A different state of mind: neural activity related to volitionally up- versus downregulating cortical excitability

To date there exists no reliable method to non-invasively upregulate or downregulate the state of the resting motor system over a large dynamic range. Here we show that an operant conditioning paradigm which provides neurofeedback of the size of motor evoked potentials (MEPs) in response to transcranial magnetic stimulation (TMS), enables participants to self-modulate their own brain state. Following training, participants were able to robustly increase (by 83.8%) and decrease (by 30.6%) their MEP amplitudes. This volitional up-versus down-regulation of corticomotor excitability caused an increase of late-cortical disinhibition (LCD), a TMS derived read-out of presynaptic GABAB disinhibition, which was accompanied by an increase of gamma and a decrease of alpha oscillations in the trained hemisphere. This approach paves the way for future investigations into how altered brain state influences motor neurophysiology and recovery of function in a neurorehabilitation context.

To achieve this goal we developed a BCI by stimulating the cortex with TMS and 124 providing neurofeedback of MEP amplitudes (Fig. 1). The feedback was designed such 125 that participants were rewarded for larger than average MEPs in one condition, and 126 smaller than average in another condition. 127 (right FDI, ADM and OP, and left FDI). The circles were red if the root mean squared (rms) EMG at rest 131 was greater than 7 microvolts. It was essential that all four circles were green for at least 500ms before 132 the trial could proceed. When this condition was met a fixation cross appeared for a random period (in 133 order to prevent anticipation of the TMS pulse). During the fixation cross, it was still essential to keep 134 the background EMG below 7 microvolts in order for a TMS pulse to be delivered. (B) The peak-peak 135 amplitude of the motor evoked potential (MEP) evoked by the TMS was calculated in real-time and 136 displayed immediately to the participant on screen in the form of a rectangular bar. 137 (C) Different feedback for UP training and DOWN training. In the UP training If the MEP was greater 138 than the baseline mean, the rectangle was green, with a green tick, a dollar sign to indicate a small 139 financial reward, a display of the current score, and a positive encouraging sound bite was heard. If the 140 MEP did not meet the criterion amplitude, the bar was red, there was no dollar sign, and a negative sound 141 bite was heard. (D) A custom 3D printed 'coil spacer' device was used to prevent direct contact of the 142 TMS coil on the EEG electrodes and allow the pre-TMS EEG period to be recorded artefact free. Bidirectional changes in corticospinal excitability were observed in the MEP 167 neurofeedback group but not in a control group. 168 169 We first tested whether participants could learn to volitionally increase or decrease 170 (bidirectional) corticomotor excitability when using a motor imagery strategy shaped 171 by neurofeedback of MEP amplitudes. Across two training sessions, we found that 172 MEP amplitudes increased during UP training ( Fig. 2A,    As part of the initial training study (see Figure 2A, 'Ses 3' for the behavioural results), 314 we investigated whether the two different activity states evoked differential cortical 315 dynamics extracted from EEG recordings which were acquired simultaneously while 316 TMS was being performed to provide neurofeedback of MEP amplitude. As distinct 317 functions have been ascribed to 8 different sub-frequency bands across the known range 318 of brain signals (0.1 -80Hz), we now probed whether volitional changes in 319 corticospinal excitability of M1, drives neural activity measured in the delta (0.1-4Hz), 320 theta (5-7Hz), low alpha (8-10Hz), high alpha (11-13Hz), low beta (14-21Hz), high 321 beta (22-30Hz), low gamma (31-50Hz) and high gamma (51-80Hz) bands. Using the 322 portion of EEG data collected in the 1.5 seconds prior to each TMS pulse, we calculated 323 relative power in the UP and DOWN states for the eight frequency bands of interest. 324   predictive of larger MEP amplitudes, and higher amplitude oscillations in low gamma 367 (pFDR=0.020) and high gamma (pFDR=0.020) were significant predictors of larger MEP 368 amplitudes. In a previous study, it was reported that a strong predictor of cortical 369 excitability was the low gamma : high alpha ratio 3 . We replicated this finding, 370 demonstrating that this ratio was a significant predictor of MEP amplitude (pFDR=0.016) 371 with larger ratios predicting larger MEP amplitudes. 372

