Flow and Optimal Difficulty in the Portable EEG: On the Potentiality

of using Personalized Frequency Ranges for State Detection

Michael T. Knierim, Mario Nadj and Christof Weinhardt

Institute of Information Systems and Marketing, Karlsruhe Institute of Technology, Fritz-Erler-Str. 23, Karlsruhe, Germany

Keywords: Flow, Optimal Difficulty, Workload, Attention, Frequency Separation, Portable EEG.

Abstract: The experience of flow has been centrally linked to peak task performances and heightened well-being. To

more effectively elicit these outcomes, flow is increasingly studied using neurophysiological measures. For

example, portable EEG is employed to enable automatic state detection required for adaptive system design.

However, so far, there is a lack of highly diagnostic findings, and moderately diagnostic ones relate more

strongly to a central flow pre-condition – namely optimal task difficulties. Unfortunately, even these metrics

might be infeasible in real-world scenarios and for portable EEG systems without midline electrodes. In this

work, we discuss how frequency band personalization and separation could provide options to overcome these

problems. Results from an experiment with a task manipulated in difficulty highlight that upper Alpha and

Beta ranges show differentiating patterns to their lower frequency counterparts (i.e. within bands). These sub-

bands could be used to detect instances of higher flow and optimized difficulty using portable EEG.

1 INTRODUCTION

Flow, the experience of effortless attention, peak task

performance and heightened well-being is deemed an

individually beneficial experience and also a

desirable state from an organizational perspective

(Ceja and Navarro, 2012). As the requirements for

flow are complex, flow facilitation at work is still a

central challenge (Ceja and Navarro, 2012). One

promising avenue to study flow and to advance the

development of flow-adaptive support systems is to

use electroencephalographical (EEG) measures

(Cheron, 2016; Harris et al., 2017). EEG provides

comparatively low cost, high portability (e.g. wireless

EEG headsets) and high temporal resolution

(Blankertz et al., 2016). Nonetheless, empiric EEG

results regarding flow have so far been short and have

shown conflicting results. For example, opposing

patterns of frontal Alpha activity with increased flow

have been reported (e.g. Berta et al., 2013; Léger et

al., 2014; Ewing et al., 2016; Katahira et al., 2018).

Overall, so far, no robust neural marker of flow has

been identified, despite shared conceptions that a

distinctive experience like flow ought to have some

representative underlying neural configuration

(Cheron, 2016; Harris et al., 2017). One of the closest,

yet not sufficient approaches to identify intensified

flow comes from the detection of situations with

moderate task demands. This is derived from research

on mental workload and the reverse inference that a

task with too low or too high demands is unlikely to

elicit flow (Csikszentmihalyi, 1996). The diagnostic

potential is given by observations of frontal midline

Theta increases with task demands, or by posterior

Alpha reductions with increasing task demands

(Borghini et al., 2014). Despite these advances, the

transferability of laboratory findings – particularly for

flow – is still limited, in particular because some of

the aforementioned features show limitations through

confounds with prolonged task exposure (e.g. Theta

power increases with task duration), natural behavior

(e.g. posterior Alpha blocking through visual

stimulation), and topographical localization (e.g.

Theta changes are strongest over midline positions).

The latter is important, as portable EEG is often using

few electrodes in non-uniformly distributed positions.

To extend the work on unobtrusive, automated

flow detection through portable EEG devices, we

propose that through a refined frequency separation

approach, (1) refined empiric contributions can be

made to the research on flow neurophysiology, (2)

new avenues to observe the concept of neural

efficiency during flow are opened, and (3)

alternatives to the prominent Theta and Alpha

markers for mental workload can be derived.

Knierim, M., Nadj, M. and Weinhardt, C.

Flow and Optimal Difﬁculty in the Portable EEG: On the Potentiality of using Personalized Frequency Ranges for State Detection.

DOI: 10.5220/0008362601830190

In Proceedings of the 3rd International Conference on Computer-Human Interaction Research and Applications (CHIRA 2019), pages 183-190

ISBN: 978-989-758-376-6

183

2 THEORETIC BACKGROUND

Flow research has found the state to occur remarkably

similar across numerous contexts like arts, gaming,

work, or sports (Csikszentmihalyi, 1996; Moller et

al., 2010). Flow theory describes the experience along

nine dimensions, that are classified in order of

occurrence (cf. Csikszentmihalyi, 1996 – Figure 1).

