Anomalies in global network connectivity associated with early recovery from alcohol dependence: A network transcranial magnetic stimulation and electroencephalography study

Abstract Although previous research in alcohol dependent populations identified alterations within local structures of the addiction ‘reward’ circuitry, there is limited research into global features of this network, especially in early recovery. Transcranial magnetic stimulation (TMS) is capable of non‐invasively perturbing the brain network while electroencephalography (EEG) measures the network response. The current study is the first to apply a TMS inhibitory paradigm while utilising network science (graph theory) to quantify network anomalies associated with alcohol dependence. Eleven individuals with alcohol‐dependence (ALD) in early recovery and 16 healthy controls (HC) were administered 75 single pulses and 75 paired‐pulses (inhibitory paradigm) to both the left and right prefrontal cortex (PFC). For each participant, Pearson cross‐correlation was applied to the EEG data and correlation matrices constructed. Global network measures (mean degree, clustering coefficient, local efficiency and global efficiency) were extracted for comparison between groups. Following administration of the inhibitory paired‐pulse TMS to the left PFC, the ALD group exhibited altered mean degree, clustering coefficient, local efficiency and global efficiency compared to HC. Decreases in local efficiency increased the prediction of being in the ALD group, while all network metrics (following paired‐pulse left TMS) were able to adequately discriminate between the groups. In the ALD group, reduced mean degree and global clustering was associated with increased severity of past alcohol use. Our study provides preliminary evidence of altered network topology in patients with alcohol dependence in early recovery. Network anomalies were predictive of high alcohol use and correlated with clinical features of alcohol dependence. Further research using this novel brain mapping technique may identify useful network biomarkers of alcohol dependence and recovery.


| INTRODUCTION
Alcohol is a major contributor to disease burden and mortality in the global community. 1 Alcohol dependence, a severe and chronically relapsing disorder, is characterised by a diminished capacity to inhibit alcohol consumption despite harms associated with continued use. [2][3][4] Difficulty inhibiting the desire to drink typically persists beyond detoxification and is related to increased rates of relapse. 5 Models of addiction have expanded the traditional neural mesolimbic 'reward' circuitry underlying addictive behaviours 6 to also include two independent, yet interconnected brain systems: the limbic system in the incentive-sensitisation of drugs 7 and the prefrontal circuitry (PFC) 4,[8][9][10] in regulating control over drug seeking behaviours. These cortical structures do not act in isolation, but rather, they recruit an interconnected network of brain regions.
There is extensive research examining alterations within cortical structures of the mesocorticolimbic 'reward' circuitry; 6,[11][12][13] however, research into the global features of this connectivity (i.e., how brain regions within the network communicate with each other) has been limited.
Network analysis (via graph theory) is a brain mapping approach which allows researchers to quantify network connectivity within the extensively interconnected human brain. [14][15][16] By using network analysis techniques, it is possible to extract topological features of brain connectivity and examine network efficiency, integration and the strength of connections within a brain network. 14,15,17 Recently, studies have begun to apply network analysis to functional magnetic resonance imaging (MRI) and diffusion-tensor imaging (DTI) data to identify topological features of network disruption associated with alcohol consumption, 18 dependence 19 and abstinence. [20][21][22] This approach has also been used to identify potential risk factors for alcohol dependence [23][24][25] and examine how potential treatment approaches may impact these networks. 26,27 Across the majority of these studies, reductions in global efficiency, 18,24 local efficiency 20,22,25,28 and clustering coefficient, 20,23,25 as well as the presence of abnormal functional hubs 29 were observed in individuals with alcohol dependence relative to healthy controls. However, while these group differences were not found significant by all studies, 19,22,30 a relationship between these network features and clinical aspects of alcohol dependence (such as alcohol use duration, severity of alcohol use and length of abstinence) has been identified. 19 The studies described above construct network connectivity graphs from high spatial resolution neuroimaging data (acquired from an entire scan session). Electroencephalography (EEG), an economic and convenient neuroimaging tool with high temporal resolution, is capable of characterising brain network connectivity within a millisecond timescale. [31][32][33] To date, only a limited number of EEG studies have explored the presence of altered network topology associated with short-term 34 and long-term 35,36 alcohol consumption, and these studies have presented mixed findings. Therefore, preliminary studies, while providing broad support for the use of EEG to identify network anomalies, examined network connectivity from multiple levels (resting-state and task-activated) and utilised a diverse range of techniques (EEG and combined EEG-MEG) and connectivity analyses (phase-synchronisation index, coherence and cross-correlation) as well as varied statistical analyses techniques (group-wise comparisons and data mining models) which may account for discrepant results. In the current study, we propose that implementation of a more targeted approach, which directly activates the addiction circuitry while the network response is quantified, may provide further insight into the network topology associated with long-term alcohol use.
Combined transcranial magnetic stimulation and electroencephalography (TMS-EEG) is a novel technique which allows researchers to non-invasively perturb the brain network (elicit a transient change in the network state) while EEG measures the network response within a millisecond timescale. [37][38][39][40][41][42][43] Healthy controls typically present with a globally efficient and highly integrated network in response to TMS perturbations, while disturbances in network response have been identified across a range of psychiatric populations. [44][45][46][47][48][49] By utilising this novel approach, researchers are now capable of applying a more robust and direct perturbation to regions within the mesocorticolimbic 'addiction' circuitry. Previously, our research group delivered a paired-pulse TMS paradigm (long interval cortical inhibition [LICI] 42,50,51 ) to the frontal cortex of individuals with alcohol dependence (ALD) post-detoxification to transiently inhibit cortical activity, while EEG measured the cortical response. 52 This study provided the first direct report of altered cortical excitability (reduced cortical inhibitory [GABAergic] neurotransmission) localised within the frontal regions. 52 However, as the frontal brain region does not act in isolation, it is possible that applying LICI to the frontal regions may also be capable of inducing a transient perturbation to the global network architecture. Therefore, the current study expands on this research and examines the global effects of the TMS-perturbation on the distributed network activity of patients with ALD in early recovery.
Therefore, the current study had three major objectives: (i) Characterise global properties of network connectivity following a perturbation (TMS pulse) in patients with ALD in early recovery when compared to healthy controls, (ii) assess whether altered network features can predict the likelihood of membership in the ALD group and (iii) examine whether these global properties are related to clinical features of alcohol dependence.

