[1] 12w Neuroimaging Study of Alpha and Beta EEG Biofeedback Effects on Neural Networks

[1] 344w Neural networks interaction was studied in healthy men (20-35 years old) who underwent 20 sessions of EEG biofeedback training outside the MRI scanner, with concurrent fMRI-EEG scans at the beginning, middle, and end of the course. The study recruited 35 subjects for EEG biofeedback, but only 18 of them were considered as "successful" in self-regulation of target EEG bands during the whole course of training. Results of fMRI analysis during EEG biofeedback are reported only for these "successful" trainees. The experimental group (N = 23 total, N = 13 "successful") upregulated the power of alpha rhythm, while the control group (N = 12 total, N = 5 "successful") beta rhythm, with the protocol instructions being as for alpha training in both. The acquisition of the stable skills of alpha self-regulation was followed by the weakening of the irrelevant links between the cerebellum and visuospatial network (VSN), as well as between the VSN, the right executive control network (RECN), and the cuneus. It was also found formation of a stable complex based on the interaction of the precuneus, the cuneus, the VSN, and the high level visuospatial network (HVN), along with the strengthening of the interaction of the anterior salience network (ASN) with the precuneus. In the control group, beta enhancement training was accompanied by weakening of interaction between the precuneus and the default mode network, and a decrease in connectivity between the cuneus and the primary visual network (PVN). The differences between the alpha training group and the control group increased successively during training. Alpha training was characterized by a less pronounced interaction of the network formed by the PVN and the HVN, as well as by an increased interaction of the cerebellum with the precuneus and the RECN. The study demonstrated the differences in the structure and interaction of neural networks involved into alpha and beta generating systems forming and functioning, which should be taken into account during planning neurofeedback interventions. Possibility of using fMRI-guided biofeedback organized according to the described neural networks interaction may advance more accurate targeting specific symptoms during neurotherapy.

[1] 122w Electroencephalographic (EEG) biofeedback (i.e., neurofeedback) is a voluntary modification of amplitude and/ or power characteristics of EEG rhythms using auditory or visual feedback signal. This technology has long been successfully used in clinical practice to treat attention deficit hyperactivity disorder (Arns et al. 2009(Arns et al. , 2013;;Lubar 2003;Sherlin et al. 2010), addictions (Peniston and Kulkosky 1989;Sokhadze et al. 2008), depression (Baehr et al. 2001;Rosenfeld et al. 1995), anxiety (Hammond 2005), post-traumatic stress disorder (Peniston and Kulkosky 1991;van der Kolk et al. 2016), chronic pain syndrome (Jensen et al. 2007), epilepsy (Sterman and Egner 2006), and other neurological and psychiatric disorders (Thompson and Thompson 2015). However, the basic neurophysiological mechanisms underlying self-control of various characteristics of EEG rhythms still remain insufficiently investigated.

[2] 140w The emergence of functional magnetic resonance imaging (fMRI), which records blood oxygenation level dependent changes during brain activity, for the first time allowed noninvasive investigation of the functional anatomy of cognitive activity with a high level of accuracy. There is a number of studies devoted to fMRI-based investigation of resting state alpha rhythm (Bowman et al. 2017;de Munck et al. 2007;Goldman et al. 2002;Laufs et al. 2003Laufs et al. , 2006)), but only few publications used fMRI to evaluate the effects of an EEG biofeedback course conducted outside of MRI scanner (Nicholson et al. 2016;Ros et al. 2013). We were, to best of our knowledge, the first group to record the process of neurofeedback in the fMRI environment at the different stages of training course, the preliminary results of which were published earlier (Kozlova et al. 2016;Shtark et al. 2015).

[3] 124w The modern approach to investigation of the cognitive processes suggests a need to consider the brain as a system consisting of a multitude of networks that are spatially distributed but at the same time are functionally connected. The theory of alpha rhythm origin and the practice of its therapeutic application in the biofeedback mode have not been subjected for a long time to a certain "renewal" process. It seems timely to revisit and update theoretical concepts using new technologies, starting with fMRI, its synergistic combination with EEG, and the neural network vocabulary. In this study we investigated the dynamic changes in neural networks activity using independent component analysis (ICA) to detect a combination of fMRI image elements (i.e., voxels) that demonstrate similar spatial dynamics.