Figure 1: Flow Theory.

Among the antecedents, the optimal balance

between perceived difficulty and an individuals’

skills has played a major role in explaining how

experiences might range from boredom in very easy

tasks to anxiety in very hard tasks (Csikszentmihalyi,

1996). Due to its central place in flow theory, the

manipulation of a task’s difficulty, has been primarily

employed for the experimental flow elicitation

(Moller et al., 2010), as it has been in workload

contexts, albeit not with a focus on optimal

difficulties. Markers from workload research have

already been employed in flow research and have for

example led to conclusions of increased frontal Theta

levels in flow (Katahira et al., 2018), potentially even

with a maximum in higher flow (Ewing et al., 2016).

A second stream in neurophysiology research has

focused on elucidating whether or not flow is

represented by neural activity reductions. In one

theoretic instance, it has been proposed that during

flow, neural activity in frontal brain regions might be

downregulated to shift explicit demand processing in

frontal regions to implicit and automated processing

of learned behaviours in posterior regions (Transient

Hypofrontality – Dietrich, 2004). While this approach

has been criticized to be overly simplistic (Harris et

al., 2017), extensions of a reductionist understanding

have proposed that still, some brain regions would

show reduced activation during flow, as only areas

crucially necessary to meet the task demands would

be activated. Such an optimization could make a way

to experience highly automated and error-free task

processing (Neural Efficiency Hypothesis – (Cheron,

2016; Harris et al., 2017). Following both

propositions, in the EEG Alpha band activity, as the

prototypical inhibitory rhythm during wakefulness,

could be the measure of choice to identify neural

configurations during flow (Cheron, 2016). In

relation to this, while some studies find increased

Alpha power with higher flow self-reports (Léger et

al., 2014), within the difficulty manipulation (DM)

group comparison studies, results point more to

Alpha activity decreases with increasing task

difficulty (Ewing et al., 2016; Katahira et al., 2018

report the inverse relationship, but use amplitudes as

unit of analysis). This highlights that there is still

much to uncover to explain Alpha patterns in flow.

In this article it is argued, that the refined

specification of frequency band ranges might explain

some of the previous differences and could open new

avenues with additional explanatory potential and

higher robustness of findings. Often in EEG research

and in flow EEG research in particular, frequency

band ranges are extracted using generalized, broad

ranges (e.g. Theta 4-7.5 Hz or Alpha 7.5-12.5 Hz),

despite evidence, that such generalized ranges can

mask frequency specific changes (Klimesch, 1999).

Importantly, evidence from laboratory experiments

has highlighted that Alpha band components can

show different and even sometimes opposing patterns

(Klimesch, 1999). For example, by segmentation of

personalized Alpha bands into three 2 Hz wide

subcomponents, lower Alpha bands (Lo1 and Lo2)

have been found to relate to general attentional

demands and alertness (over the whole scalp) and the

upper Alpha has been found to react to changes in

task-specific processes (in topographically restricted

regions) (Klimesch, 1999).

As flow is not only repeatedly associated with

cognitive demands in the form of working memory

recruitment (i.e. Theta range activity), but also often

in relation with attentional processes (Harris et al.,

2017), it would seem of high interest to employ Alpha

band segmentation to not only identify regions of

reduced neural activity, but perhaps even identify

changes in global changes in attentional demands,

and task-specific pattern changes. The need for band

personalization has already been acknowledged in

flow research (Berta et al., 2013; Ewing et al., 2016),

yet no extensive evaluations have been completed, for

example studying personalized, narrow, and multiple

sub-bands. Such approaches would seem however to

cover several conceptually related processes of flow

like workload, attention and neural efficiency.

• Difficulty – Skill Balance

• Clear Goals

• Unambiguous Feedback

• Focused Concentration

• Merging of Action & Awareness

• Loss of Self-Consciousness

• Perceived Control

• Distortion of Time

• Cognition (e.g., Ease of Use)

• Affect (e.g., Enjoyment)

• Behavior (e.g., Actual Usage)

Consequences

Flow

Experience

Antecedents

CHIRA 2019 - 3rd International Conference on Computer-Human Interaction Research and Applications

184

3 METHOD

In the herein presented study, the adaptation of a

mental arithmetic DM task was chosen due to its pre-

validated nature in flow research (Katahira et al.,

2018). In this math task, the DM can be easily

achieved by increasing or decreasing the number of

digits that have to be summed in a fixed time frame

(here 28s per trial – with 4s breaks between trials).

