We recruited 18 healthy adult participants, including eight females [mean age = 27.33 years (SD = 5.28)]. Participants did not reveal a history of neurological or psychiatric illnesses (i.e., self-reported). All participants indicated normal hearing and reported right-handedness, as confirmed by the Edinburgh handedness inventory (Oldfield, 1971). Participants were also screened to ensure they met TMS safety standards (Rossi et al., 2009, 2011). Participants provided informed written consent in accordance with the Declaration of Helsinki. Participants were compensated for their time. The research was approved by the Deakin University Human Research Ethics Committee (2015-034).

The experimental paradigm was based on an ERP investigation (Ford et al., 2010) and a TMS motor-learning study (Ticini et al., 2011). It consisted of two main stages: (1) Action and (2) Audition. During the action stage, participants produced sounds by making finger-swipes on the experimental device, while EEG and EMG techniques recorded CNS activity. The audition stage, however, used TMS and EMG to record CNS activity when participants passively listened to sounds that were played back via the device. These stages were also comprised of individual blocks to help minimize issues with participant attention waning (e.g., Finkbeiner et al., 2016). Overall, participants produced sounds (i.e., action) or heard them (i.e., audition) while EEG, EMG, and TMS recorded CNS activity within separate experimental blocks (see Figure 1 for illustration of the protocol).

Participants sat comfortably in a chair. The custom-made device, labeled the AMRJ (outlined below and illustrated in Figure 2), was placed in front of them. After each neuroscientific technique setup was completed (described below), the experiment began with the Baseline stage. This stage consisted of two blocks (B1 and B2). These blocks and the final Baseline block (B3) were designed to measure transient-state influences (Schmidt et al., 2009) and potential cumulative effects of single-pulse TMS (Pellicciari et al., 2016). Each block had 20 TMS pulses with a 4-s inter-stimulus interval and 5-s inter-block interval. Each block lasted approximately 2 min.

Following a 1-min break, the Pre-learning stage (Pre-LP) began. The Pre-LP stage was comprised of two blocks (so-called Pre-LP 1 and Pre-LP 2). Each block established a baseline of MEPs and, therefore, AMR. Participants listened to a quasi-randomized block of 48 sound samples (i.e., 48 trials in a block; sounds are described below). During listening, TMS was applied to M1, and MEPs were recorded from both the index and little finger muscles simultaneously. Each block lasted approximately 5 min, and a 1-min break between blocks was provided.

After another 1-min break, the Learning Procedure (LP) stage began. This stage was intended to associate finger-swipes and sounds, and participants produced the sounds via finger-swipes. There were four LP blocks (LP 1, LP 2, LP 3, and LP 4). Each block required participants to perform 48 swipe movements with their index or little fingers across the corresponding switch, which would generate the sounds. Beginning with the inside switch-edge, each finger would move towards the outer edge of the device (see Figure 3 for illustration). The index finger-swipe was toward the left-hand side of the device, while the little finger generated a swipe towards the right. The starting position required the middle and ring fingers to be positioned over the two home plates. This allowed the index and little fingers to be aligned with the inside edge of the respective switch. Before testing, the experimenter provided an example of both finger-swipes. Each swipe needed to be at least 300 ms in duration for sound playback to occur. Participants were instructed to modulate their speed to ensure swipe time was a minimum of 300 ms.

Once sound playback had finished after each swipe, participants returned to the original starting position. At the beginning of each LP block, participants were instructed to begin another swipe only after a self-timed 3-s break had expired. This break was designed to assuage the concern of fatigue during sensorimotor learning tasks (Bock et al., 2005; McDonnell and Ridding, 2006). If the experimenter observed that the participant began a swipe before the 3-s break had expired, participants were notified via a shoulder tap before the next swipe. This indicated they were to increase the rest period between trials.