[4] 64w The goal of the current study was to compare the effects of alpha and beta rhythm neurofeedback on neural network interactions. Considering diametrical functional difference and source localization of alpha and beta EEG rhythms, we expected to find substantial differences in underlying neural networks activities, possibly as well some common (shared) regulation elements contributing to process of exercising voluntary self-control of EEG during neurofeedback.

[1] 61w Analysis of the obtained fMRI data was conducted only for the "successful" performers (N = 13 in alpha and N = 5 in beta group) that included subjects who succeeded in increasing the power of their "target" rhythm at the intervals between the first-to-second (i.e., first to tenth neurofeedback session) and the second-to-third (i.e., eleventh to twentieth neurofeedback session) MRI scans.

[2] 17w When comparing the alpha and beta training group subjects who have successfully completed the training (Fig. 1;

[3] 96w Table 2), the following significant differences in the functional connectivity level were found. In the first session (Fig. 1a; Table 2A), a significantly higher level for alpha power compared to beta power was found for the visuospatial network (VSN) between its occipital and frontal components. In the second session (Fig. 1b; Table 2B) both alpha and beta groups showed negative connectivity between the cuneus and the right executive control network (RECN), significantly less for beta, and greater connectivity for beta group between the networks formed by primary visual network (PVN) and high level visual network (HVN).

[4] 103w In the third session (Fig. 1c; Table 2C), alpha group significantly exceeded beta group on connectivity level for the following pairs: Auditory network (AN)-RECN (occipital component), AN-Language network (LN), LN-VSN, Cerebellum-Precuneus, RECN (component comprising right frontal areas)-Cerebellum, RECN (component comprising right frontal areas)-Anterior salience network (ASN). Smaller values of functional connectivity for alpha were determined for the pair RECN (component comprising right frontal areas)-VSN. Thus, the greatest differences between subjects receiving feedback on alpha and beta-rhythm were noted in the RECN: for those subjects who underwent alpha training this network showed significantly higher connectivity with AN, cerebellum, and ASN, and significantly smaller-with VSN.

[5] 88w When assessing the dynamics of neural networks for those who successfully completed alpha training (Fig. 2; Table 3), the following results were obtained. From the first to the second fMRI session (Fig. 2a; Table 3A), there was a weakening of functional connectivity of RECN (occipital work, DMN default mode network, HVN high level visual network, LECN left executive control network, LN language network, PN precuneus network, PVN primary visual network, RECN right executive control network, VSN visuospatial network. The arrows show connectivity between the pairs of neural networks

[6] 39w areas) elements with cuneus and HVN, while the frontal parts of the right hemisphere network were desynchronized with VSN, which in turn was desynchronized with the cerebellum. In addition, the interrelation between the cuneus and precuneus networks was weakened.

[7] 118w From the second to the third session of alpha neurofeedback (Fig. 2b; Table 3B), the situation reversed and was characterized by a connectivity increase in the following neural network pairs: PVN-HVN, Cerebellum-AN, LN-DMN, LN-right occipital VSN, Precuneus-VSN, Precuneus-ASN, ASN-VSN, ASN-RECN, and component comprising the frontal parts of both hemispheres-left occipital component of VSN. When assessing the dynamics of the changes of neural networks relations from the first to the third session (Fig. 2c; Table 3C), significant differences were obtained for the following connectivity pairs: HVN-Cerebellum, HVN-VSN (occipital areas of the right hemisphere), HVN-Precuneus, HVN-Cuneus, HVN-PVN, Cerebellum-VSN (component comprising frontal sections of both hemispheres), PVN-VSN (occipital left hemisphere), Cuneus-VSN, Cuneus-RECN, Precuneus-ASN, ASN-VSN (component comprising frontal sections of both hemispheres).

