Hyperscanning Literature After Two Decades of Neuroscientific Research: a Scientometric Review

Hyperscanning, a neuroimaging approach introduced in 2002 for simultaneously recording the brain activity of multiple participants, has significantly contributed to our understanding of social interactions. Nevertheless, the existing literature requires systematic organization to advance our knowledge. This study, after two decades of hyperscanning research, aims to identify the primary thematic domains and the most influential documents in the field. We conducted a scientometric analysis to examine co-citation patterns quantitatively, using a sample of 548 documents retrieved from Scopus and their 32,022 cited references. Our analysis revealed ten major thematic domains in hyperscanning research, with the most impact-ful document authored by Czeszumski and colleagues in 2020. Notably, while hyperscanning was initially developed for functional magnetic resonance imaging (fMRI), our findings indicate a substantial influence of research conducted using electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). The introduction of fNIRS and advancements in EEG methods have enabled the implementation of more ecologically valid experiments for investigating social interactions. The study also highlights the need for more research that combines multi-brain neural stimulation with neuroimaging techniques to understand the causal role played by interpersonal neural synchrony in social interactions.


Highlights
-We analyzed the whole literature on hyperscanning with a quantitative approach -EEG and fNIRS studies allowed for implementing more naturalistic tasks -The need for more research using multi-brain stimulation emerged comparability (e.g., Bizzego et al. (2022c)).Thanks to its characteristics, 50 fNIRS allows the implementation of experimental tasks that mirror daily 51 life activities.For instance, Figure 1 shows an fNIRS hyperscanning setup 52 to measure the brain activity in mother and child during free play (Bizzego   causation-focused approaches (Novembre and Iannetti, 2021).Traditional 58 hyperscanning studies typically measure interpersonal neural synchrony 59 as a dependent variable resulting from social interactions between part-60 ners.While this approach yields interesting and valuable results, it alone 61 cannot determine whether interpersonal neural synchrony is merely an 62 epiphenomenon of the fact that participants are being exposed to the same 63 environment and the same stimuli or if it plays a causal role in facilitating 64 social exchanges.A causation approach based on brain stimulation takes 65 the opposite approach by manipulating neural activity and studying its 66 effects on social interaction (for a recent multi-brain stimulation study, see 67 (Lu et al., 2023)).Figure 2 displays an example of how sensory stimulation 68 can be used to entrain participants' brain activity during social interactions.(Hamilton, 2021).In light of this, the 74 current paper aims to identify the main domains of research as well as 75 the most impactful documents in the hyperscanning literature.To do 76 so, we will adopt a scientometric approach to reviews, as done in our 77 previous publications (e.g., Carollo et al. (2021)).A scientometric approach 78 was chosen because it merges bibliometric analysis (i.e., application of 79 quantitative techniques to bibliometric data) and scientific mapping (i.e., 80 visualization of the temporal evolution of a research domain) (Carollo et al., 81 2021).Therefore, the scientometric approach allows the use of a data-driven 82 quantitative method to uncover the main research domains and impactful 83 documents in large samples of data.

99
The current scientometric analysis goes beyond the scope of the ini-100 tial 500 documents obtained from Scopus.These documents serve as the 101 primary layer of data collection, from which additional documents (i.e., 102 all their references) will be gathered and utilized for the comprehensive 103 analysis.This approach is crucial for two key reasons.Firstly, it ensures 104 a substantial sample size, higher than previous reviews on the topic, for a robust analysis.Secondly, it facilitates the collection of literature not necessarily indexed in Scopus but still pertinent to the hyperscanning field.This is because, in principle, cited documents are selected by field experts based on their scientific relevance.
To characterize the structure of knowledge in the hyperscanning literature, we first analyzed the collected data using the bibliometrix package for R (Aria and Cuccurullo, 2017).We used bibliometrix to detect the most involved countries, scientific journals, authors, documents, and keywords in the hyperscanning literature.

Data import on CiteSpace
To conduct the scientometric analysis, all data were imported into CiteSpace (version 6.1.R6 64-bit Advanced) (Chen, 2006).A total of 32,022 cited references were identified, of which 31,802 (99.31%) were in the valid format to be included in the analysis.The data loss at this stage (n = 220 references; 0.69% of the total references) is considered acceptable and it is in line with previous scientometric works (e.g., Carollo et al. (2021)).

