Alpha Neurofeedback Training with a portable Low-Priced and Commercially Available EEG Device Leads to Faster Alpha Enhancement

Introduction Findings of recent studies have proposed that it is possible to enhance cognitive capacities of healthy individuals by means of individual upper alpha (around 10 to 13.5 Hz) neurofeedback training. Although these results are promising, most of this research was conducted based on high-priced EEG systems developed for clinical and research purposes only. This study addresses the question whether such effects can also be shown with an easy to use and comparably low priced Emotiv Epoc EEG headset available for the average consumer. In addition, critical voices were raised regarding the control group designs of studies addressing the link between neurofeedback training and cognitive performance. Based on an extensive literature review revealing considerable methodological issues in an important part of the existing research, the present study addressed the question whether individual upper alpha neurofeedback has a positive effect on alpha amplitudes (i.e. increases alpha amplitudes) and short-term memory performance focussing on a methodologically sound, single-blinded, sham controlled design. Method Participants (N = 33) took part in four test sessions over four consecutive days of either neurofeedback training or sham feedback (control group). In the experimental group, five three-minute periods of visual neurofeedback training were administered each day whereas in the control group, the same amount of sham feedback was presented. Performance on eight digit-span tests as well as participants’ affective states were assessed before and after each of the daily training sessions. Results Participants in the neurofeedback training (NFT) group showed faster and greater alpha enhancement compared to the control group. Contrary to the authors’ expectations, alpha enhancement was also observed in the control group. Surprisingly, exploratory analyses showed a significant correlation between the initial alpha level and the alpha improvement during the course of the study. This finding suggests that participants with high initial alpha levels profit more from alpha NFT interventions. digit-span performance increased in both groups over the course of time. However, the increase in individual upper relative alpha did not explain significant variance of digit-span improvement. In the discussion, the authors explore the appearance of the alpha enhancement in the control group and possible reasons for the absence of a connection between NFT and short-term memory.

158 A study of high importance for the development of IUA feedback addressed the topic by means of 159 transcranial magnetic stimulation in a within-subject design [33]. In line with the correlational findings 160 between alpha desynchronization and cognitive performance, the participants were stimulated to 161 produce more IUA activity (individual alpha peak + 1 Hz) at P6 and Fz before the execution of a task. 162 In this way, the natural desynchronization process which can be found in participants showing high 163 cognitive performance (i.e. mental rotation and short-term memory performance) was mimicked. In the 164 control condition, participants also 'underwent' transcranial magnetic stimulation, but the coil was 165 rotated by 90° so the participants did not receive any stimulation. The results show a significant increase 166 of IUA during transcranial magnetic stimulation in the experimental condition, as well as decreased test 167 power, resulting in a large event-related desynchronization. None of these changes were observed in the 168 control condition. Cognitive performance was assessed in terms of success in a Mental Rotation Task. 169 Results showed that mental rotation performance in the experimental condition was higher compared to 170 the control condition. The authors interpreted these findings as an indicator for a causal relationship 171 between IUA activity and cognitive performance in healthy subjects.