[8] 27w When analyzing the dynamics of the change in the level of neural network connectivity in five subjects who have successfully performed training with the beta rhythm feedback

[9] 158w Table 2 Comparison of connectivity coefficients of networks in two groups receiving feedback on alpha and beta rhythms, only for the "successful" performers. "alpha-beta" column refers to the difference between coefficients measured for alpha and beta groups Network pairs Beta (N = 5) Alpha (N = 13) Alpha-beta difference t Probability (p) (А) VSN-LECN 0.03 0.28 0.25 2.712 0.024 (B) HVN-PVN 0.63 0.46 -0.17 -2.685 0.016 Cuneus-RECN -0.29 -0.08 0.21 3.407 0.008 (C) AN-RECN -0.05 0.14 0.19 2.443 0.027 AN-LN 0.12 0.43 0.31 3.275 0.008 Cerebellum-PN -0.09 0.09 0.18 2.275 0.038 Cerebellum-RECN 0.10 0.30 0.20 3.592 0.005 RECN-ASN 0.41 0.53 0.12 3.101 0.007 RECN-VSN 0.40 0.23 -0.17 -2.267 0.045 VSN-LN 0.09 0.35 0.26 2.575 0.021 work, DMN default mode network, HVN high level visual network, LECN left executive control network, LN language network, PN precuneus network, PVN primary visual network, RECN right executive control network, VSN visuospatial network. The arrows show connectivity between the pairs of neural networks

[10] 96w (Fig. 3), the following connections were found. Between the first and the second neurofeedback session in the fMRI setting (Fig. 3a; Table 4A), a significant decrease in connectivity was obtained for Precuneus-cuneus and LN-RECN pairs. From the second to the third session (Fig. 3b; Table 4B), the connectivity in the LN-DMN pair was weakened. From the first to the third session (Fig. 3c; Table 4C), the interrelation in the Cuneus-PVN pair decreased, while increasing between the component comprising the frontal parts of both hemispheres and the component comprising the occipital parts of the left hemisphere VSN.

[1] 30w Based on the above findings, we consider the probable causes of the most interesting, from our point of view, trends in the formation of neural networks during the biofeedback training.

[2] 84w (1) Isolation of the cerebellum from VSN can be associated with the changes of the neural map, extracting and recalling memories in a new and more effective way. Presumably this is due to the fact that most of the subjects used the ideomotor imagination of their actions and movements to increase the alpha rhythm power, which initially triggered cerebellum activation (readiness for movement); however, as no real action was required to achieve the desired effect, the cerebellum neural network was excluded from the process.

[3] 57w (2) Desynchronization of the interconnections of the RECN, primarily with the cuneus network, is most likely caused by the decrease in the searching control, as a result of the decreased relevance (after 2 weeks of training the subjects easily reproduced their own memories), or as a result of functional antagonism of the processes of control and relaxation.

[4] 116w (3) Regarding strengthening of the modulating influence of the precuneus, the cuneus, and the network of spatial information visual processing on the network processing of high-level visual information: this set of changes partially resembles those observed in the Virtual Reality experiments (Clemente et al. 2014), and reflects the intensification of visual processing, that appears to be the most promising and significant in terms of implementing therapeutic effects of alpha training. (4) Strengthening of the interaction of anterior salience network with precuneus network is most likely due to the increased focusing on the desired memories and stimuli, both external (feedback) and internal (subjective sensations accompanying the memories that bring about an increase of the alpha rhythm power).

[5] 156w The results are consistent with earlier data on the alpha rhythm correlation with DMN (Hlinka et al. 2010;Jann et al. 2009;Mantini et al. 2007) and ASN (Sadaghiani et al. 2010) connectivity, as well as increased connectivity of ASNdorsal anterior cingular divisions (Ros et al. 2013). It is also network, HVN high level visual network, LECN left executive control network, LN language network, PN precuneus network, PVN primary visual network, RECN right executive control network, SN salience network, VSN visuospatial network. The arrows show connectivity between the pairs of neural networks consistent with fMRI studies of alpha rhythm during resting state, showing alpha activity topography localized at occipital regions (Feige et al. 2005), that involved superior temporal, orbital frontal, and cingulate cortices (Goldman et al. 2002), as well as frontal and parietal cortices (Laufs et al. 2006). The role of ASN and the precuneus in the process of alpha biofeedback was also confirmed (Ros et al. 2013).

