Cognition and Behavior
Team Flow Is a Unique Brain State Associated with
Enhanced Information Integration and Interbrain
Synchrony
Mohammad Shehata,
1,2
Miao Cheng,
1,3,4
Angus Leung,
5
Naotsugu Tsuchiya,
5,6,7
Daw-An Wu,
1
Chia-huei Tseng,
8
Shigeki Nakauchi,
2,9
and Shinsuke Shimojo
1,2
https://doi.org/10.1523/ENEURO.0133-21.2021
1
Division of Biology and Biological Engineering, California Institute of Technology, Pasadena 91125, CA,
2
The
Electronics-Inspired Interdisciplinary Research Institute (EIIRIS), Toyohashi University of Technology, Toyohashi
441-8580, Japan,
3
The University of Hong Kong, Pokfulam 999077, Hong Kong,
4
NTT Communication Science
Laboratories, NTT Corporation, Atsugi 243-0198, Japan,
5
School of Psychological Sciences and Turner Institute for
Brain and Mental Health, Monash University, Melbourne, Victoria 3800, Australia,
6
Center for Information and Neural
Networks (CiNet), National Institute of Information and Communications Technology (NICT), Suita 565-0871, Japan,
7
Advanced Telecommunications Research Computational Neuroscience Laboratories, Kyoto 619-0288, Japan,
8
Research Institute of Electrical Communication, Tohoku University, Sendai 980-8577, Japan, and
9
Department of
Computer Science and Engineering, Toyohashi University of Technology, Toyohashi 441-8580, Japan
Visual Abstract
Team flow occurs when a group functions in a high task engagement to achieve a goal, commonly seen in
performance and sports. Team flow can enable enhanced positive experiences, as compared with individ-
ual flow or regular socializing. However, the neural basis for this enhanced behavioral state remains un-
clear. Here, we identified neural correlates (NCs) of team flow in human participants using a music rhythm
task with electroencephalogram hyperscanning. Experimental manipulations held the motor task constant
while disrupting the corresponding hedonic music to interfere with the flow state or occluding the partner
’
s
positive feedback to impede team interaction. We validated these manipulations by using psychometric
ratings and an objective measure for the depth of flow experience, which uses the auditory-evoked
Significance Statement
This report presents neural evidence that teams falling into the flow state (team flow), a highly positive expe-
rience, have a unique brain state distinct from ordinary flow or social states. We established a new objective
neural measure of flow yet consistent with subjective reports. We identified neural markers of team flow at
the left middle temporal cortex (L-MTC). We showed the L-MTC had a unique causality and contributed to
information integration during team flow. Finally, we showed that team flow is an independent interbrain
state with enhanced information integration and neural synchrony. The data presented here suggest a neu-
rocognitive mechanism of team flow.
September/October 2021, 8(5) ENEURO.0133-21.2021 1
–
17
Research Article: New Research
potential (AEP) of a task-irrelevant stimulus. Spectral power analysis at both the scalp sensors and ana-
tomic source levels revealed higher
b
-
g
power specific to team flow in the left middle temporal cortex (L-
MTC). Causal interaction analysis revealed that the L-MTC is downstream in information processing and
receives information from areas encoding the flow or social states. The L-MTC significantly contributes to
integrating information. Moreover, we found that team flow enhances global interbrain integrated informa-
tion (II) and neural synchrony. We conclude that the NCs of team flow induce a distinct brain state. Our re-
sults suggest a neurocognitive mechanism to create this unique experience.
Key words:
EEG; flow; hyperscanning; in the zone; neural synchrony; teams
Introduction
Flow state, or
“
getting into the zone,
”
is a psychological
phenomenon that develops when balancing the perform-
ance with the challenge of a task and providing clear
goals and immediate feedback (
Csikszentmihalyi, 1975
;
Nakamura and Csikszentmihalyi, 2002
). The flow state is
characterized by intense task-related attention, effortless
automatic action, a strong sense of control, a reduced
sense of external and internal awareness, and a reduced
sense of time (
Nakamura and Csikszentmihalyi, 2002
).
The flow state is intrinsically rewarding and can positively
affect subsequent experiences (
Csikszentmihalyi, 1975
,
2014
;
Nakamura and Csikszentmihalyi, 2002
;
Harmat,
2016
). Because of these characteristics, flow is an in-
tensely-studied topic in sports, music, education, work,
and gaming. The flow state can develop during an individ-
ual (solo) activity or a group activity. There is a growing in-
terest in studying flow in group activities, i.e., group flow,
among several fields including psychology, sociology, or-
ganizational behavior, and business (
Sawyer, 2007
;
Walker, 2010
;
Hart and Di Blasi, 2013
;
Salanova et al.,
2014
;
Hari et al., 2015
;
Pels et al., 2018
). Team flow is a
specific case of group flow in which the group forms a
team that is characterized by a common purpose, comple-
mentary skills, clear performance goals, strong commitment,
and mutual accountability (
Katzenbach and Smith, 1993a
,
b
;
van den Hout et al., 2018
). The positive subjective experience
during team flow, as in sports teams, music ensembles,
dance squads, business teams, or video gaming teams, is
superior to everyday social interaction or experiencing indi-
vidual flow (
Sato, 1988
;
Hari et al., 2015
;
Pels et al., 2018
).
A simplistic assumption is that team flow is a simple
combination of the flow and the social states. These two
states are disparate, in other words, acting in a social
context is not necessarily sufficient to get into the flow
state, and vice versa. In prior reports, the neural mecha-
nisms underlying the individual flow state and social experi-
ence have been studied in isolation. For social information
processing, several networks have been implicated. Social
perception, empathy, mentalization, and action observation
networks may provide partially overlapping brain regions in
conjunction with the amygdala, anterior cingulate cortex
(ACC), prefrontal cortex (PFC), inferior frontal gyrus (IFG),
and the inferior and superior parietal lobule (IPL/SPL), re-
spectively (
Ongür and Price, 2000
;
Dodell-Feder et al., 2011
;
Lamm et al., 2011
;
Molenberghs et al., 2012
;
Stanley and
Adolphs, 2013
;
Yang et al., 2015
). Meanwhile, several stud-
ies of individual flow have shown increased activity in the
IFG and the IPL/SPL, and decreased activity in the PFC
(
Klasen et al., 2012
;
Ulrich et al., 2014
,
2016a
,
b
;
Harris et al.,
2017
). We cannot hypothesize that any of the aforemen-
tioned brain regions contribute to team flow since there are
concordant and discordant overlaps. Hence, we posit that
team flow is more than a combination of these two states
and may arise from a unique interaction among these brain
regions, which would reveal new neural correlates (NCs)
that create this unique team flow brain state.
Phenomenologically, the experience of team flow is sub-
jectively more intense than the individual flow state and ordi-
nary social state. However, the underlying neural mechanism
is still unclear. This study directly examines the underlying
neural activity patterns, emerging at both the intrabrain and
interbrains levels during team flow. Using an exploratory ap-
proach, we identified the intrabrain correlates in team flow
that are distinct from ordinary flow or social experiences.
Using causality analysis, integrated information (II), and neural
synchrony data, we propose a model of the neural mecha-
nisms that underlie team flow.
Received March 29, 2021; accepted September 7, 2021; First published
October 4, 2021.
The authors declare no competing financial interests.
Author contributions: M.S., M.C., and S.S. designed research; M.S., M.C.,
and D.-A.W. performed research; M.S., M.C., A.L., N.T., and S.S. analyzed
data; M.S., M.C., A.L., N.T., D.-A.W., C.-h.T., S.N., and S.S. wrote the paper.
This work was supported by the Program for Promoting the Enhancement of
Research Universities funded to Toyohashi University of Technology and
Grants-in-Aid for Scientific Research (Fostering Joint International Research
(B), Grant Number 18KK0280) (M.S. and S.N.), Sponsored Research by
Qneuro, Inc. (M.S. and S.S.), Translational Research Institute through NASA
Cooperative Agreement NNX16AO69A (M.S. and S.S.), and by the Japan
Science and Technology (JST)-CREST Grant JPMJCR14E4 (to S.S.). M.C. is
supported by the University of Hong Kong Postgraduate Scholarship Program.