| Experimental design and techniques
For the current study, network analysis techniques were applied to TMS-EEG data previously collected and published. 52 A brief overview of these data collection processes is also detailed in the current publication.

| Transcranial magnetic stimulation
Biphasic TMS pulses were applied to the cortex via a figure-of-eight cooled coil connected to the MagPro R30 stimulator with a Motor Threshold (AMT) and 1 mV peak-to-peak (measured over cap).
RMT is determined by the minimum stimulus intensity required to elicit peak-to-peak MEP > 50 μV as measured by the contralateral FDI muscle in at least 5/10 trials (Rossini et al., 1994). AMT is determined as the minimum stimulation intensity, during active FDI muscle contraction (uniformly pressing index finger against a spring), needed to induce peak-to-peak MEPs of >100 μV in 3/5 consecutive trials. For the 1-mV peak-to-peak measure, stimulator intensity is adjusted until it elicits peak-to-peak MEP of approximately 1 mV over 10 consecutive trials; this was conducted over the EEG cap (1 mV measure over cap) and measured by the contralateral FDI muscle contraction to determine the appropriate stimulation parameters for the application of frontal TMS-EEG.

| Measurement of frontal cortical inhibition
Long interval cortical inhibition (LICI) is an inhibitory paired-pulse TMS paradigm which delivers two suprathreshold pulses with the stimulus intensity based on the 1-mV peak-to-peak measure (established by the motor stimulations). The process involves applying the suprathreshold conditioning pulse, then the inter-stimulus interval (ISI) of 100 ms, which is followed by a suprathreshold test pulse. This results in the conditioning pulse suppressing the cortical response produced by the test stimulus. This process is illustrated in Figure 1.
where V i is the time average of V i . To account for cross-cortical conduction times and neurophysiological processes, 60 the delay τ ranged between 0 and 150 ms and was chosen as the delay time maximising the cross correlation. Consideration of this maximal delay τ (up to 150 ms) led to the construction of a directed network.
Notably, although quite rare, when the pairwise delay approached 150 ms (i.e., between distal electrodes), it was reasonable to extract meaningful connectivity data from 150 ms (when examining the 300-ms epoch). The global mean degree was examined across a maximal delay range of 150, 100 and 50 ms when single TMS was applied to the left PFC, see Figure 2. In the current study, the single-pulse condition reflects a baseline condition, as such, based on the data exhibited in Figure 2, the network response appears to be comparable between the groups when a maximal delay lag of 150 ms is applied.
An edge e ij is then defined as the maximum of c(τ) over all τ. If e ij > e ji , only e ij is kept and e ji is set to zero. Defining e ij in this manner represents statistical dependence of the signals between the nodes over the fixed time span. This results in a graph which is both directed (i.e. all edges are directed from one node to another) and weighted (retaining the edges' correlation co-efficient index). Individual networks were constructed for each of the participants. Please refer to For each network, a global threshold of 60% was set to remove weak or spurious correlations (i.e., those edges whose weights are close to 0) and thus allow identification of topological properties of the network. A fixed threshold allows an examination of the average number of active connections between nodes and allows us to identify how this differs between the groups. Before settling on the 60% absolute threshold, the global mean degree was examined across a range of thresholds for both stimulation sites (left and right PFC) and both pulse-type (single and pairedpulse), see Figure 4. In the current study, the single-pulse condition is considered the baseline condition, therefore, based on the data shown in Figure 4, we noted the different topological features of the network following the LICI perturbation emerged when the threshold was set at 60% (and similar patterns at 70%). Considering that this is an exploratory study, utilising novel combined TMS-EEG techniques with network analysis, it is recommended that future studies explore the optimal methods of thresholding for TMS-EEG studies, as a more extensive examination was not within the scope of the current paper.
F I G U R E 2 Phase lag: The global mean degree (mean and standard error) across a maximal delay range of 150,100 and 50 ms following single TMS to the left prefrontal cortex (PFC) are presented for Healthy Controls (HC) and individuals with Alcohol dependence in early recovery (ALD). In the current study, the network response to single pulse TMS is considered as a baseline condition and the network response appears to be comparable between groups at a maximal delay of 150 ms F I G U R E 3 Network analysis pipeline. (A) Electroencephalography (EEG) signals were acquired throughout the transcranial magnetic stimulation (TMS)-EEG session. Normalised pairwise cross-correlations between each of the signals were used to construct a correlation matrix. (B) For each participant, individual correlation matrices were constructed for each Stimulation Site (left prefrontal cortex and right prefrontal cortex) and pulse type (single pulse and paired-pulse). (C) From these correlation matrices, network measures were extracted: mean degree, clustering coefficient, distance and shortest path length (local and global efficiency). These network measures were then used to compare global features of network response to a TMS perturbation across the individuals with alcohol dependence in early recovery (ALD) and healthy controls (HC)

| Degree (k)
In a directed network, the degree k of a node denotes the number of edges, or links connecting inward (in-degree) and outward (out-degree) from the node. Nodes of high degree demonstrate increased connectivity with the other vertices and may represent hubs in the network.

| Clustering coefficient
The clustering coefficient C i quantifies the occurrence of two neighbors of the same node V i being connected, which produces a triangle in the graph. C i presents information regarding local connectivity and structure within a network. 61 If t i is the number of triangles that node V i participates in, then C i = 2t i /k i (k i À 1), where k i is the degree of node V i . Global clustering coefficient computes the average clustering coefficient of all nodes.

| Path length, local efficiency and global efficiency (measure of integration)
The characteristic path length L ij between two nodes refers to the minimum number of edges which is required to pass from node V i to node V j (i.e., topological distance) and is also described as the shortest path length. We set L ij ¼ ∞ for any disconnected node pairs V i and V j .
The global efficiency GE is defined as the average inverse shortest path length distance in the network GE ¼ The local efficiency LE is a bit more delicate, defined as the average inverse shortest path length of the sub-graph that includes all the neighbors of a node, but with the node itself removed (otherwise all shortest path lengths would be either 1 or 2). Local efficiency is thus related to the clustering coefficient, for example, if all of a node's neighbors are linked then both its clustering coefficient and its local efficiency are maximal, and equal 1. These efficiency measures reflect the transfer of information and capacity for integrated processing of the network, both globally (for the GE measurement) and locally (for the LE measurement).