The final result is in all cases a three-digit number,

yet the number of digits is adapted based on

performance (cf. Table 1). In the EASY condition,

very simple equations are shown, that could only have

the result 303 or 304. In the HARD condition, very

hard equations are shown that always are at minimum

9 levels higher (i.e. 9 digits more) than the level that

is calibrated to have the optimal difficulty for each

participant. Finally, in the OPTIMAL-Calc.

condition, the equations are of a moderate difficulty

as determined by an early calibration phase. Lastly, in

the OPTIMAL-Chos. condition, participants can

select the optimal task difficulty themselves.

Table 1: Examples of the Math task Difficulties.

EASY

(Level = 0)

OPTIMAL

(Level = 2)

HARD

(Level = 16)

100 + 1 +

100 + 2

100 + 13 +

100 + 22 +

100 + 3

100 + 35 + 22 + 16 + 2 +

100 + 64 + 45 + 26 +

100 + 25 + 51 + 31

Study participants were sampled from a public

student pool and received a compensation of 22 Euro.

After a preparation phase (consent, sensor

attachment, 5min eyes open resting and 1min eyes

closed resting), an introduction to the math task was

shown and participants could practice the task using

the EASY condition. Afterwards, the task was shown

in the OPTIMAL condition starting at level 1 in order

to calibrate the optimal difficulty. Next, all four math

task conditions were presented in randomized order

(Williams Design). The preparatory and all following

task conditions lasted for ca. 5 minutes. After each

condition, self-reports were collected for perceived

difficulty (1 item) and flow (10 item flow short scale

– FKS – both instruments by Engeser and Rheinberg,

2008). EEG data was collected with a saline-based

14-channel Emotiv Epoc+ headset (256Hz sampling).

Data was collected for 41 participants. Data from

two participants who repeatedly failed control

questions in the surveys were removed. For all self-

report variables, outliers were removed (>2 standard

deviations – SD – from the construct mean). Next,

Cronbach’s Alpha coefficients were inspected for the

flow construct and found acceptable after one item

was removed (0.80). The distribution normality was

assessed (univariate Shapiro-Wilk tests) and

supported for all conditions.

Table 2: EEG Data Pre-processing Pipeline.

Data Preparation (R)

Step

Parameters

Ref.

Data

Extraction

Baseline & Task

Phases

–

Channel

Centering

Channel Mean

Subtraction

–

Signal Processing (Matlab)

Step

Parameters

Ref.

Line Noise

Removal

50 Hz & 100 Hz

Bigdely-

Shamlo et

al., 2015

Re-

Referencing

Robust Common

Average Reference

Detrending

1 Hz High-Pass

–

Trim Outliers

800mV / 250ms

–

Channel &

Paroxysmal

Artefact

Removal

Artefact Subspace

Reconstruction

(ASR) - Burst

Criterion 10 SD

Mullen et

al., 2015

Stationary

Artefact

Removal

(Independent

Components)

AMICA - ICs:

[Horizontal &

Vertical Eye

Movement, Blinks,

Discontinuities]

Mognon et

al., 2011

Feature Extraction (R)

Step

Parameters

Ref.

Processing

Inspection

Visual Comparison

of Input & Output

–

Frequency

Power

Extraction

(Morlet

Wavelets)

55 Frequencies,

Range [3, 60],

Wavelet Cycle

Range [3,10] log.

spaced with

frequencies

Cohen,

2014

dB Power

Conversion

10*log10(mV

/Hz)

–

Frequency

Band

Extraction

Theta, Alpha &

Beta Bands from

IAF Peak

Klimesch,

1999

Change Score

Computation

∆ = Task – Eyes

Open Baseline

–

EEG data was processed primarily along the

guidelines of Cohen (2014). 7 data sets had to be

excluded due to recording errors, insufficient data

quality, or lack of survey report data. The retained,

complete sample for EEG analysis comprised 34

participants. Data preparation, feature extraction, and

analysis were conducted in R, signal processing and

Flow and Optimal Difﬁculty in the Portable EEG: On the Potentiality of using Personalized Frequency Ranges for State Detection

185

artefact removal in Matlab (EEGLab). The automated

EEG data preparation process is outlined in detail in

Table 2. Signal data was additionally screened before

and after signal processing to ensure no critical errors

occurred in the pipeline. Parameters for the

processing steps were tuned for the Epoc+ headset.