373
EEG data classification 374 375 We next tested whether the distinction between the two trained states was large enough 376 that the individual data trials could be successfully predicted as 'UP' state or 'DOWN' 377 state, using machine learning based solely on the EEG power values (relative power 378 data, scaled by 1/f transformation, in the 1.5s prior to TMS) of the 8 frequency bands 379 of interest. A linear support vector machine (SVM) was applied to the data of each 380 participant (60 UP 60 DOWN epochs). The SVM has been shown to be particularly 381 powerful on EEG data, which is noisy and contains many features that are correlated. 382 This approach additionally allowed us to perform feature selection, to quantify which 383 EEG features most heavily contributed to the distinction between the two states. Using 384 only data from the electrode closest to the hotspot (8 rhythms plus 385 LowGamma:HighAlpha ratio) the SVM was capable of classifying the brain states with 386 an average accuracy of 81.5% (±5.1%) based on 10-fold cross validation which differed 387  The UP state was associated with a significant increase of LCD while other 446 measurements probing inhibitory M1 circuits failed to reveal differential effects for the 447 UP versus DOWN state. LCD is thought to represent a read-out of the presynaptic self-448 inhibition of GABAergic neurons which is thought to be mediated by extrasynaptic 449 GABAB auto-receptors 34,35 . This mechanism is hypothesized to result in a net 450 facilitatory effect as observed during the UP condition in our study. Previously, LCD 451 was found to be elevated during motor imagery (MI), but this increase relative to rest 452 was observed irrespective of whether participants imagined voluntarily activating or 453 relaxing hand muscles 36 . However, this investigation was conducted in a single 454 session, and did not employ neurofeedback, so MEP modulation by these two 455 imagination conditions could be expected to be substantially smaller than observed in 456 our study, particularly for the voluntary relaxation condition which had a similar 457 excitability state as the rest condition. Thus, it is possible that the clear modulation of 458 LCD observed here only manifested after neurofeedback training, i.e. when the two 459 excitability states became clearly distinct. It is important to note here that in our 460 protocol, LCD was detected only during the UP state and not at rest. In our search 461 procedure to decide upon the optimum conditioning stimulus (CS) intensities, we 462 prioritized SICI and LICI, finding a CS intensity that elicited as close to 50% inhibition 463 of the test MEP as possible. We tested intensities between 106-114% RMT for LICI 464 (and above or below this if no appropriate inhibition was found), and applied these also 465 to LCD (such that the only difference between the LCD and LICI protocols was the 466 ISI). This may have simply been too low to consistently elicit LCD, which appears to 467 be more robustly evoked at higher intensities 27,37 or during mild contraction 38 , and is 468 not observed in all individuals 37 . It may also be that LCD is more readily observed 469 using triple pulse procedures in which disinhibition can more directly be measured by suggests that repeated bursts of inhibitory alpha activity serve to temporarily silence 487 gamma oscillations 1 . Thus, these two rhythms are seen to exhibit a reciprocal 488 relationship, whereby when alpha is high, gamma is low. In periods of high alpha, 489 gamma may still burst periodically, but only at the troughs of the oscillation cycle, 490 meaning that the gamma 'duty cycle' (window for neural processing) is short, and only 491 brief messages can be sent. By contrast, in periods of low alpha power, the gamma duty 492 cycle is longer, and more extensive neuronal processing and inter-regional 493 communication may occur. Our finding of increased gamma activity is also consistent 494 with previous animal literature, showing that the pharmacological removal of GABAB-495 mediated inhibition (by receptor blockage) in rats results in increased gamma 496 oscillations 40 which have been shown to be largest in M1's layer V 41 .