C.-h.T. is supported by the University of Hong Kong General Research Fund
and the Cooperative Research Project Program of the Research Institute of
Electrical Communication, Tohoku University. N.T. is supported by Australian
Research Council Discovery Projects Grants DP180104128 and
DP180100396. A.L. is supported by an Australian Government Research
Training Program Scholarship.
Acknowledgements: We thank Dr. Charles Yokoyama (University of Tokyo,
Japan), Dr. Simone Shamay-Tsoory (University of Haifa, Israel), Dr. Katsumi
Watanabe (Waseda University, Japan), and Dr. Makio Kashino (NTT
Communications Science Laboratories, Japan) for their comments on this
manuscript. We also thank Shota Yasunaga (Pitzer College, CA), Jessica Ye
(California Institute of Technology, CA), Naomi Shroff-Mehta (Scripps College,
CA), and Salma Elnagar (University of Cambridge, UK) for help with data
collection and analysis and Wenqi Yan (Monash University, Australia) for
preliminary data analysis with integrated information.
Correspondence should be addressed to Mohammad Shehata at
mohammad.
shehata@gmail.com
.
https://doi.org/10.1523/ENEURO.0133-21.2021
Copyright © 2021 Shehata et al.
This is an open-access article distributed under the terms of the
Creative
Commons Attribution 4.0 International license
, which permits unrestricted use,
distribution and reproduction in any medium provided that the original work is
properly attributed.
Research Article: New Research 2 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
Materials and Methods
Participants
We recruited 78 participants for the screening process.
In the main EEG experiment, 15 participants (five males;
age: 18
–
35 years) attended and formed 10 pairs (three
male pairs), of which five participants (one male) were
paired twice. Written informed consent was acquired from
all participants. Human subjects were recruited at a loca-
tion which will be identified if the article is published. All
the procedures were approved by the Institutional Review
Board of California Institute of Technology.
Task
We used a commercial music rhythm game called
“
O2JAM U
”
(version 1.6.0.11, MOMO Co) played on an
iPad air (model No. MD786LL/B, system, iOS 10.3.2). The
basic structure of this game follows the most common
structure in the music rhythm genre. Two consecutive
screenshots of the game are shown in
Figure 1
A
. Visual
cues (notes) moves in lanes from the top to the bottom of
the screen where the tapping area is located. There are
two kinds of cues: short and long ones. A player
’
s task is
to tap when a short cue reaches the tapping area, and to
tap and hold for the duration a long cue is at the tapping
area. The cues are designed to give the impression of
playing a musical instrument, which produces much of
the positive experience of the game. The game displays
two types of real-time feedback on the players
’
perform-
ance. The first feedback type includes a semantic judg-
ment expression (
“
EXCELLENT,
”“
GOOD,
”
or
“
MISS
”
)
together with a numerical score presented at the center
and the top corners of the screen (Extended Data
Fig.
1-1
A
). We made the first feedback type invisible to the
participants, using a privacy screen protector, to enhance
participants
’
focus on the tapping area. The second feed-
back type is a flashing visual effect that appears at the
tapping area each time the player taps at the correct tim-
ing with the cue (Extended Data
Fig. 1-1
A
). We kept this
Figure 1.
Behavioral establishment of team flow.
A
, Diagram of the finger-tapping music rhythm game. Participants must tap when
animated cues moving from the top of the screen reach the tapping area.
B
, Manipulations: team flow is predicted when the partici-
pants are playing the unmodified song and they can see the partner
’
s positive feedback (Team Flow). The flow state is disrupted
through scrambling the music (Team Only). Team interaction is disrupted by hiding the partner
’
s positive feedback using an occlu-
sion board (Flow Only). See
Table 1
for details.
C
, Sequence of the trials, showing which song and condition per trial was assigned
during the main experiment.
D
, Trial analysis: participants were sitting still while listening to a background music during the resting
phase and played the game in the playing phase. The electroencephalogram was epoched for objective assessment of flow i.e., the
AEP analysis of the task-irrelevant beeps (orange bar) and for the NCs analysis (green bar). After each trial, participants answered
the questionnaire for the subjective assessment of flow. Extended Data
Figure 1-1
shows detailed analysis pipeline.
Research Article: New Research 3 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
feedback type visible to the participants as positive rein-
forcement (
Movie 1
). The game provides two modes of
play: a two-lane or four-lane mode, in which either two or
four lanes of moving cues are presented. We used the
two-lane mode during individual screening with the par-
ticipant responsible for both lanes. We used the four-lane
mode during the main experiment in which a pair of par-
ticipants played with each participant responsible for two
adjacent lanes.
After playing each song (trial), the game displays a per-
formance report on the screen, including a final numerical
score, the total number of cues, and the number of
missed cues. The performance report of each trial was
hidden from the participants until they finish answering
their subjective experience psychometric ratings. The
percentage of the missed cues per the total number of
cues was used as a metric for the performance of each
pair of participants.
Manipulations
To create the team flow condition, the iPad was tilted
and positioned, using a custom-made holder, equidistant
from the pair of participants. Participants were instructed
to sit on two chairs at a fixed distance, to keep their heads
on chin rests, and to minimize their body movements ex-
cept for finger movement. The iPad was connected to a
pair of stereo speakers placed horizontally and equidis-
tant from the iPad and the pair of participants. A pair of
participants played in the four-lane mode.
Previous studies controlled the flow experience by ma-
nipulating the skill-challenge level in the experimental
setup, either passively by free task performance and ret-
rograde classification of certain time periods into several
flow levels (
Klasen et al., 2012
) or by actively controlling
the task level to be either too easy (boredom), adaptive
(flow), or too hard (overload;
Ulrich et al., 2014
,
2016a
,
b
).
One of the issues with modifying the skill-challenge level
to study flow is that this changes other cognitive func-
tions, such as attention, sensory information processing,
and cognitive load necessary to perform each task, as
well as gross changes in motor behavior. Therefore, ma-
nipulating the skill-challenge level complicates the ability
to distinguish the neural mechanisms underlying team
flow interaction. To avoid this complexity, we kept the
task identical in all conditions using the same sequence of
tapping cues. We manipulated the intrinsic reward/enjoy-
ment dimension of flow by scrambling the game music
and hence disrupting the pleasant experience (
Table 1
).
To create the team only condition, participants played
the same song (i.e., an identical sequence of moving
cues) as the team flow using the same setup; however, a
reversed and shuffled version of the music was played
from the same speakers (
Table 1
). The music for each
song was reversed through an online audio editing web-
site (
https://audiotrimmer.com/online-mp3-reverser/
) and
then cut into 5-s fragments through an online audio cutter
(
http://mp3cut.net/
). We randomly shuffled the fragments
and rejoined them through an online audio joiner (
http://
audio-joiner.com/
).
To create the flow only condition, participants played
the same song (i.e., an identical sequence of moving
cues) as the team flow using the same setup; however, a
black foam board (1 cm
1.5 m
75cm) was placed be-
tween the chairs to completely block the participants
’
view of each other, and a black piece of cardboard was
placed across the iPad screen with an opening to show
the visual cues but not the tapping area (Extended Data
Fig. 1-1
A
).
Screening process
The participants were first tested using a selected song
(269 cues per the two lanes) in the two-lane mode to
exclude unexperienced participants. Participants were
qualified to complete the study if they missed no more
than 10 cues. Out of the 78 participants recruited for
the first screening test, 54 participants were qualified and
the remaining 24 participants were excluded. The 54
qualified participants were also tested using other se-
lected songs with a higher number of cues to confirm their
Movie 1.
A few seconds of game-play in the Team Flow and
Team Only conditions. The
“
beeps
”
word at the bottom right in-
dicates the timing of the task-irrelevant beep sound presenta-
tion. These words are overlaid in the video for illustration and
were not present during the experiment. The scores and other
indicators at the center and at the top right and left corners
were hidden from the participants. [
View online
]
Table 1. Comparison of the stimuli across the experimental conditions
Conditions
Team Flow
Team Only
Flow Only
Cues sequence (visual stimulus)
Self
Constant (visible to both participants)
Partner
Constant (visible to both participants)
Positive Feedback
Self
Visible (performance dependent)
Partner
Visible
Visible
Not visible
Song (auditory stimulus)
Original
Scrambled
Original
Beeps (task-irrelevant stimulus)
Constant
Constant
Constant
Research Article: New Research 4 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
skill level. Because the positive experience of this game
depends on individual preference for the music rhythm,
we prepared 11 songs (500
–
960 cues per the four lanes)
in various genres. The 54 qualified participants were
asked to rate their preferences for all the 11 songs sepa-
rately on a seven-point scale (one for
“
not like it at all,
”
seven for
“
like very much
”
). Although the duration of the
songs varied from 110 to 160 s, each song was played
only for 1 min, which was long enough to ensure that par-
ticipants had heard the main rhythm of the song. We fixed
the duration for all songs to ensure the accuracy of prefer-
ence ratings. To avoid possible influence from the experi-
menter, the experimenter maintained a neutral attitude by
avoiding eye contact with participants or a physical re-
sponse to the music. Participants were paired based on
their skill level and song preference rating.