To help promote motor learning, participants voluntarily chose which movement to execute (Herwig et al., 2007). Participants were asked to perform an approximately equivalent number of index and little finger-swipes (24 each) to mitigate any potential learning biases. This was observed, and participants were made aware of the swipe distribution-count between LP blocks. During testing, 50.51% of swipes were generated with the index finger while 49.49% were produced using the little finger. Within an LP block, the minimum index swipe count was 18, and for the little finger at least 20 swipes were produced. Each block lasted approximately 5-min, and a 1-min break between each LP block was used.

Following a 4-min break, the next stage was the post-learning stage named Post-LP. Like the Pre-LP blocks, these stages required listening to the 48 sound-trials over separate blocks (Post-LP 1, Post-LP 2, Post-LP 3, and Post-LP 4). Post-LP 1 and 2 blocks used TMS, while Post-LP 3 and 4 used EEG separately. TMS and EEG were used independently to minimize interference across recordings. A 1-min break between all blocks was added.

Participants then completed the Learning Procedure-control (LP-C) stage. Here, two LP-C blocks (LP-C 1 and LP-C 2) were used to isolate the motor component within the ERP trace. Convention suggests that data associated with movement should be subtracted from the LP block ERPs. This is thought to improve the comparison with listening-derived ERPs (Martikainen et al., 2005; Ford et al., 2007, 2010, 2014; Baess et al., 2011; Van Elk et al., 2014). Therefore, the LP-C blocks used the same overall design as the LP but did not produce any sound following a swipe action. That is, swipe movements were produced; however, no sounds were played. During testing, 50.77% of swipes were generated with the index finger while 49.23% were produced using the little finger.

Lastly, another Baseline stage was completed (block B3). Like the initial Baseline blocks, four sets of five TMS pulses at the motor threshold (MT) were applied while MEPs were recorded.

Participants were asked to observe their right hand throughout the experiment to ensure a degree of uniformity across experimental stages. Throughout all listening blocks, participants were asked to pay attention to the sounds and indicate if they heard a control sound after sound playback has finished (i.e., no control sound was used during listening trials, however). In total, the experiment lasted approximately 90–120 min.

Manufactured by SPLat Controls (SPLat Controls, Seaford, VIC, Australia) and Maximum Design (Max Designs, Croydon North, VIC, Australia), the AMRJ consists of the MS121USB216 controller and MP3 Trigger printed circuit board. Intended for our experimental protocol, the AMRJ uses finger movement to activate inbuilt switches. There are two switches; one for the index finger and another for the little finger. These detect changes in electrical capacitance upon touch. They are capacitance switches and, therefore, do not require a force to activate. Mechanical button presses can produce unwanted sounds and even accentuate tactile information, which can affect ERP recordings (Horváth, 2014). The use of capacitance switches reduces the potential confound of tactile and other sound feedback on ERP recordings.

The device can also bypass switch activation and play a quasi-randomized sequence of the audio samples. In either mode, the AMRJ triggers EEG, EMG, and TMS equipment to help record data when needs require (see Figure 2 for illustration of the AMRJ, while more details regarding the switches and trigger design can be found in the Supplementary Materials).

Swipe movements were followed by one (of two) complex sounds at a stimulus onset asynchrony (SOA) of 10 ms. This delay would ensure all techniques were triggered simultaneously (where required). The sounds were recorded on Ableton 9.0 software (Ableton AG, Mitte, BER, Germany) using a Roland Juno 60 synthesizer (Roland Corporation, Hamamatsu, 22, Japan) at Otologic Studios (Toorak, VIC, Australia). One sound consisted of an approximate 500 Hz fundamental tone, as well as 250 Hz and 1,000 Hz overtones. This sound was heard as a low sound. The other sound, so-heard as the high sound, was comprised of an approximate 1,800 Hz fundamental tone, as well as 2,100 Hz and 4,000 Hz overtones (for sound spectrograms, see Supplementary Figure S2 in Supplementary Materials).