[6] 73w The decreased interaction of precuneus network and default mode network, the decrease in connectivity between the cuneus and the primary visual cortex network are presumably based on a gradual change of the training strategy. Initially based on the instruction stating that "in most people alpha rhythm occurs in a calm and relaxed state, with nice memories", the subjects gradually moved from the positive emotional images visualization towards the logical constructs that received reinforcement.

[1] 160w In summary, the formation of neural networks that accompany and provide a voluntary increase of the EEG rhythms power appears to be a long multi-stage process where key participating networks and structures can be identified. The direction of the neural networks reorganization depends on the nature of the feedback "target", and it differs under initially identical instructions and strategies. The building of stable skills of increasing alpha rhythm power is followed by a reduction of irrelevant associations between the cerebellum and the VSN, the RECN and the cuneus network, and by a formation of a stable complex comprising interacting of the precuneus, the cuneus, the VSN, and high-level visual information network, as well as by intensifying the interaction of the anterior salience network with the precuneus network. The observed changes are presumably associated with a decrease in control and a change in the modes of visual memories processing to those more effectively providing an increase in the alpha rhythm power.

[2] 140w During training, the effect noted in the beta group was almost the opposite to one of the alpha group, including a weakened interaction between the precuneus network and the default mode network, and a decreased connectivity between the cuneus and the primary visual cortex network. That was probably related to the use of the cognitive strategies that contradicted the instructions, but provided more effective increase of beta rhythm power. The differences between the alpha group and the beta group consistently increased during training course. Alpha training was characterized by a less interaction of the primary visual cortex network with high-level visual processing network, and greater interaction of the cerebellum with the precuneus and the RECN. Thus, investigation of the neural networks formed during alpha training once again confirmed the key role of visual processing in the generation of alpha rhythm.

[3] 143w Another main result of our work is that we could demonstrate the differences in the structure and interaction of neural networks involved into alpha and beta generating systems forming and functioning, which inevitably should be taken into account when performing therapeutic interventions based on the EEG biofeedback. We consider the possibility of using fMRI-based biofeedback organized according to the described neural networks interaction and aimed to target specific symptoms as a very promising approach. For example, arranging neurofeedback training of activity in the ASN in anxiety cases, or in the RECN in states accompanied by a feeling of internal tension, etc. As one of the most promising directions we consider the identification of the EEG activation patterns of specific neural networks, since focusing on these new training targets allows combining the availability of the EEG biofeedback with its selective impact on neural networks.

[4] 105w As our experience showed, the analysis of neural networks interaction, specifically those that are the product of fMRI identification of synchronously operating distributed brain structures, allowed creating a new description of alpha rhythm "territory". The neural network vocabulary allows avoiding traditional topographic and anatomical description of functions borrowed from neuropathology, moving to the analysis of a direct connection of distributed networks and complex behavioral structures. Now the neural network is becoming a direct target of self-regulatory mechanism, where the object of control is the quantified phenomenon of functional connectivity that provides the basic properties of neuronal plasticity and the emergence (construction) of new behavioral patterns.

[1] 141w The study was carried out at the Research Institute of Molecular Biology and Biophysics and at the MRI Technology Laboratory of the International Tomography Center of the Siberian Branch of Russian Academy of Sciences. The recruited subjects included 23 (alpha rhythm training) and 12 (beta rhythm control) healthy right-handed subjects (all males, mean age 27 ± 7 years old), free of any neurological or psychiatric disorders history, with college/incomplete college education, and no prior biofeedback experience. All enrolled participants had completed the neurofeedback training course and most of them were able to increase the power of target EEG rhythm across training sessions. All subjects provided informed consent to participate in the research study. The study protocol was approved by the Ethics Committee of the Research Institute of Molecular Biology and Biophysics and was carried out in accordance with the Helsinki Declaration.