| Statistical analysis
Comparability of the basic demographics of the ALD group and healthy controls was assessed using independent t-tests for continuous variables and χ 2 -tests for categorical variables (Table 1). All data analyses were performed using IBM SPSS Statistics 28 and tests were run at alpha level of 0.05. There were no significant violations of homogeneity of regression or unequal variance. Global F I G U R E 4 Determining the global threshold: To identify the optimal threshold to apply for network construction, the network response (global mean degree) to the transcranial magnetic stimulation (TMS) perturbation across stimulations site (left and right PFC) and pulse type (single and paired-pulse) were compared between healthy controls (HC) and individuals with Alcohol dependence in early recovery (ALD) for a range of global thresholds (0.5 through to 0.9). It is notable that significantly altered topological features of the network following paired-pulse (LICI inhibitory pulse) emerge at a threshold of 0.6 while no difference is observed in the single pulse condition network metrics (mean degree, clustering coefficient, local and global efficiency) were extracted for each participant (described in Table 2) and a three-way mixed measures ANOVA pipeline 62 guided by the process described in Laerd Statistics (https:// statistics.laerd.com) was conducted to identify the presence of significant interactions between group (ALC vs. HC), pulse (single pulse vs. paired-pulse) and side (left vs. right PFC stimulation).

Across these network metrics, Binomial Logistic Regression and
Receiver Operating Characteristic (ROC) Curve analysis were applied to ascertain the likelihood for these network metrics (predictors) to successfully discriminate between participants with ALD or HC.     of the variance in the ALC group and correctly classified 68% of cases. Sensitivity was 60%, specificity was 75%, positive predictive value was 69% and the negative predictive value was 67%. Local efficiency following TMS applied to the left PFC was statistically significant as an independent predictor (as shown in Table 3).
Whereby, each unit reduction in Local Efficiency increases the odds of being in the alcohol-dependent groups by a factor of 3.759 (based on inverted odds ratio). Therefore, decreasing Local Efficiency was associated with an increased likelihood of membership in the ALD group. As this is an exploratory study, it is notable that following TMS applied to the left PFC, Global Efficiency followed a remarkably similar pattern, but is not considered a significant predictor at p = 0.052. MD and CC were also independently assessed via binomial logistic regression and were both found to not contribute significantly to the model.

| DISCUSSION
This is the first study to utilise an inhibitory paired-pulse paradigm (LICI) combined with EEG to identify anomalies in network connectivity in a sample of participants with alcohol dependence during early recovery. We find it significant that following single-pulse stimulation,

