FIGURE. 1. The experimental design (1a-1c) and the analysis conditions (1d-1g) of the present tri-fMRI hyperscanning study. (a) The fixed roles of the triads respectively in the three fMRI sites across Taiwan, National Cheng-Kung University (NCKU) as Player A, National Taiwan University (NTU) as Player B, and National Cheng-Chi University (NCCU) as Player C. NTU and NCCU are in northern Taiwan, 8 km away, while NCKU is in southern Taiwan, 305 km apart from NTU and NCCU. While Player A and Player B are interacting, Player C is interacting with Player D (who is outside of the NCKU MRI scanner). (b) There are 6 runs, 3 blocks in each run, and 6 trials in each block. Each run lasts about 7~8 minutes. In each trial (≈ 19 seconds), the four players are all playing the coordination game simultaneously, with two assigned pairs playing with each other (e.g., A&B+C&D), in different orders (e.g., A&C+B&D, and A&D+B&C). (c) The coordination game with preplay communications consists of two stages, each with a decision and a feedback for every player. The feedback in Stage 2 (17~19 s) is the target TR for coherence analysis because it is the actual choice (with rewards) revealing time for the interacting pairs to form trust, and to carry this successful history into the next trial. (d) W ithin-groupI nteracting P airs (WIP) are pairs that interact in the same group (e.g., A and B in Group 1 are the interacting pair). For fMRI analysis, there are 156 interacting pairs in total (3 players x 2 pairings [pairing with D is excluded] x 26 groups).(e) W ithin-group N on-interactingP airs (WNP ) are those who belong to the same group but interact with others at the same time (e.g., in Group 1, when A and B are interacting, A&C and B&C are non-interacting pairs.). There are 312 WNPs in total (3 players x 4 pairings x 26 groups). (f)B etween-group P ermuted P airs (BPP ) pair from different groups (e.g., A from Group 1 matches with B from Group 2, or A from Group 1 matches with C from Group 2, etc.). There are 3,900 BPPs in total (3 players x 50 pairings x 26 groups).
fMRI data acquisition and preprocessing
The fMRI images were acquired simultaneously in three MRI scanners, one in NCKU Tainan, another in NTU Taipei, and the other in NCCU Taipei. NCKU is about 305 km apart from NTU and NCCU, and NTU is about 8 km apart from NCCU (Figure 1a). The MRI scanner at NCKU Mind Research and Imaging Center is a 3-Tesla General Electric Discovery MR750 (GE Medical Systems, Waukesha, WI), equipped with an 8-ch head coil. Whole-brain functional scans were acquired with a T2* EPI (TR = 2 s, TE = 33 ms, flip angle = 90°, 40 axial slices, voxel size = 3.5 × 3.5 × 3 mm3). High-resolution T1-weighted structural scans were acquired using a 3D fast spoiled grass (FSPGR) sequence (TR = 7.65 ms, TE = 2.93 ms, inversion time = 450 ms, FA = 12°, 166 sagittal slices, voxel size = 0.875 × 0.875 × 1 mm3). Another fMRI scanner, which is located at NTU (Imaging Center for the Body, Mind, and Culture Research) is a 3-Tesla PRISMA (Siemens, Erlangen, Germany) scanner equipped with a 20-channel phased array coil. Whole-brain functional scans were acquired with a T2*-weighted EPI (TR = 2 s, TE = 24 ms, flip angle = 87°, 36 axial slices, voxel size = 3 × 3 × 3 mm3). High-resolution T1-weighted structural scans were acquired using a MP-RAGE (TR = 2.0 s, TE = 2.3 ms, inversion time = 900 ms, FA = 8°, 192 sagittal slices with 0.938 × 0.938 × 0.94 mm3 voxels without an interslice gap). The other scanner fMRI scanner, which is located at NCCU (Taiwan Mind & Brain Imaging Center) is a 3-Tesla MAGNETOM Skyra (Siemens, Erlangen, Germany) scanner equipped with a 20-channel phase array coil. Whole-brain functional scans were acquired with a T2*-weighted EPI (TR = 2 s, TE = 24 ms, flip angle = 87°, 36 axial slices, voxel size = 3 × 3 × 3 mm3). High-resolution T1-weighted structural scans were acquired using an MP-RAGE (TR = 2.53 s, TE = 3.3 ms, inversion time = 1100 ms, FA = 7°, 192 sagittal slices with 1 × 1 × 1 mm3 voxels without an interslice gap). Although these three scanners vary in quality and parameters, the highly overlapping histograms in rTPJ beta Figure 2a) and coherence value (Figure 2b) among NCKU (MR750), NTU (Prisma), and NCCU (Skyra) underpin our data-pooling later.