[2] 97w The experimental protocol included three concurrent fMRI-EEG registrations with the intervals of 2 weeks, during which subjects underwent an EEG biofeedback training course consisting of twenty 30-min sessions (five sessions per week), using the BOSLAB-Professional neurofeedback device (Comsib Ltd., Novosibirsk, Russia). Biofeedback sessions were conducted outside of the MRI scanner in the morning hours. During training subjects sat in a chair with a head support. Bipolar electrodes were placed at the standard leads (F3-O1, F4-O2, 10-20 system), and the feedback was arranged from the right hemisphere. The training was conducted with closed eyes and with auditory feedback.

[3] 148w Subjects from the control group received beta rhythm feedback, though they were not informed about which rhythm was the target of the self-regulation training, while the other parts of the protocol and instructions were identical to the alpha training group. Subjects in the beta feedback control group believed that they were training alpha rhythm as it was stated in the uniform instructions. The study used deception when instructing subjects in this control group that was necessary to support design of the project. This design was approved by Institute's Ethics Committee assuming that the protocol included post-treatment debriefing of the participants in this control group about the false instructions given to them and had explanation that this was to investigate whether the ability of the subjects to voluntarily change a EEG band power and corresponding fMRI signal changes are related to a feedback or to strategy based on instructions.

[4] 185w Enrollment to the groups was not randomized and this can be considered as a certain limitation of the design of this study. At the beginning of the study all volunteers were assigned only to the alpha training group, and when that group was complete then all new volunteers were assigned to the beta training group. On the basis of the lab EEGtraining outcomes (i.e., neurofeedback sessions outside of fMRI), for data analysis purposes subjects were assigned to one of three groups depending on their performance: The first group members ("successful", N = 19) successfully completed biofeedback training and demonstrated increase of the target rhythm power both in the first and second parts of the course (i.e., at least in 1 out of 10 in each half of the course), the second group ("medium success", N = 11) showed increase of the target rhythm power only in the first half of the biofeedback course, whereas the third group ("unsuccessful", N = 5) showed gradual decrease of the target rhythm power. Table 1 describes group allocation according to success rate in the neurofeedback training in the study.

[5] 147w For EEG analysis, the mean power of the target rhythm during individual training session was used. Subdivision was based on a hypothesis that the functional foundation for the increase of the power of target rhythm should be sought in the formation of the stable neurovascular bounds. For the aims of this research study we selected subjects that were successful on the both stages of the biofeedback training. In addition, due to absence of the second fMRI recording, we excluded one subject from the "successful" control group (beta biofeedback). The final sample used for the analysis of neural networks comprised of 13 subjects from the experimental alpha biofeedback group and 5 subjects from the control (beta) training group. Hereafter participants who performed successfully in both alpha (N = 13) and beta (N = 5) groups and had all required data collected will be referred to as "successful" performers.

[6] 176w The fMRI session consisted of six blocks, each included three conditions: "Rest (Eyes open)", "Test", and "Rest (Eyes closed)". The duration of "Eyes open" and "Eyes closed" condition was 35 s, while "Test" condition was 70 s long. "Rest (Eyes open)" referred to a calm resting state without any specific task, the same referred to "Rest (Eyes closed)" but with eyes closed. During "Test" condition the subjects were instructed to increase the power of the EEG alpha rhythm (which occurs in a state of quiet wakefulness in most people), using various cognitive strategies. At the first MRI test the participants focused on their own ideas of a calm, pleasant, relaxed state, and later-on their own experience of EEG alpha training outside the MRI scanner. The first four blocks were performed with auditory reinforcement (feedback on successful target rhythm increase), and the last two blocks were conducted without any feedback. The auditory feedback modality was selected since the eyes closed mode was used during EEG biofeedback. Subjects confirmed that they could hear the feedback sound signal well.

[7] 18w Functional MRI was performed on Achieva Nova Dual (Philips) MRI scanner with a 1.5 T magnetic field induction.