498
Gamma has often been considered difficult to detect using scalp electrodes 499 because it is highly localised 42 and may also reflect non-cortical sources when recorded 500 with EEG 43,44 . However, it is tempting to speculate that, in our experiment, gamma 501 activity was strongly synchronized as a consequence of the neurofeedback training, 502 where participants learned to substantially facilitate corticomotor excitability while 503 keeping EMG activity constant, such that EMG amplitude differed only minimally 504 between the UP and DOWN conditions. This suggestion is in line with previous 505 neurofeedback studies that provided direct feedback of gamma activity, showing that 506 gamma power could be upregulated to a substantial amount which even exceeded 507 power values observed during movement execution 15,45 . By keeping the visual feedback 508 for the two conditions identical, we ensured that differences in eye movements between 509 the UP and DOWN states were minor. As we were particularly interested in gamma 510 oscillations, we additionally performed all EEG recordings in an electromagnetically 511 shielded room, using a gel-based electrode system to maximize signal to noise ratio. 512 513 Previous studies have taken a correlational approach to investigating the 514 relationships between brain rhythms and corticomotor excitability. These have shown 515 that low alpha 4,46 or beta power 47 as well as high gamma power 3 during natural 516 fluctuations at rest are associated with larger MEP amplitudes. We confirm and extend 517 these results by introducing causality to this relationship for the first time, showing that 518 experimentally driving excitability into two distinct states causes specific patterns of 519 neural dynamics in the volitionally controlled cortical area. 520 521 While changes in alpha and gamma were specific to the hemisphere from which 522 feedback was provided, theta showed a bilateral pattern of modulation, being higher in 523 the DOWN than the UP state in motor areas in both hemispheres. While mid-frontal 524 theta activity has been linked to error monitoring 48 the role of lateralized theta activity 525 close to the sensorimotor hotspot electrode and its symmetric counterpart is less clear. 526 Slower rhythms exert effects over larger distances, and are thought to be involved in 527 long-range communication 42 . A similar pattern of upregulation and downregulation 528 was observed in the homologous muscle in the opposite limb, albeit weaker and not 529 statistically significant. This is likely a reflection of the extensive transcallosal 530 structural connectivity and functional coupling of homologous regions of the cortical 531 motor network 49-51 . It is tempting to speculate that the bilateral theta activity observed 532 in the current study served to regulate the inhibition/facilitation of functional coupling 533 or 'spillover' of activation from motor areas in the target hemisphere to their 534 homologous counterparts. 535 536 537 Surprisingly we did not observe differential modulation of the Beta band, which 538 is the predominant oscillatory frequency in sensorimotor cortical regions 52,53 . It 539 typically desynchronizes (together with alpha) during motor execution and motor 540 imagery 54-57 and has been associated with corticomotor excitability at rest 3 . As our 541 results represent the direct contrast between the UP and DOWN states, the lack of Beta 542 involvement may firstly be due to the fact that both conditions involved a mental 543 strategy targeted at the sensorimotor system and, secondly, that no temporal structure 544 was imposed so that we could not perform analyses which are, for example, time-locked 545 to the potential onset of these mental strategies. However, our data further confirm that 546 the two 'inhibitory' rhythms alpha and beta might serve different functions in selecting 547 and activating the appropriate sensorimotor representations 58 . 548 549 550 Here we present an innovative approach to voluntarily and bidirectionally 551 change the state of the motor cortex, by directly targeting MEP amplitudes in a 552 neurofeedback paradigm. This method provided a unique opportunity to reveal the 553 oscillatory and neurochemical underpinnings of the two distinct trained brain states, 554 using concurrent TMS EEG measurements, and mechanistic follow-up investigations 555 using paired-pulse TMS. The results comparing UP and DOWN states indicate that 556 voluntary upregulation of corticomotor excitability causes increased presynaptic 557 GABAB-mediated disinhibition, elevated neural oscillations in the gamma frequency 558 range, and reduced alpha and theta rhythms. 559 This paves the way for new technologies that allow the user to regulate aspects 560 of their own brain function in order to reach desired states that are, for example, 561 associated with enhanced motor performance. In the context of stroke rehabilitation, 562 training volitional modulation of corticomotor excitability may hold promise as a 563 rehabilitative therapy early after stroke, i.e. when patients are deprived of rehabilitation 564 training because they are unable to execute overt movements with the impaired upper 565 limb. As it is known that LCD is recruited during actual movement 28,59,60 , the elevated 566 LCD we observed in the UP condition may reflect that the neurofeedback had engaged 567 similar mechanisms to those involved in movement execution, using only voluntary 568 endogenous processes. Furthermore, as pathological hyperexcitability of the