The fMRI data were preprocessed and analyzed using BrainVoyagerQX v. 2.6 (Brain Innovation, Maastricht, The Netherlands) and NeuroElf v1.1 (https://neuroelf.net). After slice timing correction, functional images were corrected for head movements using the six-parameter rigid transformations, aligning all functional volumes to the first volume of the first run. High-pass temporal filtering (with the default BVQX option of GLM-Fourier basis set at 2 cycles per deg, but no spatial smoothing) was applied. The resulting functional data were co-registered to the anatomical scan via initial alignment (IA) and final alignment (FA), and then both functional (T2* EPI data) and anatomical (T1 structural data) files were transformed into the Talairach space.
General linear model (GLM) and psychophysiological interactions (PPI) analyses
The GLM estimated each participant’ brain activations and compared the contrasts of interests. In order to make predictions for the blood-oxygen level dependent (BOLD) responses of various periods during the coordinating interaction, we divided each trial into four phases: 1) Stage 1 Decision, 2) Stage 1 Feedback, 3) Stage 2 Decision, and 4) Stage 2 Feedback. To test if the BOLD responses were different in successful and failed reciprocity trials (i.e., Successful vs. Failed), contrast analysis was performed on all subjects. Alphasim was adopted to correct the error rate of multiple comparisons among whole-brain voxels. Under the threshold of familywise error (FEW) p < 0.05, the minimum cluster size with a threshold of p < 0.01 was defined as 44 voxels.
In addition to the GLM contrast analysis, the psychophysiological interactions (PPI) analysis was applied to examine differences of functional connectivity between the seed region of interest (i.e., the rTPJseed) and target ROIs. The rTPJseedwas heavily involved in the ToM, attention, and social networks (Koster-Hale & Saxe, 2013; Premack & Woodruff, 1978; Wang et al., 2018; Young et al., 2010). In a social cognition review (Babiloni & Astolfi, 2014), the rTPJ activation has been consistently identified during tasks related to the process of information. In another review of more than 200 fMRI studies (Van Overwalle, 2009), the rTPJ has its crucial role in transient mental inferences about other people. A recent review of patients with damaged rTPJ even suggests its role in predicting others (Masina et al., 2022). The rTPJseed[54, -55, 22] was downloaded from neurosynth.org, comprising 103 published studies, with z scores 8.82 (corresponding to FDR=0.01, with ‘association test’ option). It is a functional mask with 239 voxels. In addition, to investigate the individual differences in the total reward under the contrast of successful and failed reciprocal interactions, and such relationship with their functional connectivity, participants’ total reward was correlated with their connectivity differences to reveal behavior-relevant brain networks. All correlation analyses were conducted using Matlab and JASP (https://jasp-stats.org/).
Data preparation for the frequency-domain coherence analysis
The procedure of converting a jittered event-related fMRI dataset into a format ready for coherence analysis was reported earlier (Wang et al., 2023). This time-frequency coherence analysis was adopted here to estimate inter-brain couplings. As coherence requires long duration in a predictable manner, we created a periodicity (18 s), targeting only on the feedback stage and lasting approximately as long as a run of the experiment (≈ 19 s). These beta-concatenated time series were extracted by iteratively estimating feedback-related blood oxygenation level-dependent responses of each trial. Details are as follows.
First, the feedback period of any given trial (here the second stage) was used (Figure 2a). Next, the protocol (or .prt file) of an fMRI run (totally 6 runs), along with the voxel time series, was created. Then, a single deconvolutional GLM was created, with 9 stick-like extensions (Figure 2b). This method allowed a least-square separation of the target trial, from which the whole brain beta series were independently estimated subsequently for the 9 TRs following the feedback-initiated event, with a finite impulse response (FIR) function (Turner et al., 2012). The feedback time, revealing the result of the actual choices from the 17th second to the 19th, was the time the dyad built up the impressions about each interacting partner. All the 9-TR volumes from each trial (6 in one run) and for the 6 runs were concatenated, resulting in 324 volumes (9 TRs x 6 trials x 6 runs), as the periodic data format ready was for coherence analysis (Figure 2c).
Time-frequency seed-based coherence analyses were conducted across pairs of choice by using the FieldTrip toolbox (https://www.fieldtriptoolbox.org/). The rTPJseed was the same as adopted in PPI analysis. Coherence analyses were performed between one’s (e.g., A) rTPJ to another’s (e.g., B) whole brain, and B’s rTPJ to A’s whole brain, and then averaged across a dyad, and then across all 26 groups under specific pairings (detailed later in 2.6) (Figure 2d). The groupwise whole-brain mask was derived from averaging over the whole 78 participants. The coherence operation yielded 162 frequency bins, out of 0.5 (fMRI sampling frequency) /324/2 (Nyquist equation) = 162 frequency bins. Since we estimated 9 beta series (each for 2 seconds), the 36th (1/18 s = 0.0556 Hz) frequency bin would be the frequency of interest. The regions with high coherence were separately mapped (Figure 2e). For the result presentation, the condition-wise whole-brain mapping was shown via NeuroElf.