172 Based on these findings, several studies examined the connection between IUA activity and cognitive 173 performance. In those studies, different aspects of cognitive performance like short-term memory 174 performance or working memory performance were assessed via a digit-span Task or a Conceptual 175 Span Task (e.g. [5,41]), or Mental Rotation Task [42]. Mental flexibility and executive functions were 176 assessed via the Trail Making Test [43], or creativity by the Unusual Uses Test [44]. Summarizing the 177 results of these studies, imply a positive connection between individual upper alpha NFT and different 178 aspects of cognitive performance like working memory/STM and visuospatial rotation. Whether the 179 relationship between IUA and STM is of causal or correlational character, which underlying 180 mechanisms lead to the enhancing effect of individual upper alpha NFT on cognitive performance and 181 whether unspecific environmental factors of the experimental setup play a key role in the process of 182 NFT is still not fully understood at the moment. In the following section, some of these aspects are 183 addressed by a comprehensive analysis of published studies addressing the link between IUA and 184 cognitive performance. 249 After being duly informed about the protocol of the study, all participants agreed to written informed 250 consent authorized by the ethical committee of the University of Fribourg. As a compensation for their 251 participation, they earned 5 credit points on a university-intern reward system. Participants were 252 assigned randomly to either the experimental neurofeedback training group (NFT group, n 1 = 17, M Age 253 = 21.29, 12 female) or the control sham feedback group (SF group, n 2 = 16, M Age = 21.13, 14 female). 254 To assess whether subjects were aware of their condition, the last experimental task was to guess which 255 group they were assigned to. The statement 'I was assigned to the control group' was answered on a 7-256 point Likert scale (ranging from 'I strongly disagree' to 'I strongly agree' 279 NFT or Sham feedback (SF) started immediately after the baseline recordings and consisted of five 3-280 min periods with a 30 second break in-between. For S1, participants first received verbal and written 281 information about alpha activity and were encouraged to be creative and come up with five personal 282 strategies for the five periods of NFT (or SF). A list with five strategies (positive thinking, evoking 283 emotions, visualizing activities, love and physiological calm) based on [5] was offered to participants 284 who had difficulty coming up with their own ideas. Participants were asked to use only one strategy 285 during each period, write it down during the break and to use every strategy only once over the course 286 of the five periods. This procedure allowed to determine the most-successful alpha-enhancing strategy 287 (one strategy, which produced the highest relative IUA for each participant). The participants were 288 instructed to use their most successful strategy during the following sessions of S2 to S4.
289 At the end of each session, participants repeated the digit span test and the MDBF. For a schematic 290 overview of the procedure during each of the sessions S1 to S4 see  Procedure within sessions S1 to S4. Procedure during each of the sessions S1 to S4 in experimental (NFT) 295 and control (SF) group.

296
-12-297 2.3. Neurofeedback Training 298 Feedback sites P7, O1, O2 and P8 were chosen for their connection to visual and attentional processes 299 (see e.g. [83,84]). Using a simple channel spectra procedure in EEGLAB (pop_fourieeg), each 300 session's baseline recordings was analysed to determine the IUA frequency band, which was then 301 used in that session. More specifically, the individual alpha peak (IAP) between 7.5 and 12.5 Hz [12] 302 was assessed from the eyes closed resting condition and the lower and upper border of the IUA 303 frequency band were defined as IAP and IAP + 2, respectively. We used the Emotiv 3D Brain Activity 304 Map standalone software to provide IUA feedback with a colour spectrum ranging from grey (low 305 IUA amplitude) over green to red (high IUA amplitude). During each session's period, participants 306 watched their real-time IUA activity at occipital and parietal sites (P7, O1, O2, and P8) colour-coded 307 on the surface of an animated head (see Fig 3) and were advised to produce as much red activity as 308 possible. Participants in the experimental group performed IUA NFT always with a real-time IUA 309 band feedback. The control group received SF by watching recordings of NFT sessions from another 310 subject not included in this sample.

314 2.4. EEG Recording and Processing
315 An Emotiv Epoc EEG headset was used for EEG baseline recordings and NFT. It has 14 channels (AF3, 316 AF4, F7, F3, F4, F8, FC5, FC6, T7, T8, P7, P8, O1 and O2, international 10-20 system) and uses passive 317 saline sensors. The device is wireless and transmits data via Bluetooth through the 2.4 GHz band, has a 318 battery autonomy of 12 hours and uses a built-in amplifier, as well as a CMS-DRL circuit for the 319 reduction of external electrical noise. It has a sampling rate of 128 bit/s, a bandwidth ranging from 0 to 320 64 Hz, automatic digital notch filters at 50 Hz and 60 Hz and the dynamic range referred to the input is 321 8400µV(pp). Moreover, a digital 5 th order Sinc filter is built-in and impedance can be measured in real-322 time. EEG was recorded using the software Emotiv TestBench, ground-reference was M1 and sampling 323 method was by default sequential sampling.