Document Co-Citation Analysis (DCA)
To identify the main research trends and impactful documents, we conducted a document co-citation analysis (DCA).DCA assesses the frequency with which two or more publications are cited together (i.e., co-cited) by other documents (Trujillo and Long, 2018).The assumption is that two or more documents are frequently co-cited because they belong to the same thematic domain of research.For this reason, co-citation patterns can provide information on the relationships between key concepts, methods, or experiments in the literature (Small, 1973(Small, , 1980).In the DCA, single documents are included as the network's nodes, co-citations are included as links, and co-citation frequencies as link weights.
The number of documents that are included in the final graph depends on the node selection criterion and its scaling factor.As done in our previous scientometric analyses (e.g., Carollo et al. ( 2021)), we optimized the parameters to obtain a well-balanced DCA network.Specifically, we compared the networks generated when using g-index, TOP N, or TOP N% as node selection criteria.The g-index is the largest number where the total

Document Co-Citation Analysis 198
The optimal DCA network, depicted in Figure 5, was made of 648 199 nodes (i.e., documents) and 3065 links (i.e., co-citations).Thus, on average,      The major citing documents in cluster #1 were authored by Nam et al. suggested that a hyperscanning approach would be ideal to assess the 380 differences between in-person and virtual social interactions, which were 381 largely used in educational settings during the COVID-19 pandemic.479 standards and quality in their designs (Callaham et al., 2002).This is par-480 ticularly true for documents published by less prestigious institutions or 481 by disadvantaged groups (e.g., researchers who cannot afford to publish 482 their documents in high-impact journals).The current manuscript does not 483 aim to perpetuate a system where quantitative measures are prioritized 484 over qualitative ones.Therefore, in the Discussion section, we included a 485 qualitative review of the clusters (Hicks et al., 2015).Additionally, when 486 discussing the clusters, we not only reviewed the content of the cited docu-487 ments but also focused more on the citing documents.This approach helps 488 mitigate biases resulting from the quantitative properties of documents 489 since citing documents are included in the thematic clusters based on the 490 citations they make, rather than the citations they receive.

Figure 1 :
Figure 1: Representation of a functional near-infrared spectroscopy (fNIRS) hyperscanning setup to monitor the brain activity of mother and child during a play session.Image from Bizzego et al. (2022b) 57

Figure 2 :
Figure 2: An example of multi-brain stimulation through rhythmic sensory stimulation.
Data Collection from Scopus 86 For the current work, all data were collected from Scopus.Scopus was 87 chosen because it has a higher coverage of indexed journals as compared to 88 other platforms (Cataldo et al., 2022).Data were collected on 04 September 89 2023 with the searching string "TITLE-ABS-KEY(hyperscan*)".A sample 90 of 548 documents published between 1998 and 2023 was retrieved.We 91 qualitatively inspected the documents' abstracts and titles to ensure that the 92 included documents were relevant to the field of hyperscanning.Through 93 this procedure, we excluded an amount of 48 non-relevant documents.94 Thus, the final sample consisted of 500 documents published between 2002 95 and 2023.
received by the top g articles equals at least g (Egghe, 139 2006).The TOP N and TOP% criteria include in the final network the N or 140 the N% most cited references for each time slice (i.e., one year in this work).

141
While the node selection criteria specify the rule determining the inclusion 142 of nodes in the final network, their scaling factors set the threshold (Cataldo 143 et al., 2022).For the analysis, we compared the networks generated using 144 g-index with k set at 10, 15, 25, 50; N set at 50; and N% set at 10.The 145 metrics for all generated networks were compared and the optimal DCA 146 was obtained with g-index with k set at 25. 147 The literature search and the generation of the DCA network are sum-148 marized in Figure 3.

Figure 3 :
Figure 3: Preferred reporting items for systematic reviews (PRISMA) flowchart for literature search and references eligibility.

Figure 4 :
Figure4: Co-occurrence analysis on the keywords in the hyperscanning literature.In the network, keywords are represented as individual nodes, with their size proportional to their degree.Co-occurrences of keywords are indicated by solid links (for co-occurrences within the same cluster) and dashed links (for co-occurrences across clusters).The width of these links is proportional to the frequency of co-occurrence.Based on the co-occurrence patterns, two clusters of keywords were automatically identified using the bibliometrix package for R (Aria and Cuccurullo, 2017) and are depicted in the figure using red and blue colors. 197

200
each document was connected with another 4.73 documents.The network 201 was moderately divisible into separate and highly homogeneous clusters 202 (modularity = 0.6495; average silhouette score = 0.8334).

Figure 5 :
Figure 5: Document co-citation analysis network of the literature on hyperscanning.In the network, documents are represented as single nodes, and co-citations are represented as links.Ten major thematic domains of research were identified.The image was generated with CiteSpace software Chen (2006).