Assignment of sounds (low or high) to each switch (index or little finger) were counterbalanced across participants (Ticini et al., 2011). For example, some participants produced the high sound by an index finger-swipe, while others produced the low sound (by that finger-swipe). For statistical purposes, however, we refer to the MEPs recorded in the index finger as the congruent sound or sound associated with the index finger as the Congruent_FDI data. Since these sounds were pinned for each participant, the Congruent_FDI sound then becomes the incongruent sound for the little finger. Therefore, the Incongruent_ADM term refers to the MEPs obtained in the little finger during playback of the index finger sound. Similarly, the Congruent_ADM sound describes the MEPs obtained during playback of the little finger-congruent sound. Therefore, this sound becomes the index finger incongruent sound, and reflects the MEPs recorded in the index finger during the little finger-swipe sound playback (i.e., Incongruent_FDI).

Sounds were played through Etymotic ER3-10 ABR insert earphones (Compumedics USA, Charlotte, NC, USA), via the AMRJ 3.5 mm stereo port, and amplified to 75 dB or slightly lower if individual comfort levels were exceeded (i.e., sound levels were determined via the inbuilt MP3 Trigger printed circuit board).

During the EEG blocks, data were recorded using 12 Ag-AgCl sintered electrodes. EEG electrode sites comprised Fz, FCz, Pz, P3, and P4. Electrodes were also placed on both mastoids, and a ground electrode was placed on the forehead for off-line referencing. Both vertical and horizontal electrooculograms (EOGs) were recorded using electrodes above and below the left eye, and on the outer canthus of each eye. EEG data were obtained via a SynAmps RT system (Compumedics USA, Charlotte, NC, USA) and Curry 7.07xs (Compumedics USA, Charlotte, NC, USA). Data were sampled at 10 kHz, and impedance was kept below 5 kΩ for each electrode.

EEG and EOG data were analyzed offline in Curry 7.07xs (Compumedics USA, Charlotte, NC, USA). A 50 Hz notch filter and a band-pass filter between 0.5 and 15 Hz was applied. Data were re-referenced to the mean combination of the left and right mastoids. To minimize the influence of eye blinks on the ERP, horizontal and vertical EOG data were corrected using Curry’s covariate analysis tool.

Bad EEG periods exceeding ±75 μV in amplitude were detected. The 700 ms epochs around these (i.e., 200 ms prior and 500 ms after trigger) were removed from further analysis to obtain conservative EEG epochs. Epochs that exceeded a signal-to-noise ratio below 0.5 and above 2.5 in a 700 ms window time-locked to triggers were also excluded from further analysis. From a possible 7114 blocks, 21.9% or 1,556 were removed.

For each participant, epochs were labeled for Sound (Congruent_FDI or Congruent_ADM), Block [(action or audition) LP 1, LP 2, LP 3, LP 4, Post-LP 3, Post-LP 4], and LP-C blocks (LP-C 1 and LP-C 2). Subsequent analyses of the epochs investigated the peak amplitude of the N1, P2, and N2 ERP components. These were respectively examined within an epoch of 50–150 ms (N1), 100–200 ms (P2), and 151–250 ms (N2) for each participant across each ERP recording.

During TMS blocks, focal TMS pulses were delivered to the scalp over the left M1. A 70 mm figure-of-eight stimulation coil was used (Magstim Company, Whitland, UK), and the coil was connected to a Magstim 200 stimulator (Magstim Company, Whitland, UK). Using self-adhesive Ag-AgCl electrodes, TMS-induced MEPs were recorded from first dorsal interosseous (FDI) muscle and abductor minimi digiti muscle (ADM) muscles simultaneously. A ground electrode was placed on the dorsal surface of the wrist (i.e., ulna bone). The EMG signal was amplified by a PowerLab/4SP (ADInstruments, Colorado Springs, CO, USA), and data were sampled via a FE135 Dual Bio-amp (ADInstruments, Colorado Springs, CO, USA). A band-pass filter between 0.3 and 1,000 Hz was applied, as well as a mains 50 Hz notch filter.