324 All analyses were carried out with MATLAB and EEGLAB [85]. The data was pre-processed using the 325 following methods: re-referencing to channel M1, automatic removal of bad epochs using the command 326 pop_autorej, calculation of IC weights with the runica algorithm. Following [34], the relative alpha 327 values for both, NFT and SF were calculated from the pre-processed EEG by dividing the mean 328 amplitude of the IUA band (defined individually in the same way as in the NFT, between the IAP and 329 IAP + 2 Hz) by the mean amplitude of the entire EEG bandwidth (Equation 1). ℎ = ℎ EEGAmplitude 0.5 -64 (1) 330 This normalization was applied to avoid variability in the absolute amplitude between trials and sessions 331 due to changes in impedance between electrodes and scalp. This way, attenuations caused by external 332 factors that affect all frequency bands are mitigated. Furthermore, we worked with amplitude instead of 333 power values to prevent excessive skewing and improve the validity of the statistical analysis [6]. 334 2.5. Subjective and Objective Measures 335 Questions about physical activities, substance intake and sleep were assessed with a self-made 336 questionnaire. Short-term memory performance was assessed by means of a forward digit span test 337 taken from the PEBL test battery [86]. During this test, digits appeared on the screen and participants 338 were advised to memorize them. On first trial, three digits appeared one after another and the participant 339 typed them into an input field in the same order as they had appeared. In case of a correct answer, a 340 positive feedback was given and the trial was repeated with the same number of digits. If the participant 341 succeeded again, the number of digit was increased by one. The test continued until the participant typed 342 in a wrong answer on two consecutive trials. Two performance indicators were assessed. One is the 343 digit span itself, defined as the highest amount of digits the participants remembered correctly. Another 344 measure is the total correct value, representing the total number of correct answers. For example, a digit 345 span of 9 indicates that the participant was able to remember 9 digits correctly. The total correct value 346 in this example however can vary between 7 and 16 because participants were allowed to continue with 347 the test if they made an error in one trial (e.g. they remembered 8 digits only once). In the statistical 348 analysis of the present study, only total correct values are reported. 349 2.6. Statistical Analyses 350 All analyses were carried out with the IBM Statistical Package for Social Sciences (SPSS version 24) 351 and R [87]. If not indicated differently, the chosen level of significance for all analyses was α = .05 352 (5%). Data were analysed with several mixed-measures design ANOVAs and corresponding contrast 353 analyses using either a polynomial or a simple algorithm. Concerning the mixed design ANOVAs, 354 Mauchly's Test of Sphericity was taken into account. If Mauchly's Test was not significant (p ≥ .20), 355 sphericity was assumed. When Mauchly's Test was significant (p < .20) and Greenhouse-Geisser 356 Epsilon was smaller than .75, Greenhouse-Geisser corrected results were reported. When Mauchly's 357 Test was significant (p < .20) and Greenhouse-Geisser Epsilon was larger than .75, Huynh-Feldt 358 corrected results were reported. 359 The general connection between alpha and digit-span was assessed with linear regressions. 360 Additionally, for a more detailed picture, paired-samples t-tests with Bonferroni adjustment were 361 conducted. More specifically, during each analysis (e.g. the 20x2 mixed designs ANOVA), Bonferroni 362 correction was applied by multiplying the p-value of all associated t-tests by the number performed t-363 tests. 364 For the present study, only the change of alpha and digit-span and not their general level was of interest. 365 Hence, all alpha measurements were standardized by subtracting the mean value of the first 366 measurement. By applying this standardisation to experimental group and control group separately, it 367 was assured both groups had the same initial value of alpha and digit-span, respectively. Digit-span 368 values were not standardized because they did not differ during the first measurement (M NFT = 8.76, 369 SD NFT = 1.82; M SF = 7.94, SD SF = 1.81; t(31) = 1.31, p = .20). 370 2.6.1. Complementary analyses for selected extraneous variables 371 Session related changes in mood. Thirteen extraneous variables related to mood and effort were 372 collected before and after each experimental session (see Figure 7B for details). We computed session 373 related changes for each one of these variables and used principal component analysis (PCA) for feature 374 extraction. The number of principal components (PCs) was chosen by interpreting the scree plot, and 375 choosing the number of components until when diminishing returns would be obtained, guaranteed that 376 at least 60% of variance could be explained. Each PC was then used as the response variable of a linear 377 mixed effects model, resulting in one linear model for each PC. Each model had two random variables: 378 subject as the random intercept and session number as the random slope. The fixed effect term was a 379 triple interaction between period, experimental group and changes in relative alpha. Changes in relative 380 alpha at each period were computed using an area under the curve with respect to increase (AUCi) 381 formula described in [88], and these values were averaged for each session. Satterthwaite approximation 382 for the degrees of freedom was used to compute p-values with the lmerTest package in the R 383 programming environment [89].