The second step of the process was to screen paired
participants for their preference to the team set-up (team
flow) versus playing in the team with board set-up (flow
only) or a single-player set-up (solo flow) in a behavior
pilot experiment. For the solo flow condition, we used the
two-lane mode of the game. At the end of the pilot experi-
ment, we presented the following question:
“
Based on
how much you enjoyed the performance and you want to
play it again, please rank the following experiences: sin-
gle-player set-up, team set-up, the team with board set-
up.
”
We presented a scale from 1 (least preferred) to 7
(strongly preferred) in front of each set-up, and the rank-
ing was not a forced-choice one. We excluded partici-
pants who ranked the team with board set-up or single-
player set-up higher than the team set-up. Out of the 54
participants from the first screening step, 38 prosocial
participants were invited for the main experiment on an-
other day, and the remaining 24 participants were ex-
cluded. For both the screening and the main experiment,
we paired only same gender participants, and the main
pairing criteria for qualified participants were their skill
level and song preferences. As friends were encouraged
to pair up to participate in this experiment, we preferred
pairing friends into the main experiment if they satisfied
the main criteria. Participants reported their relation with
the partner by answering the question
“
Generally, how
much time do you spend with the other player?
”
Pairs
who answered
“
first time to meet
”
or
“
only meet in the last
experiment
”
were categorized as strangers; pairs who an-
swered
“
less than 5 h per week,
”“
5
–
20 h per week,
”
or
“
more than 20 h per week
”
were categorized as friends. In
total, 17 participants consisted of six pairs of strangers
and five pairs of friends.
Main experiment
After setting the EEG cap, the electrode positions were
co-registered with the T1-magnetic resonance imaging
(MRI) using the Brainsight TMS Navigation system (Rogue
Resolutions Ltd). Then the paired participants, seated
on two chairs at a fixed distance, underwent a beep-only
trial. In this trial, they were instructed to passively listen to
the task-irrelevant beep stimulus for 2.5 min, to keep their
heads on chin rests and their eyes open, and to minimize
their body movements. This trial was to check the EEG
recording quality and verify that we could obtain a clear
AEP response. The paired participants then performed a
practice trial in the flow only condition to become familiar-
ized with the procedure. Then each pair of participants
was required to play six songs each at the team flow,
team only, or flow only conditions forming 18 trials (
Fig.
1
C
,
D
). One pair played only five songs because of time
availability. The sequence of songs and conditions were
pseudorandomized (
Fig. 1
D
). To keep participants
’
con-
tinuous interest, the consecutive songs were always differ-
ent (
Fig. 1
D
). To control practice and carryover effects, we
arranged each condition to have an equal chance of being
before or after the other two conditions (
Fig. 1
D
). All trials
included a resting phase and a playing phase (
Fig. 1
C
).
During the resting phase, participants were instructed to
passively listen to the task-irrelevant beep sound and the
game background music for 30 s, to keep their heads on
chin rests and their eyes open, and to minimize their body
movements. Then, the experimenter asked participants to
click the game-play icon on the iPad to start the playing
phase.
During the playing phase, participants were instructed
to keep their heads on chin rests, to minimize their body
movements except for finger movement, and to minimize
vocal sounds that could distract their partner. The partici-
pants were allowed to give verbal comments related to
the game. The participants did not comment while playing
the game. They were only allowed to give verbal com-
ments after answering the psychometric ratings and re-
vealing the participants
’
final scores. We video recorded a
top-view of the iPad and the participants
’
hands using an
iPhone fixed
;
50cm above the iPad where all the types
of feedbacks were visible. After the playing phase of each
trial, participants were given access to private screens
and keyboards to freely answer the psychometric ratings
on the flow experience and team interaction experience.
Then, the pair were allowed to jointly view the perform-
ance report. The experimenter asked the participants
whether they wanted to proceed to the next trial or if they
needed some rest to minimize the effect of fatigue on per-
formance or EEG recording quality.
Task-irrelevant stimulus
A task-irrelevant auditory stimulus (a beep sound) was
pseudorandomly presented to probe the strength of the
participants
’
selective attention to the game and was
used as an objective measure of flow. We presented beep
trains played at 5Hz for 1 s (i.e., each train consisted of
five beeps). Each beep was at 500 Hz and lasted for 10
ms. The beep trains simulated the sound of someone
knocking on a door to make the stimulus as natural as
possible. The interval between the beep train varied from
4 to 8 s. The beeps were generated by MATLAB 2012
(The MathWorks) and delivered through another pair of
speakers placed equidistant from the iPad.
Anatomical MRI acquisition
To increase the accuracy of source estimation for corti-
cal activity, individual head anatomy from each
Research Article: New Research 5 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
participant, who passed the screening and agreed to par-
ticipate in the main experiment, was acquired with MRI. A
3 Tesla Siemens Trio scanner and standard radio fre-
quency coil was used for the entire MRI scanning. High
resolution structural images were collected using a stand-
ard MPRAGE pulse sequence, providing full brain T1-
weighted 3D structural images.
Psychometric ratings and calculation of experience
indices
From subjective reports, we calculated the flow, team,
and team flow indices, by calculating the arithmetic mean
of the ratings for each trial, to estimate subjective experi-
ence for flow state, positive team interaction, and team flow,
respectively (Extended Data
Fig. 2-1
). For assessing flow ex-
perience, we used psychometric ratings related to the skill-
demand balance (Q1 and Q2), feeling in control (Q3), auto-
maticity (Q4), enjoyment (Q5), and time perception (Q6) di-
mensions of flow (
Nakamura and Csikszentmihalyi, 2002
).
For assessing team interaction, we used psychometric rat-
ings related to awareness of partner (Q7), teamwork (Q8),
and coordination (Q9) dimensions of positive team interac-
tion. Psychometric ratings assessing competition (Q10) and
distraction (Q11) were used to confirm the absence of nega-
tive team interactions and were not included in any index.
The team flow index was calculated by averaging the flow
and team indices. In addition, we tested the effect of friend-
ship between the two players (friend or stranger) on the
subjective rating of flow. There was no significant differ-
ence between friend-pairs and stranger-pairs in flow
index, team index or team flow index (two-way repeated
measures ANOVA, main effect of relation for flow index,
F
(1,18)
=0.6853,
p
= 0.4186; for team index,
F
(1,18)
=
0.1557,
p
=0.6978; for team flow index,
F
(1,18)
=1.4992,
p
=0.2366). Therefore, we combined friend-pairs and
stranger-pairs in the following analysis.
Hyperscanning EEG recording and preprocessing
Electroencephalogram (EEG) was recorded simulta-
neously from both participants using a dual BioSemi
ActiveTwo system (BioSemi Inc.). Each participant wore
a cap holding 128 scalp Ag/AgCl electrodes. Signals
were amplified by two daisy-chained ActiveTwo AD
boxes where one AD box was connected to the control
PC and worked as a master controlling the other AD
box to ensure synchronization. Electrode impedance
waskeptbelow10k
V
. For each cap, an active common
mode sense (CMS) electrode and a passive driven right
leg (DRL) electrode positioned near the vertex served
as the ground electrodes. EEG signals were recorded at
asamplingrateof2048Hz(laterdown-sampledto
256Hz). During recording, the A1 electrode, or A2 elec-
trode in three participants served as a reference. In the
ABC layout (a Biosemi designed equiradial system),
these electrodes overlap with the Cz location of the in-
ternational 10
–
20 system. Signals were recorded and
saved using ActiView/LabView software (version 8.04,
BioSemi Inc.) installed on the control PC. Another mas-
ter PC was used to generate the task-irrelevant beep
sound and to send signals to the EEG data receiver
marking the onset of each beep train (event triggers).