The site on the scalp that produced the largest median (peak-to-peak) MEP in five consecutive trials from the right-FDI while at rest was defined as M1. Stimulation to M1 determined the MT. MT was defined here as the stimulation intensity that evoked a median peak-to-peak MEP of approximately 1 mV. The acceptable range was between 0.8 and 1.3 mV. Once a suitable stimulator output had been determined (e.g., indicating MEPs from M1 were within the 0.8–1.3 mV range), recordings were obtained from 10 trials using the FDI muscle while the hand was at rest. If the median MT MEP was not found to fall within the accepted range across those 10 trials, the MT process began again. The stimulator output ranged from 33% to 60% [M = 46% (SD = 7.6%)] of the maximum stimulator output. We also used the Baseline blocks (B1, B2, and B3) to ensure no significant changes in background corticospinal excitability occurred. Corticospinal excitability across the paradigm was stable as measured via a mean of TMS Baseline blocks (see Supplementary Materials for test details).

Due to the uniqueness and duration of swipe movements, we opted for a broad range of trigger-points to obtain MEPs during listening. We were unsure where peak AMR would be recorded during listening. It was foreseeable the sensorimotor association might involve an internal mapping that encodes the sound with either movement commencement, movement transition, swipe termination, or a variation of each (for related discussion, see Horváth, 2013). We could not, however, determine swipe time precisely, which might help calculate a suitable AMR time-window or trigger point for participants. For example, an individual might produce slower swipes than another, which suggests a longer TMS trigger latency is appropriate to assess the sensorimotor association. Thus, a variety of static TMS-trigger time points were used to overcome this issue.

Focal-TMS pulses were delivered to M1 at 50, 150, 300, or 450 ms from SOA in each of the 48 trials (i.e., 12 pulses at each 50, 150, 300, or 450 ms from SOA). While these triggers do not consider individual variability, we considered the AMR time-window suitable for swipes 300+ ms in duration. Additionally, we considered the time points helpful in exploring some basics regarding timing during the association process. Potentially, that is, how the brain overcomes sensory delays while integrating a present sound with a past action (for discussion on this point, see Hanuschkin et al., 2013; Giret et al., 2014; Keysers and Gazzola, 2014; Burgess et al., 2017).

As is recommended for MEP data (Schmidt et al., 2009), individual median peak-to-peak MEP amplitudes (mV) were extracted for each TMS block (Pre-LP 1, Pre-LP 2, Post-LP 1, Post-LP 2, B1, B2, and B3) across Muscle (FDI or ADM), Sound (Congruent_FDI or Congruent_ADM), and Time point (50, 150, 300, and 450 ms). Approximately 252 MEPs were obtained for each participant. Missing data points in baseline blocks, due to faulty TMS-based triggers, were replaced by the median MEPs of the remaining blocks for those participants (this required the alteration of two data points or 3.7% of Baseline MEPs collected across all participants). Also, one participant’s Post-LP 1 block was corrupt, requiring the (presumably comparable) Post-LP 2 dataset to be used (i.e., alteration of 1.4% of total MEPs averaged across all participants).

To minimize the influence on tests of normality, extreme outliers at a sample level were reduced to one value above the next highest data point (Tabachnick and Fidell, 2006). From the FDI muscle, 1.2% of MEP recordings were altered, while 1.9% of the ADM raw data. Tests of normality, histograms, as well as stem and leaf plots, were inspected and were considered satisfactory for parametric data analyses.

First, we investigated changes in sensory prediction. Figure 3 displays the grand-average ERP traces from SOA. Regarding the N1 peak, a three-way interaction between Finger, Stage, and Electrode was revealed (F_(1,17) = 6.031, p = 0.025, ηp2 = 0.262) requiring further investigation (see Supplementary Materials). However, no main effects were present, suggesting action and audition peaks did not differ within this time window (50–150 ms).

Regarding the P2 peak, ANOVA revealed a main effect for Stage as hypothesized (F_(1,17) = 4.471, p = 0.050, ηp2 = 0.208). Post hoc PCs, which were Bonferroni corrected, indicated that the action stage produced larger P2 peaks (M = 2.704, SE = 0.653) than the audition stage (M = 1.438, SE = 0.486). This suggests that some sensory prediction outcomes during action are being reflected in this peak.