For almost all feature aggregation steps median

averaging was used to reduce the impact of outliers in

the data (Cohen, 2014).

Frequency bands were extracted following

Klimesch (1999). To account for inter-individual

differences, individualized Alpha frequency (IAF)

peaks were identified. As Alpha is known to also vary

regionally (being slower at anterior sites), yet as not

all participants showed such clear peaks for all sites,

a global IAF maximum was determined as lying 0.5

Hz below the occipital Alpha maximum during an

eyes-closed resting phase (cf. Figure 2).

Figure 2: PSDs for one participant during an eyes-closed

resting phase with pooled anterior (AF3, AF4, F3, F4, F7,

F8, FC5, FC6), temporal (T7, T8), parietal (P7, P8), and

occipital electrodes (O1, O2). Dots show regional peaks.

Figure 3: Grand Average PSD for all participants during an

eyes-closed resting phase with all 14 electrodes pooled,

demonstrating the frequency band decomposition into

narrow Theta, Alpha and Beta bands.

Based on this IAF, 2 Hz Theta and Alpha sub-

bands were extracted (cf. Figure 3). To extend the

personalized and band-refined approach, the Beta

band was similarly decomposed. In line with previous

research that has extracted low, mid, and upper Beta

bands with 3 Hz, 5 Hz, and 10 Hz ranges respectively

(Berta et al., 2013), the previous IAF-based

decomposition was continued using these ranges.

4 RESULTS

In the following analyses, one-way repeated measures

analyses of variance (ANOVA) with Greenhouse-

Geisser (GG) correction were used to assess main

effects, followed up by pairwise Welch’s t-Tests with

Benjamini-Hochberg (BH) correction. Error bars in

all figures show standard errors. To check the

manipulation success, perceptions of task difficulty

and flow are evaluated. For task difficulty, a

difference between the conditions is found (F(3, 102)

= 161.81, p < .01, 

= .71) with stepwise increases

in perceived task difficulty per condition (p < .01).

Also, for flow, a main effect for condition is found

(F(3, 99) = 29.63, p < .01, 

= .30). Follow-up tests

show a stepwise increase in flow from the EASY to

the OPTIMAL conditions, with OPTIMAL-Cal.

being increased from EASY at trend level (p = .085),

and OPTIMAL-Chos. being increased to both former

conditions (i.e. maximal) (p < .01). In the HARD

condition, flow is decreased compared to all other

conditions (p < .01), showing the expected inverted

U-shape pattern of flow with increasing difficulties

(cf. Figure 4). Together, the findings lead to the

assumption, that flow is increased in the OPTIMAL

conditions, with a maximum in OPTIMAL-Chos.

Figure 4: Perceived difficulty & flow reports.

To assess the changes in frequency band activity

over several scalp locations, separate one-way

repeated measures ANOVAs (GG corrected) were

conducted for each available electrode pair (i.e. data

CHIRA 2019 - 3rd International Conference on Computer-Human Interaction Research and Applications

186

were pooled for electrodes at AF3 & AF4 (= AF), F3

& F4 (= F-M), F7 & F8 (= F-L), FC5 & FC6 (= FC),

T7 & T8 (= T), P7 & P8 (= P), O1 & O2 (= O) (cf.

Table 3). All ANOVA significance levels were BH-

adjusted, as were the significance levels in the follow-

up pairwise Welch’s t-tests.

Table 3: EEG Frequency Power ANOVAs – Only showing

significant results. Exact p-values when p > .01.