| Altered network metrics and alcohol dependence
The first network feature explored in the current study was Degree Centrality (hub-like structure) which represents the integration and processing of network information and communication within the network. 63 From a clinical perspective, it has been found that a global decrease in nodal degree is associated with longer duration 19 and with greater severity of alcohol dependence. 34 In the current study, the ALD group exhibited altered degree centrality in response to the LICI perturbation compared to HC. Moreover, following paired-pulse stimulation to the left PFC, degree centrality as network classifier could adequately discriminate between the groups. This classification was not observed following single-pulse or right-side stimulation. Additionally, in the ALD group, reductions in degree centrality were associated with an increase in alcohol dependence severity. Previous research has demonstrated that short term inebriation elicited increased degree centrality of various nodes in the resting-state condition, 34 while long-term alcohol use was associated with the presence of abnormal functional hubs (varied across regions within the network). 29 However, emerging neuroimaging studies (fMRI, DTI and EEG) into long-term alcohol use are still mixed regarding the direction of these network anomalies, with some studies reporting decreased nodal degree in various brain regions 29 and other studies reporting increased nodal degree. 30 Another key topological feature of the brain network that has been examined by addiction studies is locally specialised connectivity of the network. The clustering coefficient, which indexes the number of connections existing between a node's nearest neighbours, provides a measure of locally specialised connectivity. 14 Complex networks present with high clustering, while reduced clustering may indicate a shift towards random organisation. Reduced clustering has been previously associated with increased severity of alcohol use and duration of alcohol dependence. 19 In the current study, the ALD group presented with a lower clustering coefficient across the network in response to the LICI perturbation but not in response to single pulse stimulation. Following paired-pulse stimulation (only) to the left PFC, clustering coefficient network metrics were capable of discriminating between the groups. Additionally, in the ALD group, a decrease in clustering coefficient was associated with an increase in alcohol dependence severity. Currently, there are only limited fMRI and DTI findings regarding group differences in clustering coefficient following alcohol exposure. Rank-ordered differences were found with values of nodal clustering being lowest in the ALD group (compared to healthy controls and unaffected siblings of the ALD participants). 25 Reduced clustering has also been observed in an abstinent addicted population 20 and in individuals with familial risk of developing alcohol dependence. 23 However, a number of other studies failed to observe any significant group differences on global measures of clustering. 19,30 With regard to the preliminary EEG studies, data modelling was applied to EEG data to ascertain the predictive value of detecting an ALD group compared to healthy group based on network features.
Reduced clustering coefficient was found to reliably predict the ALD group under various stimulus conditions. 64 This was supported by one EEG study in males with alcohol dependence after approximately 28 days of detoxification, which identified reduced clustering (in the low beta band), 36 while the second EEG-MEG study observed group difference in local connectivity (decreased clustering at the posterior sites and increased clustering at the frontal sites) but no significant differences in global connectivity. 35 These studies, when combined with the current study, provide initial evidence of compromised locally specialised connectivity associated with long-term alcohol use. Additional studies are required to further characterise how these anomalies relate to relapse and recovery.