As hypothesized, a main effect for Stage was demonstrated with the N2 peak data (F_(1,17) = 15.296, p = 0.001, ηp2 = 0.474). PCs (Bonferroni corrected) revealed audition generated larger (more negative) peaks (M = −9.354, SE = 0.738) in comparison to action (M = −6.676, SE = 0.435). This suggests that a predictive process is being undertaken during action which suppresses the incoming sounds.

Altogether, we found evidence for some sensory prediction outcomes during finger-swipe movements.

Next, we determined if AMR developed (i.e., sensorimotor associations). Comparisons between Pre and Post-LP MEPs, which were recorded upon hearing congruent and incongruent sounds using the static time points, were explored.

The four-way ANOVA revealed a main effect for Time point (F_(3,51) = 4.809, p = 0.005, ηp2 = 0.220). PCs (Bonferroni corrected) indicated the MEPs at the 50 ms time point (M = 0.681, SE = 0.060) were significantly larger than the 150 ms time point (M = 0.612, SE = 0.051; p = 0.003). This suggests motor-brain activity is high during the early stages of swipe-sound listening.

ANOVA also revealed a main effect for Audition blocks (F_(3,51) = 3.799, p = 0.016, ηp2 = 0.183). Although PCs did not survive Bonferroni corrections, estimated means indicated that Post-LP 1 MEPs were, unexpectedly, the smallest recorded [Pre-LP 1: M = 0.708 (SE = 0.076); Pre-LP 2: M = 0.704 (SE = 0.068); Post-LP 1: M = 0.538 (SE = 0.063); Post-LP 2: M = 0.645, SE = 0.060]. When both congruent and incongruent MEPs are examined, the reduction in MEP size immediately post-training is in direct contrast with our expectations. Indeed, it will be difficult to show AMR across pre- and post-training comparisons if, overall, post-training MEPs are reduced when compared with baseline measurements (see Supplementary Materials for discussion of repetition suppression during the LP blocks, which might explain this unanticipated finding).

Regarding AMR illustration, no Audition block × Sound × Time point interactions were present. This suggests that AMR did not develop. Trained sounds did not increase finger-corticospinal networks beyond baseline measures when all blocks and time points are considered.

However, we were concerned with the use of static TMS trigger points, which do not consider individual variability and learning. We suspected these triggers might be censoring the AMR illustration. If sensorimotor associations are experience dependent, and a participant learns to complete the swipe in 450 ms, then the largest MEPs for the congruent sounds might be generated at this 450 ms time point. Another participant, however, may produce a swipe duration of 300 ms. Thus, the 300 ms time point might be better suited at revealing AMR for this person. Others still, might encode the swipe initiation with the sound, which suggests the 50 ms time point might be suitable. Therefore, examining the MEPs without regard for individual variability could conceal the AMR illustration. Add to this the surprising finding regarding the reduced MEPs immediately post-training, and it perhaps explains why AMR was not revealed.

Accordingly, we examined changes in MEPs across congruent and incongruent sounds for both fingers via pre and post-learning blocks separately (for a related examination, see Ticini et al., 2011, 2019). Furthermore, we selected a single time point to overcome the stated challenges with individual variability. We supposed, if (a) individuals learn to associate a finger movement with a sound (e.g., the index finger with the Congruent_FDI sound), and (b) behavioral learning variabilities cause changes in the timing of the sensorimotor association process. Perhaps, then, (c) we should explore the inhibition of the incongruent sound MEPs relative to maximum congruent sound MEP at a given TMS time point. In other words, we were interested in the maximal dissociation of congruent vs. incongruent sound-generated MEPs across post-blocks within a time point.

We determined where a participant’s maximal congruent sound MEP in either Post-LP block was recorded. This time point then became the guide. We obtained the congruent and incongruent MEPs for both Post-LP blocks at this time point only. A 2 (Sound: MAX-congruent and incongruent) × 2 (Block: Post-LP 1 and 2) ANOVA for each muscle was run. Demonstration of AMR would show that listening to incongruent sounds generates smaller, perhaps inhibited, MEPs when compared to congruent, trained sounds.