Site

Test Result

Theta (IAF-6 to IAF-4)

F(3, 75) = 3.72, p = .0686, 

= .02

Lo2Alpha (IAF-2 to IAF)

F(3, 87) = 11.69, p < .01, 

= .04

F(3, 90) = 9.24, p < .01, 

= .02

HiAlpha (IAF to IAF+2)

F-M

F(3, 87) = 7.04, p < .01, 

= .03

F(3, 84) = 12.86, p < .01, 

= .04

F(3, 87) = 14.16, p < .01, 

= .04

Alpha (IAF-4 to IAF+2)

F(3, 87) = 11.52, p < .01, 

= .04

F(3, 90) = 8.89, p < .01, 

= .02

LoBeta (IAF+2 to IAF+5)

F-M

F(3, 87) = 10.65, p < .01, 

= .05

F(3, 90) = 7.70, p < .01, 

= .03

F(3, 93) = 11.55, p < .01, 

= .03

MidBeta (IAF+5 to IAF+10)

F-M

F(3, 84) = 3.34, p = .0840 , 

= .03

HiBeta (IAF+10 to IAF+20)

F(3, 81) = 5.97, p = .0213, 

= .08

F-M

F(3, 81) = 4.03, p = .0524, 

= .05

F-L

F(3, 78) = 7.00, p < .01, 

= .06

F(3, 75) = 5.39, p = .0196, 

= .05

F(3, 75) = 10.14, p < .01, 

= .09

F(3, 84) = 5.96, p = .0152, 

= .06

F(3, 87) = 5.65, p = .0193, 

= .04

Altogether, Theta power shows almost no

difference across the conditions, with only a trend

level effect being present at temporal locations. For

this location, post-hoc tests showed trend level

increases of OPTIMAL-Cal. (p = .0537) and

OPTIMAL-Chos. (p = .0952) from EASY as the only

differences. To complete this assessment, the

neighbouring Lo1Alpha band did not show changes

at any site, and neither did the Theta band when

extracted for a more traditional and non-

individualized 4-7.5 Hz range (all p > .1).

While no effects were found in the lowest Alpha

range (Lo1Alpha), both Lo2Alpha and HiAlpha

showed Alpha suppression at posterior regions (both

at P & O) with similar effects. Namely, a decrease in

Alpha from EASY, with all other conditions being on

the same level (all p < .01). For Lo2Alpha at occipital

sites (p = .0247) and HiAlpha at occipital sites (p =

.0513 - trend level), there was also a lower level in

HARD than OPTIMAL-Cal. indicating that occipital

alpha suppression was somewhat in line with

perceived difficulty increases. In relation, for LoBeta

at posterior sites, power reductions are found. At

parietal sites LoBeta power drops from EASY to

OPTIMAL-Cal. (p < .01) and OPTIMAL-Chos. (p =

.0321). At occipital sites, LoBeta power drops from

EASY compared to all other conditions (all p < .01)

(cf. Figure 5).

Figure 5: Posterior Alpha decreases from EASY are visible

in all Alpha frequencies. Occipital sites show more

sensitivity to difficulty changes.

While the same pattern of posterior Alpha

reductions is also generally visible in the broad 6 Hz

Alpha band (also here with occipital Alpha being

slightly higher at OPTIMAL-Cal. than HARD – p =

.0174), the most noticeable difference revealed

through the Alpha split is a reduction of frontal Alpha

power at sites closer to the midline (F-M), that is

visible in HiAlpha, but not in the lower Alpha bands,

nor the broad 6 Hz Alpha band. HiAlpha at F-M is

suppressed in all other conditions when compared to

EASY (all p < .05) (cf. Figure 6). This effect is

identically found in the adjacent LoBeta band (and in

MidBeta at trend level – here the post-hoc contrast is

visible for EASY compared to OPTIMAL only – and

HiBeta at trend level, p = .0893 and p = .0701 – also,

for F-M HiBeta a minimum in OPTIMAL is visible

due to an increase in HARD compared to OPTIMAL-

Cal. p = .0302 and OPTIMAL-Chos. p = .0185).

Furthermore, the (almost complete) absence of

effects in the MidBeta range further highlights the

utility of the narrower frequency inspection. In

particular, HiBeta is found to be increased at AF sites

in HARD compared to EASY (p = .0411), and

Flow and Optimal Difﬁculty in the Portable EEG: On the Potentiality of using Personalized Frequency Ranges for State Detection

187

OPTIMAL-Cal. compared to EASY at trend level (p

= .0721), at F-L sites in HARD compared to EASY

(p < .01), to OPTIMAL-Cal. (p = .0744 – trend level),

and to OPTIMAL-Chos. (p = .0150), with EASY

showing lower power than the OPTIMAL conditions

at trend level (p = .0608 and p = .0.798). At FC sites,

HiBeta is found to be increased in HARD compared

to all other conditions (all p < .05). At T sites, HiBeta

shows an increase from EASY to OPTIMAL-Chos.