384 Analysis of the extraneous variable pre-session relaxation. Here, we aimed at investigating if the 385 inclusion of pre-session relaxation increases the predictive validity of the experimental group in changes 386 for relative alpha for each session. A mixed effects model was used to predict the session average 387 relative alpha AUCi, using each subject as the random intercept and session number as the random 388 slope. The fixed effect term was the moderation between experimental group and pre-session relaxation. 389 The moderation was introduced to understand if pre-session relaxation increases in alpha would be 390 specific to one of the experimental groups. If the interaction term was not significant, we would test the 391 additive model. For the latter, pre-session alpha would be tested as a suppressor variable. Satterthwaite 392 approximation for the degrees of freedom was used to compute p-values with the lmerTest package in 393 the R programming environment. 528 levels are presented for uncorrected p-values. When Bonferroni correction is applied, to control for Type I errors 529 due to the comparing models for five PCs, the triple interaction term in panel C is no longer significant (p corrected = 530 0.152). For panels D and E, when applied, Bonferroni correction controls for two comparisons made in the simple 531 main effects analysis resulting in a significant interaction effect 'Session x relative alpha' AUCi (p corrected = 0.028). 8 HERE   533 Fig 8. Linear mixed effect model for relative alpha AUCi as the response variable and the interaction between how 534 relaxed participants were at the beginning of the Session and experimental group as predictors. Coefficient 535 estimates and standard errors (SE) depicted as dot and line respectively. Red and blue colors represent positive and 536 negative coefficient estimates, respectively. Significance levels: ** p < .01, * p < .05. 537 4. Discussion 538 The present study investigated the connection between IUA alpha NFT, relative alpha and short-term 539 memory performance using a commercially available BCI device (Emotiv Epoc) in a single-blind 540 design. In line with previous results [5], an enhancing effect of the training on the relative alpha and on 541 short-term memory performance (digit-span Task) was expected. 542 Our analyses showed a significant improvement in relative alpha in the neurofeedback group between 543 period 1 and period 20 which could not be observed in the sham feedback group. Moreover, contrasts 544 showed that the increase in alpha was obtained much earlier (period 3) for participants who saw their 545 real-time brain activity compared to participants who followed a sham-feedback intervention (period 546 11). Additionally, we also observed that if the level of relaxation before each Session is taken into 547 account, it is possible to observe a clear effect of NFT on the production of relative alpha. All in all, the 548 results regarding relative alpha indicate that up-training alpha with a real-time NFT procedure facilitates 549 the process of enhancing alpha activity. Thus, the results of the present study were in accordance with 550 hypothesis H1. 551 Furthermore, we hypothesised the control group would not show an enhancement in alpha activity. 552 Interestingly though, contrast analyses showed the opposite was true and mixed measure longitudinal 553 analysis indicated that the slope for both groups were equal. One possible explanation for this finding 554 is that the control group was given feedback during the first session. Although sham feedback was used 555 for this group, by the end of this session they were informed of the most successful mental strategies 556 for alpha upregulation. This would imply that it is possible to infer appropriate mental strategies within 557 one session and that coherent visual feedback might not be necessary for ensuing sessions to upregulate 558 alpha to a certain degree, provided the adequate mental strategy is used. This interpretations seems to 559 be in line with studies showing an alpha enhancing effect of certain types of meditation (e.g. [48][49][50]). 560 A replication study with an additional control group that would not be informed about which strategies 561 are generally linked to alpha enhancement might address this question. Another possible explanation 562 resides in the framework of socio-physiological processes. Alpha is enhanced by being in a calm and 563 resting state and in the course of the current study, participants became more and more familiar with the 564 experimental environment, as well as with the experimenter. It is likely that the participants became 565 more and more relaxed, comfortable and calm during the later sessions of NFT/SF, which might have 566 led to the observed enhanced level of alpha in the control group. This interpretation finds some support 567 in the analysis of extraneous variables, which suggests that being relaxed is an important factor for