[8] 224w The basic operating T2*-weighted images were obtained using the SSH-EPI (Echo-planar imaging) method with a 64 × 64 × 35 matrix and a voxel (3D image element) size of 4 × 4 × 4 mm 3 , repetition time TR = 3500 ms, TE = 50 ms, 260 dynamic frames in a series. The reference anatomical image was obtained using the T1 TFE method with a 256 × 256 × 64 matrix and a voxel size of 1 × 1 × 2 mm 3 . Statistical processing of the results and fMRI images obtained was performed using MATLAB (MathWorks) software with SPM8 package. The processing included matching the relative positions of the frames, normalizing the images to a standard form (MNI space), and smoothing out using the Gaussian function with an isotropic 8 mm core. EEG was recorded simultaneously with fMRI scanning in a dark room with Brain Vision (Brain Products GmbH) electroencephalograph using an MR-compatible EEG cap BrainCap MR (EASYCAP), 128 electrodes, including the reference electrode, according to the extended 10-20 system (i.e., 10-10 system for 128 channel EEG). The electrode for electrocardiogram (ECG) recording was placed under the subject's shoulder blade (scapula). Before placing a participant in the MRI scanner, the level of the electrode impedance was kept< 20 kOhm. EEG processing was performed using Brain Vision Analyzer 2.1.1 (Brain Products GmbH).

[9] 140w After the automatic removal of the MR artifacts, EEG data were resampled at a sampling rate of 250 Hz and filtered with a bandwidth of 1-40 Hz. The cardioballistic artifacts were determined using a moving average over 21 periods. Manual removal of electrooculographic (EOG) and motor artifacts was performed using WinEEG 2.103.70 package (Mitsar). The EEG recording processing was divided into 2 s analysis epochs with a 50% overlap. To provide feedback, fast Fourier transformation was applied, in particular, the EEG spectral power in 8-13 Hz range was averaged at O1, O2, P3, and P4 leads where alpha rhythm amplitude characteristics are usually higher and are closer to maximum. The obtained value was compared with the individually set threshold (30% of time when the signal was above the threshold), and if it was exceeded, an auditory feedback signal was delivered.

[10] 249w ICA was used for reconstruction of the neural networks. The method involved decomposition of fMRI data according to a spatial criterion. Identification of components was performed using a software package GIFT 3.0.a that provides the possibility of group ICA studies. Estimation of the optimum number of components was carried out on the entire set of fMRI data using the minimum description length criterion. ICA was performed using the Infomax algorithm, and to compensate the stochasticity it was repeated 100 times. The individual dynamics of each component were reconstructed from the group data using the GICA reverse reconstruction procedure (GICA is back-reconstruction approach for group ICA) for each subject. Components that correlated well with the white matter and the cerebrospinal fluid, and poorly with the gray matter (correlation coefficient < 0.05), were excluded. For each of the remaining components their spatial coordinates within "standard brain" and their positions in relation to relevant anatomical structures were defined. Correlation with 14 known neural networks [according to the list of "individual networks" (Greicius 2017)] allowed for identification of the components' functional specialization. A voxel-by-voxel multiplication of the components' spatial maps and network maps was performed, and the correlation coefficients were calculated. Network coherence was estimated using the functional neural connectivity (FNC, MIALAB) toolbox to find timing correlations. The strength of functional relations and the time of shifting for each pair were calculated using the Lag-Shift algorithm. Group differences were assessed using Student's t test for independent or paired samples depending on a case.

[1] 126w 1. An increase of the differences was manifested as a change in the strength and the nature of neural networks interaction during biofeedback training. 2. A relatively smaller interaction between the primary visual cortex network and the high-level network of visual processing was noted during alpha training. Probably it was related to the hierarchy of processing visual memory stimuli of different origin, i.e., subjects' preferred reconstructing specific memories and images in order to generate a corresponding rhythm. 3. An interaction of the cerebellum with the precuneus and the RECN was significantly greater during alpha training, most likely due to the findings that the subjects in beta group followed the strategies that involved higher cognitive load (beta rhythm enhancement), but less attention to internal stimuli and sensations.

[2] 91w The absence of randomization in this study and unequal sample sizes of experimental and control groups (13 vs. 5 subjects) should be considered as a limitation that may affect some of above interpretations. In particular, absence of randomization could lead to the pre-training difference between groups in alpha and beta EEG power, brain networks intercorrelations, biofeedback learning abilities, and level of motivation. Further replications and randomized trials are needed to establish our findings. Analysis of fMRI data only in "successful" performers can be considered as yet another limitation of the study.