203
In the network, ten major thematic clusters of research were identified.204 The clusters that included the highest number of documents were cluster 205 #0 (size = 88; silhouette = 0.729; mean publication year = 2010), cluster #1 206 (size = 85; silhouette = 0.820, mean publication year = 2014), and cluster 207 #2 (size = 83, silhouette = 0.781, mean publication year = 2017).Based on 208 their silhouette score, the most homogeneous thematic clusters of research 209 were cluster #9 (size = 30; silhouette = 1.000; mean publication year = 2004), 210 cluster #12 (size = 13; silhouette = 0.991; mean publication year = 2007), and 211 cluster #5 (size = 37, silhouette = 0.961, mean publication year = 2008).After 212 inspecting the documents included in the clusters, we manually labeled 213 the clusters to reflect their thematic content as in Carollo et al. (2021).The 214 metrics of all the major thematic clusters of research are presented in 219 that were not recognized as duplicates by CiteSpace.Based on the strength 220 of their citation burstness, the most impactful documents in the network 221 were authored by Czeszumski et al. (2020) (citation burstness = 13.93;222 duration = 2 years), Lindenberger et al. (2009) (citation burstness = 13.71;223 duration = 6 years), and Hari and Kujala (2009) (citation burstness = 12.11; 224 duration = 5 years).In particular, Czeszumski et al. (2020) provided a 225 comprehensive review of the state-of-the-art in the hyperscanning literature, 226 focusing the discussion on the main neuroimaging techniques, types of 227 analysis, and research outcomes.A summary of the metrics of the top ten 228 documents with a citation burst is provided in

383
In cluster #3, the major citing documents were authored by Nam et al. 384 (2020) (coverage = 27; GCS = 21), Czeszumski et al. (2020) (coverage = 385 23; GCS = 153), and Schwartz et al. (2022) (coverage = 19; GCS = 5).In 386 line with Balters et al. (2020), the focus of cluster #3 appears to be the 387 investigation of interpersonal neural synchrony in virtual environments 388 (e.g., Barde et al. (2020)).With the increase in the use of virtual environ-389 ments for collaboration and social interactions (Barde et al., 2020), several 390 documents in the cluster examined whether in-person and virtual social 391 interactions share the same neural underpinnings.The preliminary studies 392 by Gumilar et al. (2021) and by Wikström et al. (2022) showed that inter-393 personal brain synchrony in virtual interactions mimics the one observed 394 in in-person interactions.However, contrasting results were reported by 395 Schwartz et al. (2022), who observed lower interpersonal neural synchrony 396 in technologically-assisted interactions as compared to in-person interac-397 tions between mother and child.

399
The major citing documents in cluster #2 were authored by Nam et al. 400 (2020) (coverage = 35; GCS = 21) and Kelsen et al. (2022) (coverage = 22; 401 GCS = 26).The common thematic interest of research in this cluster regards 402 the use of hyperscanning to investigate the neural mechanisms of verbal 403 communication (e.g., Jiang et al. (2021); Wang et al. (2022)).Particularly, 404 Kelsen et al. (2022) observed that patterns of interpersonal brain synchrony 405 among communicators predominantly emerge from frontal and temporo-406 parietal brain regions.As for the authors, synchronization in these regions 407 might reflect the activity of the mirror and mentalizing system.

492
The literature on hyperscanning has been strongly influenced by ten 493 key research trends.Over time, the field has evolved from discussions 494 about methods for EEG hyperscanning studies to the exploration of more 495 ecologically valid paradigms that simulate everyday life.The introduc-496 tion of fNIRS has enabled the implementation of hyperscanning designs 497 in more naturalistic settings (e.g., musical settings, educational settings, 498 parent-child interactions) while providing spatial information about the 499 brain activity under investigation.According to scientometric analysis, 500 the use of neuroimaging techniques in hyperscanning is well-established, 501 while the use of brain stimulation has not yet emerged, likely due to its 502 recent introduction as an approach.This finding suggests the need for 503 further studies that combine multi-brain stimulation with neuroimaging 504 techniques to begin interpreting interbrain neural synchrony in a causal 505 manner.506 Data and code availability statement 507 All data were retrieved from Scopus using the string "TITLE-ABS-508 KEY(hyperscan*)".CiteSpace software (Chen, 2006) was used for the scien-509 tometric analysis and bibliometrix package (Aria and Cuccurullo, 2017) was 510 used for the bibliometric analysis., GE; Methodology: AC; Formal Analysis: AC; 513 Investigation: AC; Writing-original draft preparation: AC; Writing-review 514 and editing: AC, GE; Supervision: GE.All authors have read and agreed to 515 the published version of the manuscript.