(p = .0335) and from there an increase to HARD (p =

.0385). HARD is also higher compared to EASY (p <

.01) and OPTIMAL-Cal. (p = .0335). OPTIMAL

Chos. is found to be higher than OPTIMAL-Cal. at

trend level (p = .0778), which means that temporal

HiBeta slightly indicates a stepwise increase in power

that would be in line with difficulty perception

changes. HiBeta is also elevated in HARD at parietal

and occipital sites when compared to all other

conditions (all p < .05), with the other conditions

being equal. Altogether, this means that HiBeta

mostly reveals maxima during HARD task conditions

over the whole scalp (cf. Figure 7), which represents a

useful contrast to the lower frequency effects, in

particular the Alpha increase at posterior and fronto-

medial sites during the EASY condition.

Figure 6: Fronto-Medial Alpha and HiBeta progressions.

This points to an interesting potential of

combining (Hi-)Alpha and (Hi-)Beta powers, to

identify states of OPTIMAL difficulty. Such

approaches have been previously undertaken on the

Workload Index (Berka et al., 2007), that have also

been employed in flow research (Chanel et al., 2011)

and are traditionally either used for pooled electrodes

over the whole scalp or for central midline electrodes.

For the sake of comparison this Workload Index (WI

= Beta / (Theta + Alpha)) was also computed for

pooled electrodes here using non-individualized

broad bands (Theta = 4-7.5 Hz, Alpha = 7.5-12.5 Hz,

Beta = 12.5-30 Hz). No significant effect was found

in a one-way repeated measures ANOVA using the

difficulty conditions as within-subjects factor.

Figure 7: Whole Scalp HiBeta progressions – for better

visibility only the most distinctive patterns are shown (O

was very similar to P and FC very similar to F-L).

5 DISCUSSION AND

CONCLUSIONS

Of particular importance is the finding that through

frequency band separation, HiAlpha suppression in

frontal medial sites emerged independent of LoAlpha

frequencies, similarly to the work by Ewing et al.

(2016). This is an interesting finding that not only

points to the utility of frequency band separation, but

might provide an alternative to detection of increased

task demands in real-world scenarios where posterior

Alpha blocking has been named as a prominent

confounding factor (Blankertz et al., 2016).

Furthermore, through segmentation of the Beta band

it was also found that while a larger similarity is

visible between HiAlpha and LoBeta ranges, the

MidBeta range showed mostly no variation across

conditions, and the HiBeta range primarily showed

increases with very high task demands.

Using the refined approach, we did not find

features that clearly reflect the variation in reported

flow. This is not completely surprising given that

related work has not uncovered such markers using

frequency power comparisons in DM tasks.

Although, here, two observations tentatively indicate

such reactivity, namely temporal Theta and frontal

medial HiBeta activity at trend levels. Both are in line

with previous findings of increased Theta activity

during moderate to high task difficulty (Ewing et al.,

2016), and with documentations of negative

correlations between frontal Beta and flow self-

reports (Léger et al., 2014). Therefore, these results

would appear to warrant further investigation. As an

CHIRA 2019 - 3rd International Conference on Computer-Human Interaction Research and Applications

188

initial proposition, it would be plausible to consider

simultaneous, localized HiAlpha suppression

together with HiBeta reductions as a sign of neural

efficiency. The reasoning behind this thought builds

on the proposition of HiAlpha reflecting task-specific

information processing (Klimesch, 1999), and Beta

reflecting increased local communication (Buzsáki

and Draguhn, 2004). Therefore, HiAlpha suppression

could be indicative of a cortical region being recruited

to process a particular task, while at the same time a

reduction in local Beta would reflect a reduction in

communication among local neuron populations. In a

similar manner as frontal Theta and widespread Beta

increases are considered as coping mechanisms

during (too) hard tasks (Sauseng et al., 2005), a

reduction of local Beta might be indicative of the fact

that local regions are coping well and only recruit

absolutely required neuron groups.