non-569 damaged hemisphere has been hypothesized to limit recovery in some patients 61  MEPs were larger than baseline (the 'UP' condition) and on the other two days, the 598 rewarding stimulus was displayed when MEPs were smaller than baseline (the 599 'DOWN' condition). On each of the training days, 4 separate blocks of neurofeedback 600 were preformed, each comprising of 30 individual MEP feedback trials (total 120 trials 601 per day). The format of individual trials and feedback is described in more detail below. The stimulation intensity used to evoke MEPs during the experiment was chosen 622 using the following procedure in order to obtain a baseline MEP amplitude that was 623 50% of the participant's maximum. A recruitment curve eg. 63 was performed at the 624 beginning of the first experimental session, whereby 6 TMS pulses were applied at 10 625 different intensities relative to RMT (90%, 100%, 110%, 120%, 130%, 140%, 150%, 626 160%, 180%, 190%) in a randomized order. MEP amplitude at each intensity was 627 plotted to determine the point on the curve at which plateau occurs and the MEPs do 628 not continue to increase. Maximal MEP amplitude was recorded, and the intensity 629 required to evoke 50% of this amplitude was used for all subsequent testing. With this 630 approach, there was scope for MEP amplitude to both increase and decrease to similar 631 extents from this 'intermediate' value. Post-hoc analyses revealed that this procedure 632 resulted in an average stimulation intensity corresponding to 130% RMT. Immediately 633 following this procedure and prior to the first block of neurofeedback, 20 MEPs were 634 collected at rest at the chosen intensity to determine 'baseline' corticospinal 635 excitability. The mean MEP amplitude at baseline was recorded and used during 636 neurofeedback to establish the criterion amplitude that determined whether participants 637 received either positive or negative feedback. 638

Format of neurofeedback 639
Neurofeedback was performed using custom written MATLAB software. 640 Participants kept eyes open with attention directed to a monitor in front of them. They 641 were instructed to relax their limbs and avoid tensing any muscles throughout the 642 experiment. In order to ensure that MEP amplitude could not be influenced by 643 background muscle activation, the root mean square (rms) of the EMG signal for each 644 muscle for the previous 100ms of data was calculated and displayed in real-time on 645 screen at the beginning of each trial in the form of four coloured 'traffic lights', 646 representing each muscle (Fig. 1A). If the background EMG in a muscle exceeded 7µV, 647 the corresponding light turned red. Participants were instructed that a trial could not 648 begin unless all four lights were green (all muscles relaxed below 7µV) for at least a 649 continuous 500ms period. When a trial commenced the traffic lights disappeared, but 650 background EMG continued to be monitored and the trial was automatically paused if 651 any muscle exceeded the threshold. At the beginning of each trial a fixation cross 652 appeared in the center of the screen. After a variable period of time (between 5.5 -8.5 653 seconds, or longer if muscle activation delayed the trial) a TMS pulse was fired. The 654 MEP amplitude for the target muscle (right FDI) was immediately measured and 655 displayed to the participant on screen within 500ms. The display consisted of a vertical 656 bar indicating MEP amplitude relative to a horizontal line in the middle of the screen 657 representing the mean recorded at baseline (Fig. 1B). In 'UP' sessions if the MEP was 658 larger than the criterion amplitude, the bar was shown as green with a tick beside it, a 659 positive soundbyte was heard, and a number adjacent to a dollar sign incremented to 660 indicate that a small financial reward had been gained. If the MEP was smaller than the 661 criterion amplitude, the bar was red with a cross beside it, a negative soundbyte was 662 heard, and no financial reward was shown. The reverse was true in the 'DOWN' 663 sessions (Fig. 1C). The feedback remained on screen for 4 seconds, before being 664 replaced by the traffic lights display preceding the next trial. Participants were 665 instructed to attend to the feedback and that the goal was to increase (or decrease) the to minimize artefacts associated with the direct contact of the TMS coil resting on the 684 electrodes of the EEG cap, we designed and 3D-printed a custom plastic 'coil spacer' 685 device 64 , which has four wide legs positioned to provide a platform distributing the 686 weight of the TMS coil, so that it hovers over the electrodes without contact (Fig. 1D). 687 This allowed quality recordings to be obtained even from the channel of interest closest 688 to the participant's 'hotspot'. The participants RMT was established while wearing the 689 EEG cap with TMS coil spacer, and the same % above threshold that was used for all 690 previous sessions was applied for neurofeedback. Impedances were monitored 691 throughout and maintained below 50kΩ. EMG data from all four hand muscles were band-pass filtered (30-800 Hz), 722 separately for the portion of data containing the 100ms of 'pre-TMS' background EMG, 723 and for the portion of EMG containing the MEP, in order to prevent smearing of the 724 MEP into the background EMG data chunk. The root mean squared (rms) of the 725 background EMG was calculated, and peak-peak MEP amplitude was measured. 