The event triggers were used to align the EEG data with
the resting and the playing phases by using a real-time
projection of the top-view video recording to the control
PC. The experimenter confirmed that all the onsets of
the beep trains happened during the resting or the play-
ing phase periods.
To analyze the auditory-evoked potentiation (AEP), EEG
data were epoched
0.5
–
1 s (1.5-s total) flanking the
beep train onsets (AEP epochs). To analyze the NCs of
game play experience, EEG data were epoched 2
–
5 s (3-s
total) after the beep train onset (NC epochs;
Fig. 1
C
). EEG
data were bandpass filtered at 0.5
–
50Hz, using the
Parks-McClellan FIR filter, and re-referenced to the aver-
age of all channels. After this initial preprocessing, we did
a visual inspection for artifacts, including EMG, then per-
formed artifact-rejection using automatic independent
component analysis (ICA) rejection using the FASTER
toolbox (
Nolan et al., 2010
). Bad channels showing line
noise noted during recording sessions were rejected and
interpolated during the FASTER preprocessing.
Auditory-evoked potential (AEP) analysis
To select the channels maximally responsive to the
task-irrelevant auditory stimuli, we analyzed the AEP
epochs during the resting phase. We calculated the
event-related spectral perturbation (ERSP) and the intere-
poch coherence (IEC) using the EEGLAB toolbox (version
14.1.1;
Brunner et al., 2013
). Both ERSP and IEC showed
changes in
u
activity (3
–
7 Hz) at 100
–
350 ms postonset,
with a peak increase at 150
–
250ms postonset (Extended
Data
Fig. 2-2
A
,
B
). Topographical analysis in the
u
band
showed strong positive activity in the 14 central channels
from 200
–
260 ms postonset (Extended Data
Fig. 2-2
C
).
The frequency, time, and topographical frames of our
AEP were consistent with previous reports (
van Driel et
al., 2014
;
Stropahl et al., 2018
). For each trial, we used
IEC in the
u
band during the resting phase to select chan-
nels showing stable AEP. IEC was averaged across the 14
central channels, and channels showing IEC lower than
one standard error below the mean were excluded from
further AEP analysis for that trial. We then analyzed
u
power from
200 to 500ms flanking the beep train onsets
during the resting and the playing phase (Extended Data
Fig. 2-2
A
,
B
). AEP peak amplitude was calculated accord-
ing to the method described by a previous simulation
report showing that event-related potential measured
based on the mean amplitude surrounding the group la-
tency is the most robust against background noise
(
Clayson et al., 2013
). Therefore, we calculated the N1,
P2, and N2 peak latencies averaged across all conditions
during the resting phase (Extended Data
Fig. 2-2
D
). The
individual N1, P2, and N2 mean peak amplitudes
6
40ms
surrounding the calculated peak latencies were obtained
during the playing phase (Extended Data
Fig. 2-2
E
). This
resulted in the following time windows: N1 (110
–
150ms),
P1 (210
–
250 ms), and N2 (310
–
350ms). The amplitude
peaks at these time windows were averaged, considering
polarity [i.e., (P2-N1-N2)/3], and used as AEP (
Fig. 2
F
,
G
).
Research Article: New Research 6 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
Anatomically-defined source estimations
FreeSurfer (
Reuter et al., 2012
) was used for automatic
segmentation and reconstruction of the MRI images. MRI
images were used to compute each individualized head
model using the boundary element model (BEM) imple-
mented in OpenMEEG withinthe BrainStorm software
package (version 3.4) using the default parameters
(
Gramfort et al., 2010
;
Tadel et al., 2011
). MRI registration
with EEG electrode-positions were aligned with each par-
ticipant
’
s BEM model, and sources were computed (ver-
sion 2018) using BrainStorm for each NC epoch in the
playing phase. Maps of cortical activity density were ob-
tained across the BEM mesh using the distributed mini-
mum-norm estimate (MNE) method, with constrained
dipole orientations and no baseline noise correction. For
cortical region-based analysis, brain regions were defined
according to the anatomic parcellation of the Destrieux
atlas as implemented in FreeSurfer and available in
BrainStorm (
Destrieux et al., 2010
). The time series of
source activities from the 15,002 vertices and the aver-
aged activity of the predefined 148 regions of interest
(ROIs) were exported for further analysis.
Power spectrum analysis
The power spectral density (PSD) estimate was calcu-
lated using Welch
’
s overlapped segment averaging
estimator as implemented in the MATLAB 2016a signal
processing toolbox within the EEGLAB toolbox using de-
fault parameters (
Welch, 1967
;
Brunner et al., 2013
). The
normalized PSD was calculated for each NC epoch then
averaged within each trial yielding trial PSD data at each
of the 128 channels, the 148 brain region sources, and the
15,002 mesh vertex sources. For each song played, the indi-
vidual
’
s mean PSD across the three conditions was calcu-
lated. The normalized power was calculated by subtracting
the individual
’
smeanPSDfromthePSDateachcondition.
The normalized power was averaged within the following fre-
quency bands:
d
(1
–
3Hz),
u
(4
–
7Hz),
a
(8
12 Hz),
b
(13
–
30 Hz),
g
(31
–
120Hz), and lower
g
(31
–
50 Hz). We started
with exploratory analysis by checking the normalized power
grand-averaged across all channels for each frequency band
(Extended Data
Fig. 3-1
). We found significant differences
across conditions in the
g
(31
–
120Hz; one-way repeated
measures ANOVA,
F
(2,57)
=5.1445,
p
= 0.0105) band, and
showing a trend in the
a
(8
12Hz; one-way repeated meas-
ures ANOVA,
F
(2,57)
= 2.3661,
p
=0.1075) and
b
(13
–
30 Hz;
one-way repeated measures ANOVA,
F
(2,57)
= 2.0504,
p
=
0.1427) bands. The
d
(1
–
3Hz; one-way repeated measures
ANOVA,
F
(2,57)
=0.5378,
p
=0. 5884) and
u
(4
–
7Hz; one-
way repeated measures ANOVA,
F
(2,57)
=0.1129,
p
=0.8936)
bands were not significant. For the topographical analysis,
the normalized power for the 128-channel data and the per-
mutation statistics with Bonferroni multiple comparison
Figure 2.
Assessment of the flow state.
A
–
C
, Subjective assessment of flow: psychometric rating indices as a measure of subjective
flow (flow index;
A
), team interaction (team index;
B
), or team flow (team flow index;
C
) experiences (Extended Data
Fig. 2-1
shows
the detailed psychometric ratings for each question). Friedman test with Conover
’
s
post hoc
test; *
p
,
0.05, **
p
,
0.01, ***
p
,
0.001.
Error bars represent mean
6
SEM;
n
= 15.
D
,
E
, Objective assessment of flow.
D
, The mean AEP calculated by averaging the follow-
ing time windows: N1 (110
–
150ms), P1 (210
–
250ms), and N2 (310
–
350ms), considering polarity. The non-flow condition (Team
Only) showed statistically significant higher AEP than the flow conditions. One-way repeated measures ANOVA with Bonferroni
post
hoc
test; *
p
,
0.05. Error bars represent mean
6
SEM;
n
= 15.
E
, Spearman
’
s correlation between AEP and flow index. AEP is nega-
tively correlated with the flow index in the team flow condition (Spearman
’
s Rho =
0.48,
p
= 0.03), showing a negative correlation
trend in the flow only condition (Spearman
’
s Rho =
0.29,
p
= 0.22), and no correlation in the team only condition (Spearman
’
s
Rho= 0.11,
p
= 0.64). The lines indicate the regression lines. Shaded areas indicate a 95% confidence interval;
n
= 20. Extended
Data
Figure 2-2
shows the detailed AEP analysis.
Research Article: New Research 7 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
correction were projected to topographical maps using
EEGLAB toolbox. As detecting high-
g
power (
.
50 Hz) using
noninvasive EEG might be prone to artifacts (
Völker et al.,
2018
), we only considered the combined
b
and low-
g
(
b
-
g
)
band (13
–
50Hz) for further analysis. We used one-way re-
peated measures ANOVA across conditions for determining
the significance in each anatomic-source
b
-
g
power effect.