Besides these theoretical potentials, the remaining

present findings show an alternative practical

potential to differentiate levels of perceived difficulty

(and thereby indirectly flow) through feature

combination. In this regard, it should be pointed out

first, that barely any significant changes were

detected in the Theta band. This was an unexpected

finding, as a large amount of literature is available

documenting frontal Theta increases with increasing

task demands (Klimesch, 1999; Borghini et al., 2014).

As to why such a pattern is here only reflected in trend

level temporal Theta power changes, several

explanations are possible. Theta changes have been

documented to occur strongly with prolonged task

exposure (e.g. in studies with airplane simulators –

Borghini et al., 2014). By adding the first task round

(the task introduction using the EASY treatment), to

a one-way repeated-measures ANOVA we do find a

difference for frontal Theta at F3 & F4 (F-M) (F(4,

104) = 4.33, p < .01, 

= .05), with the lowest Theta

power in the introduction phase. This points to the

fact that time might have acted as a confounding

effect on Theta power changes. Further, frontal Theta

effects are typically identified in midline positions

(e.g. Fz – see Ewing et al., 2016). Thus, as these

electrodes are missing for the Epoc+ headset, it might

not be possible to find Theta effects in some EEG

devices. This is a limitation that could easily occur in

real-world measurement scenarios using portable

EEG systems with few electrodes.

However, employing the frequency separation

approach, an interesting pattern emerged in that some

narrow frequency features showed significant

reaction to changes in perceived difficulty, with some

indicating stepwise, monotonous increases, some

indicating increases with moderate level plateaus for

OPTIMAL conditions and some indicating maxima

for either the EASY or the HARD condition when

compared to all other conditions. The functional

explanation of these patterns should be the subject of

future work as would be the development of sensitive

and robust compound indices, ideally based on more

than one task type (see e.g. Berka et al., 2007).

Presently, it is primarily argued that these patterns

allow to discuss flow related changes in a refined

manner and that they pose interesting alternatives for

the detection of situations of optimal difficulty,

especially in scenarios where less information might

be available than typically is in laboratory setups (e.g.

fewer and unevenly distributed electrodes). When

considering how for example neuro-adaptive systems

employ thresholds to inform adaptation rules (cf. e.g.

Ewing et al., 2016), features indicating maxima

during EASY or HARD conditions could be valuable

indicators, given that they would be subject to lower

variation except in the boundary cases. In this regard

they could firstly be employed to robustly identify

when difficulty is unbalanced and flow unlikely.

In conclusion, this study posits that flow research

could benefit from nuanced frequency power

analyses, in general by identifying Alpha and Beta

power changes that could relate to neural efficiency

(in a local form), and in particular when (portable)

EEG systems are used that lack midline electrodes. In

line with previous research (Klimesch, 1999), the

herein presented initial analyses support the

understanding that a personalized and narrow

frequency power analysis helps to avoid to miss

frequency specific effects. Specifically, this research

contributes to the literature on flow by highlighting

that frontal medial HiAlpha decreases in increased

task difficulty (and HiBeta decreases during

optimally balanced task difficulties) as well as

widespread HiBeta increases in very hard task

conditions provide additional avenues to

automatically and unobtrusively detect boundary

situations to flow. Thus, these metrics, indirectly

allow to improve the identification of situations with

optimal preconditions for flow to emerge.

Importantly, the frequency segmentation approach

would appear to be a valuable alternative when

portable EEG systems are used that don’t include

midline electrodes. As an additional effect, Alpha and

Beta bands appear unaffected by influences from task

exposure durations providing an interesting

alternative to the more established metric of frontal

midline Theta power increases with higher task

demands (e.g. Ewing et al., 2016). Ideally by taking

narrow frequency power analysis into account, future

flow research will thus move closer to identifying

Flow and Optimal Difﬁculty in the Portable EEG: On the Potentiality of using Personalized Frequency Ranges for State Detection

189

robust concomitants and markers of flow that can be

employed in neuro-adaptive systems using portable

EEG in real-world scenarios. Eventually, systems

able to adapt to flow intensities could then reduce

flow interruptions (e.g. by blocking incoming

messages) or provide feedback information to

improve flow self-regulation (e.g. by self-adjusting

task difficulty, or by optimizing arousal levels and

catalysing task focus through EEG-neurofeedback).

REFERENCES

Berka, C. et al., (2007) ‘EEG Correlates of Task

Engagement and Mental\nWorkload in Vigilance,

Learning, and Memory Tasks’, Aviation, Space, and

Environmental Medicine, 78(5), pp. B231–B244.

