of these neural underpinnings to the SUD phenotype is not yet well understood.It is worth noting that the evidence suggested that they may be associated with early and less severe stages of SUD leading to the interpretation of a correlate that underlies motor-related and repetitive addictive behaviours that could be essential in earlier stages of the disorder.Furthermore, considering the network level one might note that sensorimotor circuits including subcortical structures could be involved in mediating psychomotor disturbances in SUD that are often observed especially in states of withdrawal and craving.Identifying their true behavioural correlate and whether they are specific to SUDs or contextualise in psychopathology more broadly could be crucial for advancing this research question.

Limitations
This work has some limitations that require consideration.First, our approach to sample overlap assessments was conservative, aimed at minimizing potential false positives resulting from the double consideration of overlapping samples.Unfortunately, the absence of validation for these assessments means they remain as suspected cases, introducing the risk of false negatives.As the literature on this topic is relatively new and meta-analyses should guide future studies, we have been rigorous to avoid misleading future research.
Second, the current literature is heterogeneous.For instance, it encompasses various SUDs with different demographics, clinical features and across different VIMs.Some of these metrics may not yet have adequate power for meta-analyses when considered individually.Therefore, our approach has a broad and exploratory nature.Although involving 1439 SUD participants, the possibility of substance-specific or metric-specific effects, or interactions between them to be currently underpowered is a consideration to make.
Third, the ALE method used is limited by its dependence on significant peak coordinates.Although only one study in our systematic literature search did not survive correction for multiple comparisons, it is uncertain whether there are any unpublished null findings.Furthermore, sensitivity analyses may inflate results when assessed within directionality.Indeed, our main analysis did not yield any significant clusters.ALE tends to misconceive many true effects as a pattern of random noise.For VIM data, the foci are distributed broadly across the brain, which should be considered a limitation.Although the main effect of ReHo cluster C1x is significant irrespective of the directionality of effect sizes, this was not the case for cluster C2, as the main effect of ALFF was not significant.However, the FSN indicated that our findings show a certain degree of robustness.

Conclusion
The primary motor and sensory cortex are two of the most extensively researched and well-understood structures of the human brain.Although they may be perceived as supporting relatively basal and lower-order functions of human behaviour, their involvement in SUD symptoms leaves many questions unanswered.It is important to acknowledge that our voxel-wise and data-driven approach has brought attention to sensorimotor sites as brain nodes of interest.This certainly represents a shift from the previous emphasis on brain networks such as salience, default mode, and limbic networks.Unfortunately, sensorimotor cortices are brain regions that receive less attention in SUD research, leading to chronic under-discussion.To act on a solid foundation of evidence, it is crucial to replicate these findings and to decipher their exact role within the neuroscience of addiction.
Abbreviations.ALFF amplitude of low frequency fluctuations, cFWE family-wise error correction, cFDR false-discovery rate correction, dALFF dynamic amplitude of low frequency fluctuations, DC degree centrality, GRF gaussian random field correction, MDD major depressive disorder, PerAF percent amplitude fluctuations, ReHo regional homogeneity, VIM voxel-wise intrinsic measure, VMHC voxel-mirrored homotopy connectivity a Voxel-mirrored-homotopy-connectivity always generates two coordinates of contralateral equivalent foci sights of altered intrinsic brain activity.b The sample inLuo et al. (2017) seem to overlap with either sample 1 or sample 2 reported in Guo et al. (2019) or both.Therefore, the Luo et al. sample and the Guo et al. sample 1 have been pooled in the final ALE. c Study reports median and IQR; a sample mean and a standard deviation has been estimated for better comparability using the calculation provided by Luo et al. (2018) and Wan et al. (2014).

Table 1
Demographic and clinical characteristics of included studies.Studies where sample overlap was suspected share the same sample number and are treated as one sample.Their contrasts are considered as one experiment in the ALE analysis.

Table 2
ALE results show significant clusters of above-chance convergence between SUDs and controls.Depicted are the results for the main analysis and subsequent sensitivity analyses.