Table 1 :
Metrics for the ten thematic clusters of research identified in the network.

Table 2 :
Summary metrics of the ten with the strongest citation burstness.Citation burstness is defined as an abrupt increase in the number of citations received by a document.In the case of a repeated entry, only the reference with the highest value of citation burstness is presented in the table.The increased adoption of neuroimaging techniques with higher spatial 308 resolution in hyperscanning stemmed from and aimed to enrich years of 309 neuroscientific research on the neural correlates of social behaviors (e.g.,(2009), in one of the documents with the highest citation burst, used EEG 231work was to identify the main thematic domains of research and the most 307 313 documents were authored by Acquadro et al. (2016) (coverage = 20; GCS 314 = 38), Burgess (2013) (coverage = 19; GCS = 127), and Cornejo et al. (2017) 315 (coverage = 17; GCS = 43).After the pioneering studies by Funane et al. 316 (2011) and Cui et al. (2012), fNIRS gained momentum as the elective tech-317 nique for conducting ecological hyperscanning studies (Scholkmann et al., 318 2013).However, some solutions to enhance the ecological validity of EEG 319 hyperscanning were proposed too (e.g., Astolfi et al. (2012); Toppi et al. 320 (2016)).321Theinterest in ecologically valid experiments directed the research fo-322 cus to a core and largely neglected side of daily life social interactions: 323 the affective component (Balconi and Vanutelli, 2017; Cornejo et al., 2017).324Affectand emotions influence people's daily lives in multiple ways.For 325 instance, they are crucial in creating social cohesion as well as determining 326 the willingness to undertake joint actions or prosocial behaviors (Czeszum-327 ski et al., 2020; Konvalinka et al., 2011; Lopes et al., 2005; Twenge et al., 2007).328Despite the importance of emotions in daily life, they had been scarcely 329 investigated in hyperscanning studies due to the complexity of the setups 330 (Czeszumski et al., 2020).Acquadro et al. (2016) proposed that using hyper-331 scanning in musical settings represents one possible solution to investigate 332 the affective component of social interactions.For the authors, the use of 333 musical settings guarantees high ecological validity, the emotional compo-334 nent (which is a catalyst for social interactions), and an enactive view of 340 et al. (2011, 2012); Sänger et al. (2012, 2013)).Notably, Lindenberger et al.341 350 et al. (2018b) (coverage = 9; GCS = 7).In this cluster of research, initial 351 studies helped identifying some regions of interest from which patterns of 352 interpersonal neural synchrony frequently emerge (e.g., Astolfi et al. (2015); 353 Balconi et al. (2018c)).For instance, Nozawa et al. (2016) showed enhanced 354 interpersonal neural synchrony in frontopolar brain regions during natural 355 verbal exchanges between people.Interestingly, the same brain regions are 356 typically involved in social communication.Similarly, Hirsch et al. (2017) 357 found an increased interpersonal neural synchrony in the left superior 358 temporal gyrus, middle temporal gyrus, supramarginal gyrus, pre-motor 359 cortex, and supplementary motor cortex during mutual gaze.In light of the 360 increased interest in the neural basis of real-life social interactions, Koike 361 et al. (2015) suggested employing EEG-fMRI hyperscanning to combine 362 the high temporal resolution of EEG with the high spatial resolution of 363 fMRI.As argued by the authors, such an approach would allow the use of 364 inter-brain dynamics as a neuro-marker of real-life social interactions.3654.6.Cluster #1: Educational Settings

Limitations of the Study 472
(Carollo et al., 2022)d 2008, some scholars sought to gain more insight 459 into the brain regions supporting individuals involved in naturalistic inter-460 actions.fNIRSemergedas the elective tool to do so, as it has higher spatial 461 resolution and is less sensitive to movements than EEG(Carollo et al., 2022).462Theinterest in ecologically valid experiments led to considering more com-463 plex and previously neglected components of social interactions, such as 464 affect and communication patterns.465More recently, hyperscanning research has shown translational poten-466 tial in educational settings and in investigating the underlying mechanisms 467 supporting interactions in virtual environments.The most recent cluster 468 focuses on the use of hyperscanning to investigate parent-child interac-469 tions using real-life tasks, often employing a multimodal approach and 470 integrating behavioral, physiological, and neural data.471 5.