Berta, R. et al., (2013) ‘Electroencephalogram and

Physiological Signal Analysis for Assessing Flow in

Games’, IEEE Transactions on Computational

Intelligence and AI in Games, 5(2), pp. 164–175.

Bigdely-Shamlo, N. et al., (2015) ‘The PREP pipeline:

standardized preprocessing for large-scale EEG

analysis’, Frontiers in Neuroinformatics, 9(June), pp.

1–20.

Blankertz, B. et al., (2016) ‘The Berlin Brain-Computer

Interface: Progress Beyond Communication and

Control’, Frontiers in Neuroscience, 10, pp. 1–24.

Borghini, G. et al., (2014) ‘Measuring neurophysiological

signals in aircraft pilots and car drivers for the

assessment of mental workload, fatigue and

drowsiness’, Neuroscience and Biobehavioral Reviews.

Elsevier Ltd, 44, pp. 58–75.

Buzsáki, G. and Draguhn, A. (2004) ‘Neuronal Oscillations

in Cortical Networks’, Science, 304(5679), pp. 1926–

1929.

Ceja, L. and Navarro, J. (2012) ‘“Suddenly I get into the

zone”: Examining discontinuities and nonlinear

changes in flow experiences at work’, Human

Relations, 65(9), pp. 1101–1127.

Chanel, G. et al., (2011) ‘Emotion Assessment From

Physiological Signals for Adaptation of Game

Difficulty’, IEEE Transactions on Systems, Man, and

Cybernetics – Part A: Systems and Humans, 41(6), pp.

1052–1063.

Cheron, G. (2016) ‘How to measure the psychological

“flow”? A neuroscience perspective’, Frontiers in

Psychology, 7(DEC), pp. 1–6.

Cohen, M. X. (2014) Analyzing neural time series data:

theory and practice. MIT press.

Csikszentmihalyi, M. (1996) Flow and the psychology of

discovery and invention. 1st edn. New York: Harper

Collins.

Dietrich, A. (2004) ‘Neurocognitive mechanisms

underlying the experience of flow’, Consciousness and

Cognition, 13(4), pp. 746–761.

Engeser, S. and Rheinberg, F. (2008) ‘Flow, performance

and moderators of challenge-skill balance’, Motivation

and Emotion, 32(3), pp. 158–172.

Ewing, K. C. et al., (2016) ‘Evaluation of an Adaptive

Game that Uses EEG Measures Validated during the

Design Process as Inputs to a Biocybernetic Loop’,

Frontiers in Human Neuroscience, 10, pp. 1–13.

Harris, D. J. et al., (2017) ‘Neurocognitive mechanisms of

the flow state’, Progress in Brain Research, 237(1), pp.

221–243.

Katahira, K. et al., (2018) ‘EEG correlates of the flow state:

A combination of increased frontal theta and moderate

frontocentral alpha rhythm in the mental arithmetic

task’, Frontiers in Psychology, 9, pp. 1–11.

Klimesch, W. (1999) ‘EEG alpha and theta oscillations

reflect cognitive and memory performance: a review

and analysis’, Brain Research Reviews, 1(29), pp. 169–

195.

Léger, P. M. et al., (2014) ‘Neurophysiological correlates

of cognitive absorption in an enactive training context’,

Computers in Human Behavior, 34, pp. 273–283.

Mognon, A. et al., (2011) ‘ADJUST: An automatic EEG

artifact detector based on the joint use of spatial and

temporal features’, Psychophysiology, 48, pp. 229–240.

Moller, A. C. et al., (2010) ‘Developing an experimental

induction of flow: Effortless action in the lab’, in Bruya,

B. (ed.) Effortless Attention, pp. 191–204.

Mullen, T. R. et al., (2015) ‘Real-time Neuroimaging and

Cognitive Monitoring Using Wearable Dry EEG’,

IEEE Transactions on Biomedical Engineering, 62(11),

pp. 2553–2567.

Sauseng, P. et al., (2005) ‘Fronto-parietal EEG coherence

in theta and upper alpha reflect central executive

functions of working memory’, International Journal

of Psychophysiology, 57(2), pp. 97–103.

CHIRA 2019 - 3rd International Conference on Computer-Human Interaction Research and Applications

190