We set the significance threshold to
p
,
0.00034 (i.e., 0.05/
148 ROIs) to correct for multiple comparisons (Bonferroni-
corrected critical value).
Unsupervised clustering analysis
We clustered the 15,002 mesh vertex sources based on
their
b
-
g
power. We used scikit-learn, a Python machine
learning library, and implemented the unsupervised ag-
glomerative clustering approach (
Abraham et al., 2014
).
Agglomerative clustering uses a bottom-up hierarchical
approach where vertices are progressively linked together
into clusters based on their feature similarity. We used
three features for clusters which are the grand averaged
b
-
g
normalized power at each of the three conditions.
We used the Euclidean distance as a similarity measure
and the complete linkage criteria, which minimizes the
maximum distance between observations of pairs of clus-
ters. We have tried setting the number of clusters into 3
–
40 clusters. We selected the minimum number of clusters,
seven in our case, that shows trends for flow-related and
team-related clusters. When we set the number of clusters
to three up to six clusters, the anatomic resolution was not
clear. When we set the number of clusters to more than
seven clusters, the anatomic resolution was more clear,
and we obtained higher significant clusters even after mul-
tiple comparison. When we set the number of clusters to
seven clusters (cls; Extended Data
Fig. 4-2
), we detected
two cls distributed over the anterior part of the frontal cor-
tex, where the
b
-
g
power was higher in the team only con-
dition than the other conditions (Extended Data
Fig. 4-2
C
,
D
; cls 1 and 2). This pattern was significant in cl 2 (one-way
repeated measures ANOVA,
F
(2,57)
= 3.6125,
p
=0.033),
while cl 1 showed a trend (one-way repeated measures
ANOVA,
F
(2,57)
=1.5916,
p
=0.2125). The suppressed activ-
ity in these clusters is specific to the flow experience, re-
gardless of the social context, which is consistent with a
neural representation of the automaticity-dimension of flow
(
Klasen et al., 2012
;
Ulrich et al., 2014
). We also detected
two clusters distributed mostly over the middle and inferior
frontal cortex and the left occipital cortex (OC), where the
b
-
g
power was lower in the flow only condition than the
other conditions (Extended Data
Fig. 4-2
C
,
D
; cls 3 and 4).
This pattern was significant in cl 4 (one-way repeated
measures ANOVA,
F
(2,57)
= 7.4841,
p
= 0.0013), while cl
three showed a trend (one-way repeated measures
ANOVA,
F
(2,57)
=2.4288,
p
=0.0972). The increased ac-
tivity in these clusters is specific to team interactions,
regardless of the flow state. The remaining clusters
were distributed mostly over the temporal, parietal, and
occipital cortices, where the
b
-
g
power was higher
in the team flow condition than the other conditions
(Extended Data
Fig. 4-2
C
,
D
;cls5
–
7). This pattern was
significant in all three cls: cl 5 (one-way repeated measures
ANOVA,
F
(2,57)
=11.8753,
p
=0.000049), cl 6 (one-way re-
peated measures ANOVA,
F
(2,57)
=9.548,
p
=0.00027), and
cl 7 (one-way repeated measures ANOVA,
F
(2,57)
=6.9256,
p
=0.002). The increased activity in these clusters was spe-
cific to team flow.
Grouping of ROIs
First, the anatomically-defined ROIs that showed signif-
icant
b
-
g
normalized power across conditions, as shown
in Extended Data
Figure 3-2
A
–
C
, were grouped as RG7
regardless of their cluster composition. Second, for the
remaining anatomic-defined ROI, we calculated the clus-
ter composition as the percentage of the flow-related
clusters (cls 1
–
2), team-related clusters (cls 3
–
4), and
team flow-related clusters (cls 5
–
7). We checked whether
the anatomically-defined ROIs can be spatially subdi-
vided into smaller ROIs with clear tendencies for a certain
activity-dependent cluster composition (Extended Data
Fig. 4-1
A
). This check was done by calculating a cumula-
tive cluster composition curve to define a threshold for
subdividing the ROIs (Extended Data
Fig. 4-1
B
). We pre-
sented the superior frontal cortex as an example of the
subdivided ROIs (Extended Data
Fig. 4-1
A
,
B
). Finally, we
grouped anatomically-defined ROIs or their subdivisions
into six regions (RGs) per hemisphere based on the major
activity-dependent cluster composition (Extended Data
Fig. 4-1
C
). Therefore, the total number of RGs was 14
RGs (seven RGs per hemisphere). For each of the 14 RGs,
the activity-dependent cluster composition is summarized
in Extended Data
Figure 4-3
and the anatomic composi-
tion is summarized in Extended Data
Figure 4-4
. The time
series from all the 15,002 vertices were averaged based
on the new 14 RGs and hence reduced into 14 time series
for each trial per participant.
Intrabrain causal interactions analysis
We used the Source Information Flow Toolbox (SIFT) to
fit an adaptive multivariate autoregressive (AMVAR)
model for the 14 RGs activities for each subject
’
s trial
using the Vieira
–
Morf algorithm (
Delorme et al., 2011
). We
fitted the NC epoch with a sliding window length of
500ms and a step size of 25 ms (
Wang et al., 2014
).
Model order was selected by minimizing the Akaike
Information criterion. We validated each fitted model
using tests included in SIFT for consistency, stability, and
whiteness of residuals. To estimate causal interactions,
we used three directed model-based linear frequency-do-
main Granger-causality (GC) measures (
Wang et al.,
2014
). These measures are the normalized partial directed
coherence (nPDC;
Baccalá and Sameshima, 2001
), the
direct directed transfer function (dDTF;
Korzeniewska et
al., 2003
), and the Granger
–
Geweke causality (GGC;
Geweke, 1982
;
Bressler et al., 2007
). For each connectiv-
ity measure, we averaged across trials for each partici-
pant per condition, then averaged across the NC epoch
time interval (3 s) and across the
b
-
g
(13
–
50Hz) fre-
quency. Finally, to quantify the degree by which an RG
sends or receives information, we calculated the ratio of
sending (to) divided by receiving (from) for each RG-RG
Research Article: New Research 8 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
interaction and then average these ratios for each RG per
condition per participant (to/from ratio). A two-way re-
peated measures ANOVA was used as statistical test. To
calculate the information senders for RG-RG causal inter-
actions, we used the Log to/from GGC ratio for each RG-
RG connection. Top information senders were calculated
by setting a threshold with a
p
value of 0.064. The RG-RG
connections above this threshold were represented on a
circular graph.
Integrated information analysis
Integrated information (II) was used as a measure of
inter-RG bidirectional causal interaction. For every pair of
time courses of the RGs activities, within and between
participants, we operationalized the
“
state
”
of the pair of
RGs by discretizing time-samples into binary values. To
roughly match the frequency range of 13
–
50Hz, we first
down-sampled the RGs activities to give timesteps of
12.8, 17.1, 25.6Hz or 51.2Hz (that is, a time step of 19.5,
39.1, 58.6, or 78.1ms). Using the down-sampled RGs ac-
tivities, we then converted each pair of consecutive time
samples to
“
on
”
if the RGs activity
’
s voltage was increas-
ing over two time steps and
“
off
”
otherwise. Using the
time series of binarized states, we computed the probabil-
ities of each state transitioning into each other state, con-
structing a transition probability matrix (TPM) which
describes the evolution of the pair of RGs activities across
time. To ensure accuracy of transition probabilities, we
computed these across all trials. As lower time resolutions
give fewer observations with which to compute the proba-
bilities, we repeated the down-sampling for each possible
“
start
”
(i.e., for each time-sample in the first time-bin) and
used all transitions from all shifted-down-sampled time
series to build the TPM. We then submitted the TPM to
PyPhi (1.2.0;
Mayner et al., 2018
), which then constructs a
minimally reducible version of the TPM, assuming inde-
pendence of RGs activities, and compares the original
TPM to the minimally reducible version to compute II
(
Oizumi et al., 2014
;
Mayner et al., 2018
). For each actual
pair, we calculated the normalized II value by subtracting
the absolute value from the average across all conditions
for each RG-RG connection. A three-way repeated meas-
ures ANOVA (condition
RG1
RG2) was used as a sta-
tistical test for normalized II at each RG-RG connection.