726 Trimming (removal) of the maximum and minimum MEP in each block was performed 727 in order to screen out extreme values. MEP amplitude is known to be modulated by 728 EMG background activation 23,24 . Therefore, the rms pre-stimulus EMG recordings 729 were used to assess the presence of unwanted background EMG activity in the period 730 110 to 10ms preceding the magnetic pulse. MEPs preceded by background EMG higher 731 than 0.01mV were excluded. For each subject and over all trials we calculated the mean 732 and standard deviation of the background EMG. MEPs that occurred when the 733 background EMG value exceeded 2.5 standard deviations above the mean, and MEPs 734 with a peak-to-peak amplitude which exceeded Q3 + 1.5 x (Q3 -Q1) were removed 735 from further analysis, with Q1 denoting the first quartile and Q3 the third quartile 736 computed over the whole set of trials for each subject. Signals from all 64 channels were first epoched to extract only the data on each 757 trial from the 4 seconds before the TMS pulse. This was to remove the substantial 758 artefacts that arise during the magnetic pulses, prior to conducting any filtering or 759 further processing. These separate chunks of unpolluted data were then concatenated 760 into one continuous epoch, and highpass filtered at 1Hz, prior to conducting an 761 independent components analysis (ICA). Independent components were visualized and 762 those containing artefacts arising from eye movements, facial EMG, cardiac signals, 763 bad channels or other non-brain activity related signals were removed. 764 The cleaned data were average-referenced, and re-epoched into chunks of data 765 containing only the 1.5s on each trial prior to the TMS pulse (ie. to capture the ongoing 766 oscillatory activity at the instance in which the TMS occurred, while the fixation cross 767 was on screen and the 'traffic lights' had disappeared). 768 A power spectrum was computed (Welch periodogram/FFT) for each single epoch and 769 the mean power (and relative power) in each of the relevant bandwidths were extracted 770 (delta (0.1-4Hz), theta (5-7Hz), low alpha (8-10Hz), high alpha (11-13Hz), low beta 771 (14-21Hz), high beta (22-30Hz), low gamma (31-50Hz) and high gamma (51-80Hz). 772 Power values were computed separately for UP and DOWN trials, and non-parametric 773 Wilcoxon signed rank tests (with FDR correction) were used to compare neural 774 oscillations in these two states. 775

776
We also analysed whether trial-by-trial variation of EEG data was associated with trial-777 by-trial variation of MEP amplitudes. Therefore, relative power in each bandwidth for 778 each epoch was entered into a multiple regression model with MEP amplitudes 779 measured in the muscle from which neurofeedback was provided (right FDI A sub-set of 11 participants from the experimental group returned approximately 6 800 months later to participate in a follow-up experiment probing retention and mechanisms 801 underlying the two distinct states. This was conducted over a further 4 days of testing. 802 On one day, retention, aftereffects, and excitability in the opposite 'untrained' 803 hemisphere were tested for the 'UP' condition. On another, neurophysiolocial 804 mechanisms were probed using paired pulse TMS. These two days were repeated for 805 the 'DOWN' condition, and the order of these sessions was counterbalanced. We 806 additionally tested whether trained participants were able to upregulate and 807 downregulate when feedback was temporarily removed. 808 809

Retention testing & aftereffects measurement 810
After a 6-month break and no top-up training, participants were tested with one block 811 of TMS-neurofeedback (20 MEPs) in order to assess retention of learning. All other 812 procedures were identical to those carried out in the main experiment. 813 Following this block, 12 MEPs were collected at rest after 5 and 10 minutes. 814 815 Excitability in the opposite hemisphere 816 817 During one block, two TMS coils were used, placed over the right and left motor 818 hotspots (as described previously). This block contained 40 trials, 20 of which were 819 normal TMS neurofeedback trials. The other 20 were trials where TMS was applied to 820 the opposite hemisphere, rather than to the hemisphere that was the target for 821 neurofeedback. No feedback was given in these trials. intensity was identical to that which had been chosen for the TMS neurofeedback (ie. 842 that produced MEPs that were 50% of the maximum on the recruitment curve). On 25% 843 of trials Short Interval Intracortical Inhibition (SICI) was measured. This was with a 844 conditioning stimulus intensity that was chosen using a personalized search procedure 845 testing intensities ranging from 50%-90% RMT, to achieve as close to 50% inhibition 846 as possible, and an inter-stimulus interval of 1.97ms 67 . The reduction in the size of the 847 test MEP is believed to represent postsynaptic GABAA inhibition 25 . On 25% of trials 848 Long Interval Intracortical Inhibition (LICI) was measured. This was with two 849 suprathreshold pulses, with the conditioning stimulus intensity chosen using a search 850 procedure between 106-114% RMT, and an inter-stimulus interval of 100ms 27 . This is 851 believed to reflect postsynaptic GABAB inhibition 68 . On the remaining 25% of trials, 852 Late Cortical Disinhibition (LCD) was tested. This was with the exact same pulse 853 intensities as used for LICI, but with a 220ms inter-stimulus interval 27 , and is thought 854 to measure presynaptic GABAB disinhibition 26-28 . The order of presentation of paired 855 pulses and single pulses was randomized.