A prominent focus of addiction network studies is the efficiency (a measure of information transfer) of the brain network. Measure of Efficiency, which is represented by a network with short characteristic path length, provides a measure of parallel information transfer between nodes and integrated processing. 14 High level functioning relies on efficient information transfer, which is also beneficial for cognitive control and executive function. 18,65 From a clinical perspective, decreased global efficiency has been found to relate to a longer history of alcohol dependence, 19 while increased local efficiency was related to greater duration of abstinence. 28 In the current study, participants with ALD presented with reduced local efficiency in response to the LICI perturbation but not in response to single pulse stimulation. Moreover, it was found that as local efficiency reduced, the predictive odds of being in the ALD group increased. With regards to global efficiency, participants with ALD demonstrated reduced global efficiency in response to left sided stimulation. Following paired-pulse stimulation to the left PFC (only), local and global efficiency network classifiers were able to adequately discriminate between the groups. This is consistent with fMRI and DTI studies which have generally identified that decreased efficiency is associated with chronic alcohol use 18,28 and persists during abstinence, 20 while increased efficiency was found to be related to remission. 22 Additionally, individuals considered at risk of developing alcohol dependence (familial risk and children with foetal alcohol spectrum disorder) also presented with significantly reduced efficiency. 23,24 However, these significant group differences were not identified by all studies. 19,30 While many fMRI and DTI studies predominantly identify reduced global efficiency, only a small number of EEG studies have characterised these network anomalies and present mixed findings. In terms of short-term alcohol consumption, a preliminary resting-state EEG study examined network connectivity in social drinkers following alcohol administration as compared to placebo in healthy social drinkers. 34 Short-term inebriation resulted in increased global efficiency in the resting-state network. 34 With regards to long-term alcohol exposure, two studies examined network connectivity of males with alcohol dependence after approximately 28 days of detoxification. 35,36 The first EEG study examined network response during a working memory task and identified shorter characteristic path length and increased global efficiency. 36 The second study utilised EEG-MEG during resting-state and observed no group differences in global efficiency. However, differences in local efficiency were identified, specifically, the presence of decreased efficiency at the posterior sites and increased efficiency in the frontal sites of the ALD group. 35 Therefore, while fMRI and DTI neuroimaging studies provide compelling evidence of reduced local and global efficiency associated with long-term alcohol use, the findings from EEG studies are less consistent. In the current examination of local efficiency, there were no significant group differences observed in response to the single stimuli; the ALD group exhibited reductions in local efficiency, which are comparable with the fMRI and DTI studies, only when the paired-pulse inhibitory paradigm was applied to the addiction network. Notably, while the findings regarding global efficiency followed a similar pattern (but did not reach significance regarding the comparison between LICI and single pulse administration), there were significant differences in network response to the TMS perturbation between groups, it is anticipated that a future study with a larger sample size is required to further explore the response to LICI (as compared to the single pulse).
Therefore, the current study identified a shift in the balance of network integration and segregation following a paired-pulse TMS perturbation to the left PFC. Human brain networks have been found to exhibit small world networks which rely on an efficient balance between information segregation and integration (via high clustering and high global efficiency). 66 Across a range of psychiatric disorders, a loss to small-world organisation have been identified, with shifts in small-world features depending on the pathology of the disorder. 67 In the current study, we have identified that ALD patients exhibit altered small-world organisation (reductions in both efficient segregation and integration), which persist beyond detoxification. Therefore, we suggest that high-cost elements of brain network organisation (such as high clustering and high efficiency) 68 appear to be compromised in the ALD patient group, and moreover, the persistence of this altered topological organisation (beyond detoxification) may indicate an enduring risk to the economic organisation of the brain network.