The global normalized II was calculated through averaging
normalized II values across all possible RG-RG connec-
tions. A one-way repeated measures ANOVA was used as
a statistical test for global normalized II.
Phase synchrony analysis
The phase-locking value (PLV), or intersite phase clus-
tering (ISPC), was used as an index of neural synchrony.
The distribution of the phase angle differences between
sources was generated at each time point (within the NC
epoch 3-s window) then averaged over (ISPC-trial;
Lachaux et al., 1999
;
Cohen, 2014
). ISPC-trial was calcu-
lated at each frequency and then averaged across the
frequency band of 13
–
50Hz. For each condition, we cal-
culated the ISPC-trial between all sources for the actual
pairs or for each of 10 randomly-assigned pairs. For each
actual or random pair, we calculated the normalized PLV
value by subtracting the PLV value from the average
across all conditions for each RG-RG connection. The
global normalized PLV was calculated through averaging
normalized PLV values across all possible RG-RG con-
nections. A two-way repeated measures ANOVA was
used as a statistical test.
Statistical analysis
All statistics were done using the Statistics and Machine
Learning Toolbox within MATLAB 2016a and JASP (Version
0.14.1). We compared non-overlapping dependent correla-
tions, as described in the article (
https://garstats.wordpress.
com/2017/03/01/comp2dcorr/
), using the Robust Correlation
Toolbox in MATLAB (
http://sourceforge.net/projects/
robustcorrtool/
) which was validated for Spearman
’
s
correlation (
Wilcox, 2016
). In this section, we give a
parameter justification for each analysis based on the
rationale for doing the analysis.
Screening process
The screening process is necessary in this study to at-
tain reasonable team flow behavioral response. In the
first screening process, we needed to assure that partici-
pants who signed up for this study have enough skill to
fall into the flow state. In the second screening process,
we needed to match participants based on their skill and
song preference. We assumed that this screening would
maximize the chances of finding pairs of participants
who can reach the team flow state.
Sample size
The final number of participants was mainly constrained
by availability after the screening process. We tried to
kept the final number of participations similar to the sam-
ple sizes reported in similar publications (
Yun et al., 2012
).
Note, during the main experiment, the data collection pro-
cess for one male pair of participants was interrupted be-
cause of a technical error, and the collected data were
excluded from data analysis.
Trial numbers
We limited the number of trials to six per condition to
avoid fatigue which might have compromised the possi-
bility of falling into the flow state in later trials. For one
pair, we could only collect five trials per condition be-
cause of time constraints. For another pair, one of the
trials contained excessive noise, and hence, we ex-
cluded this trial and all corresponding trials in the other
conditions.
Units of analysis
Unless otherwise described, the unit of analysis is par-
ticipant, i.e.,
n
= 15. For the five participants invited twice,
we averaged the results from the two experiments giving
one data point. In some analyses, the unit of analysis was
participation, i.e.,
n
= 20. For the performance analysis,
the unit of analysis was the final score for the pair, i.e.,
n
=10. Data collection was not performed blind to the
conditions of the experiment. Experimental blinding
Research Article: New Research 9 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
was not possible because of the overt and obvious na-
ture of the experimental setup for each manipulation.
Data in all conditions were subjected to identical analy-
sis algorithms.
Data availability
Analysis codes used in the preparation of this article are
available at
https://osf.io/3b4hp
.
Results
Behavioral paradigm for team flow
We designed a behavioral paradigm to assess team
flow, in which a pair of participants played a popular
music rhythm game. The game
’
s task required responses
by tapping a touch screen when animated visual cues
reached a designated area and delivered instantaneous
positive feedback. The game created the impression of
playing a musical instrument, which increases the likeli-
hood of entering a flow state. Each pair of participants
played as a team by splitting the tapping area and sharing
in task completion with the common goal of obtaining the
best score for the team. We simultaneously recorded their
brain activities using electroencephalogram (EEG;
Fig.
1
A
; Extended Data
Fig. 1-1
;
Movie 1
). Participants were
screened to select prosocial highly-skilled participants in
this game and were matched according to their skill level
and song preference (for more details, see Materials and
Methods).
In the primary experimental condition, the team flow
condition, teams played the unmodified songs in an open
interpersonal setting to maximize the team flow experi-
ence (
Fig. 1
B
, left panel). To fulfill the team characteristics:
(1) common purpose: we instructed each pair of partici-
pants (team) to get the highest score for the team; (2)
complimentary skills: we matched participants based on
skill and song preference; (3) clear performance goals: we
provided the performance feedback at the end of each
trial; (4) keep a strong commitment: we allowed for the
visibility of teammate
’
s instant feedback; and (5) mutual
accountability: we explained that a decrease in perform-
ance from any teammate would affect the total score. We
designed two control conditions to manipulate either the
flow or the social states. To disrupt the flow state, we
manipulated the team only condition by modulating the
intrinsic reward/enjoyment dimension for flow by scram-
bling the game
’
s music. This procedure then interrupted
the sense of immersion and the continuity of the game
(
Fig. 1
B
, middle panel). To disrupt the social state in the
flow only condition, we used an occlusion board be-
tween the two participants that occluded the partner
’
s
positive feedback and bodies while leaving all of the cues
visible to both players (
Fig. 1
B
, right panel; Extended Data
Fig. 1-1
A
). We designed the manipulations to disrupt one
component of team flow per condition to ensure that any
discovered NC for team flow did not arise from only one of
these components.
To control for stimuli-related neural activities, we kept
all stimuli constant across conditions by asking the teams
to play the same song at the three conditions (
Fig. 1
C
).
For each song, the visual stimulus (cue sequence), the
total auditory stimulus presented, task difficulty, and the
sequence of the task-irrelevant beeps were kept constant
across conditions (
Table 1
). Then we normalized the neu-
ral signals per song. The only remaining variables across
conditions were the song
’
s pleasance and the visibility
of the partner
’
s positive feedback. There were no differ-
ences in the participants
’
performances across condi-
tions (repeated measures one-way repeated measures
ANOVA,
F
(2,27)
= 0.02437,
p
= 0.976), which ensure no
differences in gross motor responses.
Subjective assessment of team flow
To validate our manipulations, participants performed
psychometric ratings after each trial (
Fig. 1
D
; Extended
Data
Fig. 1-1
B
). To assess the dimensions of the flow
state, we presented the participants with the following
psychometric ratings: (1)
“
I had the necessary skill to play
this trial successfully
”
; (2)
“
I will enjoy this trial more if it
has less/more notes
”
; (3)
“
I felt in control while playing this
trial
”
; (4)
“
I made correct movements automatically with-
out thinking
”
; (5)
“
I love the feeling of this trial and want to
play it again
”
; and (6)
“
How time flies during this trial.
”
To
assess positive social interaction for teams, we pre-
sented the following: (7)
“
I was aware of the other play-
er
’
sactions
”
;(8)
“
I felt like I was playing with the other
person as a team
”
;and(9)
“
I was coordinating my fingers
with the other player
’
sfingers
”
(Extended Data
Fig. 2-1
).
Responses were collected on a seven-point Likert scale
and averaged into a flow index by averaging responses
across (1) to (6), a team index by averaging across (7) to
(9), and a team flow index by averaging across (1) to (9).
As expected, the flow index decreased significantly in
the team only condition than the other two conditions
(Friedman test, non-parametric repeated measures ANOVA,
x
2
=20.133,
p
,
0.001,
n
=15;
Fig. 2
A
). The team index de-
creased significantly in the flow only condition than the other
two conditions (Friedman test,
x
2
=20.373,
p
,
0.001,
n
=15;
Fig. 2
B
). The team flow index was significantly higher
in the team flow condition more than the other two condi-
tions (Friedman test,
x
2
=22.933,
p
,
0.001,
n
=15;
Fig. 2
C
).
The results of the psychometric assessment confirmed ef-
fectiveness of our manipulations to achieve the desired sub-
jective experience for each condition.
Objective assessment for the depth the flow state
To provide objective evidence for the flow state, we de-
veloped a novel neurophysiological measure of flow. We
used the intense task-related attention and the reduced
sense of external awareness dimensions of flow (
Nakamura
and Csikszentmihalyi, 2002
), and the well-known effect of
selective attention on the AEP (
Picton and Hillyard, 1974
).