Provided the AMR illustration is time-locked to maximal trained muscle activity during audition, the index finger ANOVA with post-learning blocks demonstrated a significant main effect for Sound (F_(1,17) = 26.987, p < 0.001, ηp2 = 0.614). PCs (Bonferroni corrected) showed hearing the congruent sound produced larger MEPs (M = 1.404, SE = 0.143) than the incongruent sound within a time point (M = 0.744, SE = 0.101). This indicates that AMR develops, and trained sounds generate larger MEPs in corticospinal tracts than those recorded upon hearing untrained (incongruent) sounds.

There was also a significant main effect for Post-LP block (F_(1,17) = 12.322, p = 0.003, ηp2 = 0.420). PCs (Bonferroni corrected) revealed the Post-LP 1 block generated smaller MEPs (M = 0.847, SE = 0.125) than those recorded within the Post-LP 2 (M = 1.301, SE = 0.123). Perhaps this modulation of MEPs across blocks reveals a memory consolidation period, which is facilitated by the TMS pulse and sound playback. Sound playback during TMS to M1 over the FDI region could aid the formed sensorimotor association. In other words, Post-LP block 1 might act like another training block (although see the “Discussion” section for a caveat to this explanation).

The same procedure followed for the little finger. Despite using the higher threshold muscle to generate the MT, ANOVA revealed a significant main effect for Sound (F_(1,17) = 19.062, p < 0.001, ηp2 = 0.529). Bonferroni corrected PCs indicated the congruent sound produced larger MEPs (M = 0.313, SE = 0.234) than those recorded when the incongruent sound is played (M = 0.234, SE = 0.034). This data supports the index finger AMR illustrations, and suggest sensorimotor associations developed after training.

To limit concerns with post hoc statistical biases, we used the respective (i.e., individual’s) congruent time point as the guide and explored Pre-LP MEPs, too. For a powerful AMR illustration to exist, we did not expect to find the post-training disassociation between congruent and incongruent sound-derived MEPs to be present in Pre-LP blocks.

Indeed, separate ANOVAs for each finger do not reveal any main effects or interactions when the Pre-LP MEPs are examined using the guide time point. As shown in Figure 4, a disassociation between congruent and incongruent sound MEPs is only present after learning. This indicates that AMR is a trained effect. Together, this suggests that a bidirectional sensorimotor association developed.

Having established sensory prediction mechanisms during action and AMR during audition, we undertook some nonparametric Spearman correlational analyses to explore how these mechanisms are associated putatively. First, negative correlations between the maximum Post-LP 1 MEPs in the index finger (i.e., audition) and N1 peak data during the little finger-swipe (i.e., action) are present (Fz, r = −0.655, p = 0.003; FCz, r = −0.544, p = 0.020). This relationship is mirrored at the Fz electrode with the Post-LP 2 MEPs (r = −0.598, p = 0.009). Also, the maximum Post-LP 1 MEPs in the index finger and the P2 peak during a little finger-swipe are negatively correlated (Fz, r = −0.544, p = 0.020; FCz, r = −0.610, p = 0.007). These data suggest that large index finger MEPs during congruent sound listening are associated with recordings of larger (i.e., less suppressed and more negative) ERP data during the little finger-swipe. Together, this highlights the close ties between a sensorimotor association and sensory prediction during action learning.

In support of this close relationship, behavioral activity and sensory prediction markers also demonstrate some correlation. EMG activity in the index finger during the swipe is negatively correlated with the N1 peak data during audition of the incongruent little finger-sound at Fz (r = −0.513, p = 0.030) and FCz electrodes (r = −0.548, p = 0.019). It would seem that efficient index-swipe movements are associated with the absence of early sensory prediction data when hearing the different (related) sound.

Therefore, negative correlations across sensory prediction, sensorimotor association, and behavioral data highlight an interdependent nature of these components. Simply put, the disassociation of closely related motor control plans (e.g., sensorimotor associations) might be achieved via sensory prediction processes, which are learnt during behavior.