| Altered brain networks, brain stimulation paradigms and models of addiction
Recent studies have provided preliminary evidence for the utility of combined neuroimaging-brain stimulation techniques to map and explore network change. 48,69 The present study is the first to apply LICI to examine global network anomalies underlying alcohol dependence. According to neurobiological models, acute alcohol exposure facilitates GABAergic inhibitory activity resulting in an overall inhibitory effect. [70][71][72][73] Chronic alcohol exposure elicits a compensatory response, whereby the brain seeks to restore equilibrium in neural function, which leads to neuroadaptation in the mesocorticolimbic 'addiction' circuitry, and subsequent suppression of GABAergic neurotransmission and facilitation of glutamatergic neurotransmission. 13,74,75 Cortical inhibition (CI), the neurophysiological process wherein GABA inhibitory interneurons selectively dampen activity of other neurons in the cortex, is critical to understanding the development of alcohol dependence. 76 It has been found that application of LICI, a paired-pulse inhibitory TMS paradigm, is capable of noninvasively characterising CI of the prefrontal cortex. 41,42,77 This approach has been utilised to identify altered cortical excitability within the frontal cortex of participants with alcohol dependence post-detoxification. 52 However, as the frontal cortex is intricately interconnected with widely distributed brain regions, it was important to explore whether the compensatory response (i.e. the neurophysiology of widespread suppression of GABAergic neurotransmission) may also be identified at the network level.
Recent studies have examined whether the neurophysiology of short-term alcohol exposure can be reflected by alterations in global connectivity. One preclinical study identified alterations in network organisation following acute ethanol exposure. 78 This was also observed in a study of social drinkers, whereby acute ethanol had a measurable effect on brain networks and appears to increase network density (hub-like behaviour) and global efficiency (global integration). 34  Furthermore, as per the nature of cross-sectional addiction studies, it is very difficult to assert whether the identified altered network connectivity is a direct result of chronic alcohol exposure, or due to preexisting vulnerabilities, or quite possibly, the combination of both. We anticipate that future longitudinal studies could expand on these findings and examine whether altered brain networks are a possible predictor of subsequent vulnerability to alcohol dependence, or whether it occurs predominantly as a consequence of long-term alcohol exposure. Regardless, results from the current study are promising and provide initial proof of concept for utilising TMS-EEG paradigms to examine anomalies in brain connectivity associated with alcohol dependence; however, these findings are preliminary and further studies are required to confirm these findings.

| Conclusion
The current study is the first to directly target the addiction circuitry (via an inhibitory TMS paradigm) whilst simultaneously utilising EEG to quantify anomalies in network topology associated with persisting features of alcohol dependence. Identified network anomalies included reduced global integration, locally specialised connectivity, and hub-like structures. These network alterations were related to clinical features of alcohol dependence, whereby reduced global integration was predictive of high alcohol use, while decreased locally specialised connectivity and hub-like structures were associated with increased alcohol use severity. Therefore, whilst preliminary, the current study provides compelling evidence of the potential efficacy of TMS-EEG combined with network science to identify network biomarkers associated with alcohol dependence and early recovery.
Further studies are required to confirm and extend these preliminary findings. A better understanding of the global features of the network may contribute to current neurobiological models of alcohol dependence. Additionally, it is anticipated that these network biomarkers may be used to identify risk factors associated with the development of alcohol dependence and as potential markers of treatment efficacy.

ACKNOWLEDGMENTS
Sincere appreciation is expressed to Nigel Rogasch, Jerome Maller, Neil Bailey and Karyn Richardson for their support and contribution with recruitment and data acquisition during the study and Dominik Freche for his assistance with assessing the TMS-EEG data. Appreciation is also directed towards the Israel Science Foundation grant (No.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.