During each trial, we presented task-irrelevant beeps to the
participants (
Fig. 1
D
; Extended Data
Fig. 1-1
B
). The more
the participants were immersed in the game, the weaker the
strength of the AEP in response to the task-irrelevant beeps.
Thus, this AEP constitutes an objective measure for flow
(
Fig. 2
D
,
E
; Extended Data
Fig. 2-2
). The mean AEP re-
sponse was significantly higher in the team only (mean=
Research Article: New Research 10 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
0.29, 95% confidence interval (CI) [0.22, 0.35]) condition
more than the other two conditions (team flow mean: 0.19,
95%CI [0.13, 0.25]; flow only mean: 0.17, 95%CI [0.11,
0.24]; one-way repeated measures ANOVA,
F
(2,42)
=6.149,
p
=0.006,
h
2
= 0.305;
Fig. 2
D
). The higher AEP in the team
only condition indicated that the participants where not fully
engaged, and hence their brains responded more to the
task-irrelevant beep sound. Notably, the AEP was negatively
correlated with the flow index in the team flow condition
(Spearman
’
sRho=
0.48 [
0.76,
0.03],
p
=0.03), while it
wasonlyweakly(Spearman
’
sRho=
0.29 [
0.66, 0.14],
p
=0.22) or not correlated (Spearman
’
sRho=0.11[
0.35,
0.55],
p
= 0.64) with the flow index in the flow only and the
team only condition, respectively (
Fig. 2
E
).Thenegativecor-
relation of AEP with the flow index was significantly stronger
in the team flow condition than the team only condition
(Spearman
’
s Rho difference =0.59 [
1.05,
0.05],
p
=0.04).
These results indicate that the experimental manipulations
did produce a deeper flow state in the team flow and flow
only conditions than the team only condition.
b
-
c
Power at the middle temporal cortex (MTC) as a
neural signature for team flow
To detect specific NCs for team flow, we used power
spectral analysis at various domains (Extended Data
Fig.
1-1
B
). In the frequency domain, we started with explora-
tory analysis by checking the normalized power grand-
averaged across all channels for each frequency band.
We found significant differences across conditions in the
a
(8
–
12 Hz),
b
(13
–
30Hz), and
g
(31
–
120 Hz) bands (for
details, see Materials and Methods, Power spectrum
analysis; Extended Data
Fig. 3-1
). At the topographical
domain level,
a
power analysis did not show specific sur-
face channels significantly different across conditions
(data not shown). Topographical
b
-power and
g
-power
analysis showed four channels at the left temporal
area with significantly higher
b
and
g
power in the team
flow condition, more than the other two conditions
(
Fig. 3
A
,
B
). The power spectral analysis, averaged from
these four channels, showed a clear higher normalized
b
-power and
g
-power in the team flow condition than the
other conditions (
Fig. 3
C
). As there are some limitations in the
capability of EEG to accurately detect high-
g
(
.
0Hz)power,
we used the combined
b
and low-
g
(
b
-
g
)band(13
–
50 Hz)
for further analysis. The
b
-
g
band showed significantly high-
er normalized power in the team flow condition, more than
the other two conditions (
Fig. 3
D
, one-way repeated
measures ANOVA,
F
(2,42)
=6.335,
p
=0.005,
h
2
= 0.312;
team flow mean: 0.77, 95%CI [0.32, 1.23]; team only
mean:
–
0.27, 95%CI [
0.72, 0.19]; flow only mean:
–
0.51, 95%CI [
0.96,
0.06]).
At the anatomic-source domain level, we performed a
cortical source localization method, using co-registration
with the individual
’
s structural MRI. The brain was seg-
mented into 148 ROIs based on the Destrieux brain atlas
(
Destrieux et al., 2010
). The anatomic-source
b
-
g
power
analysis, after multiple comparison correction, showed 16
ROIs in the left and right temporal areas with a signifi-
cantly higher
b
-
g
power in the team flow condition com-
pared with the other two conditions (Extended Data
Fig.
3-2
A
,
B
). As a representative example, the normalized
b
-
g
power for the left middle temporal gyrus (L-MTG) is
shown in
Figure 3
F
(one-way repeated measures ANOVA,
F
(2,42)
= 6.744,
p
= 0.004,
h
2
= 0.325; team flow mean:
0.76, 95%CI [0.32, 1.2]; team only mean:
–
0.2, 95%CI
[
0.63, 0.24]; flow only mean:
–
0.56, 95%CI [
1.0,
0.12]). Also, the
b
-
g
power of these brain regions
showed higher correlation tendencies with the team
flow index only in the team flow condition. The L-MTG
showed the highest
b
-
g
power correlation with the team
flow index in the team flow condition (Spearman
’
s
Rho =0.59 [0.21, 0.84],
p
=0.006) but not in the team
only (Spearman
’
s Rho =
0.19 [
0.67, 0.34],
p
=0.43) or
the flow only (Spearman
’
sRho=
0.02 [
0.46, 0.42],
p
=0.95) conditions (
Fig. 3
G
). The positive correlation of
the
b
-
g
power with the team flow index was significantly
higher in the team flow condition than the team only con-
dition (Spearman
’
s Rho difference=0.78 [
0.01 1.40],
p
= 0.05) and the flow only condition (Spearman
’
sRho
difference =0.61 [0.00, 1.16],
p
=0.05).
We note that some ROIs showed a trend unique to the
team only or the flow only conditions, but they did not sur-
vive after the multiple comparison correction. Since the
anatomic-source localization averages source vertices
based on a predefined parcellations method, we devel-
oped a method to give more weight to the distribution of
activity rather than anatomy. We used unsupervised ma-
chine learning to cluster (cl) the source vertices based on
their similarity in the
b
-
g
power pattern (Extended Data
Fig. 3-2
C
,
D
). Using the unsupervised clustering analysis,
we detected cls specific to team flow, team only and flow
only conditions (for details, see Materials and Methods,
Unsupervised clustering analysis). These results indicate
that even during team flow, the brain shows neural activ-
ities related to each isolated experience: the flow and the
social states.
Theresultsfromthepowerspectralanalysesatevery
tested domain provided the first neural evidence that the
team flow experience is a qualitatively different brain state
distinguishable from the flow or social states. In other words,
the team flow state does not result from a simple combination
of the flow and the social states, but it has its own neural sig-
nature, which we posit accounts for the superiority in the sub-
jective experience. Next, we checked for possible unique
interactions between these brain regions during team flow.
Before performing further analyses, we grouped the 148
ROIs into 14 brain regions, seven per hemisphere, using a
combination of the standard anatomic definition and the
functional activity revealed through the cluster analysis
(Extended Data
Figs. 3-2
C
,
D
,
4-1
,
4-3
,and
4-4
). These 14
brain regions (RGs) are: the PFC (RG1), the ACC (RG2), the
inferior frontal cortex (IFC, RG3), the superior temporal cortex
(STC; RG4), the central and parietal cortex (CPC; RG5), the
OC (RG6), and the MTC (RG7). The MTC included all the
ROIs that showed a significant effect on team flow (Extended
Data
Fig. 3-2
B
), regardless of the cluster composition.
The left MTC (L-MTC) receives and integrates information
from brain areas encoding flow or social states
We tested whether the neural signature of team
flow detected in the MTC upstream or downstream in
Research Article: New Research 11 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
information processing. We analyzed the causal infor-
mation interactions across all the brain regions (RGs),
using three frequency-domain GC measures: the GGC,
the dDTF, and the nPDC (
Wang et al., 2014
). In all GC
measures, the causal interaction matrix showed that
MTC receives information (from) more than sending in-
formation to (to) other RGs (Extended Data
Fig. 4-2
A
).
Also, we quantified the global to/from ratio for each
Figure 3.
Higher
b
-
g
power at the L-MTC revealed as a unique neural signature for team flow.
A
, The topographies of the
b
and
g
frequencies (13
–
120Hz) computed as the average over normalized power.
B
, Permutation statistical significance across conditions
with Bonferroni multiple comparison corrections. The black crosses indicate channels with
p
,
0.05.
C
, The normalized power spec-
tral analysis averaged from the four channels in the left temporal area identified in
B
. Shaded area represent mean
6
SEM;
n
= 20.
Extended Data
Figure 3-1
shows the power difference spectral analysis grand averaged across all the 128 channels.
D
, Averaged
normalized power for the
b
-
g
(13
–
50 Hz) frequency band showing power enhancement in the team flow condition. One-way re-
peated measures ANOVA with Bonferroni
post hoc
test; *
p
,
0.05. Error bars represent mean
6
SEM;
n
= 15.
E
, The brain regions
(highlighted in green), as defined by the Destrieux atlas and showing significant
b
-
g
normalized power difference across conditions.
Extended Data
Figure 3-2
shows the average normalized
b
-
g
power for each significant region.
F
, The average normalized
b
-
g
power at the L-MTG. One-way repeated measures ANOVA with Bonferroni
post hoc
test; **
p
,
0.01. Error bars represent mean
6
SEM;
n
= 15.
G
, Condition-specific Spearman
’
s correlations between
b
-
g
power and team flow index at L-MTG as a representative
region. Positive correlation was found in the team flow condition (Spearman
’
s Rho= 0.56,
p
= 0.006), but not in the team only condi-
tion (Spearman
’
s Rho =
0.19,
p
= 0.43) or in the flow only condition (Spearman
’
s Rho =
0.02,
p
= 0.95). The lines indicate the re-
gression lines. Shaded areas indicate a 95% confidence interval;
n
= 20.
Research Article: New Research 12 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
RG per condition. In all GC measure, global to/from
ratiofortheL-MTCwassignificantlylessthanany
other RG except for the right MTC (R-MTC; Extended
Data
Fig. 4-2
B
; for GGC, two-way repeated measures
ANOVA,
F
(26,494)
= 2.9768,
p
= 0). Hence, the detected
b
-
g
power in L-MTC is a downstream in information
processing during the team flow experience. We then
checked the most important upstream brain regions
that sent information to L-MTC. For each RG-RG
causal interaction, we applied a global threshold to
leave only the top (
;
10%) information senders (
Fig.
4
A
). Only in the team flow condition, the top informa-
tion senders to L-MTC include the contralateral R-PFC
and R-IFC. The team only and flow only conditions
showed a similar causality pattern, yet different from
the team flow condition, in which the top information
senders to L-MTC include the contralateral R-CPC.
Interestingly, the differences between the conditions
were observed in the interhemispheric connectivity
rather than the intrahemispheric one.
Figure 4.
Causality and II analyses during team flow.
A
, Causality analysis showing the top information senders among all RG-RG
causal interactions. For each RG-RG connection, the line color matches the color of the RG name which sends the information.
Notably, only in the team flow condition, L-MTC receives information from R-PFC and R-IFC. Extended Data
Figures 4-1
,
4-3
,
4-4
show the method for grouping of ROIs. Extended Data
Figure 4-2
shows detailed causality analysis.
B
, The mean normalized II
value (Norm II) connectivity matrix for the brain regions (RG1
–
RG7). Normalized II is calculated by subtracting the mean per condi-
tion from the average II across conditions for each RG-RG connection across conditions.
C
, The mean global Norm II averaged
across all RG-RG connections showing significantly higher interbrain (left panel) and intrabrain (right panel) mean during team flow
condition. One-way repeated measures ANOVA with Bonferroni
post hoc
test; **
p
,
0.01. Error bars represent mean
6
SEM;
n
= 15.
D
, RG-RG connections that shows significant (
p
,
0.05) Norm II in the team flow condition compared with other conditions. Three-
way repeated measures ANOVA with Bonferroni
post hoc
test. Black lines indicate intrabrain and green line indicates interbrain RG-
RG connections. D-L, dorsal-left.
Research Article: New Research 13 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org
The results above indicate that the
b
-
g
power detected
in the L-MTC during team flow might arise from informa-
tion processing that happened earlier in time, and the
sources of this information include brain areas that enco-
des flow state (namely, PFC) and social state (namely,
IFC). We suspected that a downstream brain region plays
a role in integrating information from the brain regions en-
coding each isolated experience. To test this hypothesis,
we used the II theory (
Tononi, 2004
;
Oizumi et al., 2014
).
We calculated the normalized II value (Norm II) as a metric
for the II. In both intrabrain and interbrain calculations,
there was a general tendency for Norm II to be higher in
the team flow condition than the other conditions (
Fig.
4
B
). When we averaged the Norm II across all the RG-RG
connections (global Norm II) from the left hemisphere, the
team flow condition showed significant higher interbrain
global Norm II than other conditions (one-way repeated
measures ANOVA,
F
(2,42)
=9.310,
p
,
0.001,
h
2
= 0.399;
team flow mean: 0.56, 95%CI [0.3, 0.83]; team only mean:
–
0.33, 95%CI [
–
0.6,
–
0.06]; flow only mean:
–
0.23, 95%CI
[
–
0.5, 0.04]), while showing a similar trend at the intrabrain
level (one-way repeated measures ANOVA,
F
(2,42)
=2.496,
p
= 0.101,
h
2
= 0.151;
Fig. 4
C
).
Next, we checked for the specific RG-RG connections
that showed a significant Nom II at the team flow condi-
tion compared with other conditions using the three-way
repeated measures ANOVA with Bonferroni
’
s multiple
comparison correction (condition
RG
RG interaction
for intrabrain:
F
(26,10133)
= 4.7622,
p
=0, and for interbrain:
F
(26,10959)
= 3.676,
p
= 0). Among all RG-RG connections,
we detected significant connections only at the left hemi-
sphere. These connections formed an intrabrain L-IFC-
STC-CPC-MTC subnetwork and an interbrain L-MTC-to-
L-MTC link that showed significantly higher Norm II in the
team flow than the other conditions (
Fig. 4
D
). These re-
sults indicate that during team flow, the team members
exhibited higher information integration not only within
each player
’
s brain, but also between their brains. More
specifically, L-MTC was the only brain region that showed
significantly higher interbrain II during team flow. These
results indicates that L-MTC plays a critical function in in-
formation integration during the team flow state.
Team flow is associated with higher interbrain neural
synchrony
Enhanced interbrain II might concur with enhanced
neural synchrony between the team
’
s brain regions. To
test for this hypothesis, we calculated the interbrains nor-
malized PLV (Norm PLV) across all the RG-RG connections
for each condition (
Fig. 5
A
). The results showed a general
tendency for Norm PLV to be higher in the team flow condi-
tion than other conditions. The interbrain Norm PLV calcu-
lated using a randomly shuffled pairs did not show any
difference across conditions (
Fig. 5
A
). To quantify this effect,
we averaged the Norm PLV for all RG-RG connections (global
Norm PLV) in the left hemisphere. The team flow showed a
significantly higher global Norm PLV than other conditions
only in the actual paired participants but not in randomized
pairs (
Fig. 5
B
; two-way repeated measures ANOVA, condi-
tion
randomness interaction
F
(2,84)
= 3.317,
p
=0.05,
h
2
=
0.092; condition effect
p
=0.015,
h
2
= 0.135, team flow
mean: 0.002, 95%CI [0.0007, 0.003]; team only mean:
–
0.001, 95%CI [
–
0.002, 0.0003]; flow only mean:
–
0.001, 95%
CI [
–
0.002, 0.0003]). Collectively, these results indicate that
during team flow, the team members exhibited higher inte-
gration and neural synchrony between their brains. This en-
hancement in information integration and neural synchrony is
consistent with the phenomenological experience during
team flow, and it might be the neurocognitive basis for the
superior subjective experience of team flow.
Discussion
In summary, we established a new objective neural
measure of flow, consistent with subjective reports. We
Figure 5.
PLVs show enhanced interbrain synchrony during team flow.
A
, The mean PLV connectivity matrix for the brain regions
(RG1
–
RG7). Normalized PLV is calculated by subtracting the mean per condition from the average PLV across conditions for each
RG-RG connection across conditions.
“
Paired
”
indicates the actual experimental pair,
“
random
”
indicates randomly selected pairs.
B
, The mean global normalized PLV averaged across all RG-RG connections showing significantly higher interbrains during the
team flow condition. Two-way repeated measures ANOVA for interbrain comparison with Bonferroni
post hoc
test; *
p
,
0.05. Error
bars represent mean
6
SEM;
n
= 15. ns, not significant.
Research Article: New Research 14 of 17
September/October 2021, 8(5) ENEURO.0133-21.2021
eNeuro.org