We performed waveform analyses on 2 midline electrodes, Fz and Cz. First, these 2 electrodes are the most frequently used sites for auditory evoked potentials because sources in the Sylvian sulcus project maximally to the top of the head (Hall, 2007). Second, the P300 auditory evoked potential in response to deviant stimuli is classically reported at Fz and Cz (Duarte et al., 2009). In addition, in studies that reported the ERAN (early right anterior negativity) in response to musical syntactic transgressions, this component clearly manifested at Fz and Cz (Koelsch et al., 2001, 2007). Finally, because we used the Geodesic 256 Sensor Net, Cz and Fz are among the most reliable to be placed according to the 10/10 system. To compare the 3 expertise levels for all transgression conditions, we computed Bonferroni corrected contrasts (t-tests) for independent samples on mean voltage amplitude over time between grand-average ERPs of Non-musicians, Amateurs and Experts, over the full −200 to 750 ms window, applying a minimum time constraint of 28 consecutive milliseconds (7 time frames at 250 Hz) for all 3 transgression conditions at both electrodes.

Because our main goal for ERP analyses consisted of microstate analyses comprising all electrodes (see Step 2 for an explanation) as a function of expertise level (EX) and syntactic transgression (T), the waveform analyses principally served to enable comparison to the literature.

The insight that spontaneous EEG and evoked potentials manifest over time as stable voltage topographies lasting tens to hundreds of milliseconds, separated by short periods of transition, led to the concept of “microstates” (Lehmann et al., 1987; Murray et al., 2008; Brunet et al., 2011). Microstate analysis enables studying these stable scalp voltage topographies over time, comprising all recorded electrodes of different groups and conditions in one unified analysis, resulting in a more global assessment than with classical ERP waveform analyses. Such a multivariate approach is particularly suited in the context of a complex experimental design, such as in the current study, with 3 groups, 3 experimental conditions and 204 electrodes. Indeed, it condenses a great amount of information into one series of microstates, while conserving a maximum of the variance present in the data. One of the benefits of this approach is that these different microstates can be compared between groups and conditions in a multifactorial manner.

In the first stage, a k-means cluster analysis on topographic dissimilarity served to segment the grand-mean ERP data of all 204 electrodes of all 3 groups and all 3 transgression conditions over time into one series of k microstates or scalp topographies (brainmapping.unige.ch/cartool). The time frame of analysis extended from 0 to 750 ms after stimulus onset (onset of the terminal chord of the musical piece). Topographic dissimilarity in the context of surface EEG is an indicator of spatial difference between 2 voltage scalp configurations (Murray et al., 2008). An optimal choice of k is essential in order to build meaningful clusters (Fischer, 2011). For that matter, we used cross-validation criteria, minimizing residual variance, as well as a modified Krzanowski–Lai criterion, which provides a quality measure for each clustering (Krzanowski and Lai, 1988; Pascual-Marqui et al., 1995; Tibshirani and Walther, 2005; Murray et al., 2008; Michel and Brandeis, 2009; Brunet et al., 2011). This approach generated an a priori hypothesis in the form of a microstate series occurring over time, based on this multivariate analysis of grand mean ERPs of all groups and conditions.

In the second stage, this a priori hypothesis on the appearance of microstates over time was statistically verified by means of a procedure that fits the microstates to the time series of each individual over a certain time period during which differences occur (Murray et al., 2008; Brunet et al., 2011). The result consists of measures of microstate presence or “duration” (in milliseconds) that can be statistically compared between groups and conditions. Moreover, such an analysis of changes in scalp voltage topographies directly informs on underlying generators in the brain. By physical laws, changes of scalp voltage topographies indicate a change of underlying generators (Vaughan, 1982). Therefore, the time periods during which different microstates appear simultaneously in different groups or conditions provide us with proper time windows to estimate and compare neuronal generators between groups or conditions. We only retained microstates manifesting over periods longer than 28 ms (7 time frames at 250 Hz) because processing of cognitive information in time intervals shorter than several tens of milliseconds is not plausible (Lehmann et al., 1987; Koenig et al., 2002; Brunet et al., 2011).

Microstate durations compared over conditions and groups are commonly not normally distributed. For example, in a certain group, a certain microstate may occur only sporadically, and in some individuals, it may not occur at all, which may constitute simply an experimental effect because a specific stage of information processing may strongly differ as a function of expertise level or transgression condition. Therefore, we applied nonparametric analyses (Kruskal-Wallis, one-way ANOVAs on ranks and Mann-Whitney U-tests).

Because the inverse problem is underdetermined with scalp EEG (multiple solutions would be possible), a priori constraints are necessary to reduce this dimensionality (Michel et al., 2004; Grech et al., 2008). Distinct microstates that occur simultaneously in different groups or experimental conditions, if statistically verified, provide such a priori constraints. We therefore performed the estimation and statistical comparison of underlying neuronal generators over a time window during which significant differences occurred in the presence (duration in milliseconds) of the microstates between the different groups and conditions. Refer to the Results section for precise information on this time window.

The electrical source imaging analysis comprised 2 stages.

In the first stage, we estimated the intracranial current density (CD) distribution (in micro amperes per cubic millimeter; μA/mm³) for the evoked potential of each participant and condition, averaged over a time window that manifested strong microstate differences as a function of expertise level (EX) and syntactic transgression (T), deriving from the second stage of the microstate analyses (see Results section). These source estimations were computed from averaged ERP waveforms over this time window at all 204 electrodes into a solution space represented by a 3D grid composed of 3005 nodes. These 3005 nodes were selected from a 6 × 6 × 6 mm grid equally distributed over the gray matter of the average brain provided by the Montreal Neurological Institute.

From these average ERPs expressed in microvolts (μV), we computed the CD (μA/mm³) in each node of the 3005 3D grid solution space, using a depth-weighted minimum norm (WMN) distributed linear inverse solution (Hamalainen and Ilmoniemi, 1994; Michel et al., 2004). After applying a homogeneous transformation operation to the volume that rendered it to the best fitting sphere (Spherical Model with Anatomical Constraints, SMAC; Spinelli et al., 2000) we used a 3-shell spherical head model to compute the lead field for the 204 electrodes and the inverse solution based on the WMN constraint.

In the second stage, we compared these brain sources by means of statistical parametric mapping (SPM), i.e., comparing the CD for each voxel/node of the 3D grid of the brain between groups and conditions. We computed an ANOVA EX (3, inter) × T (3, intra) at all nodes of the 3D-grid representing the grey matter of the brain using the Statistical Toolbox for Electrical Neuroimaging (STEN; http://www.unil.ch/line/en/home/menuinst/about-the-line/software--analysis-tools.html). To cope with multiple comparisons, only globular clusters composed of minimum 10 contiguous nodes with an individual p-value of < 0.005 met our significance criterion (Knebel et al., 2011; De Meo et al., 2015). Moreover, the spatial criterion of 10 contiguous nodes was verified using the AlphaSim program (http://afni.nimh.nih.gov) under the assumption of a spatial smoothing of 7 mm FWMH (full width at half maximum) and a maximum cluster connection radius of 10 mm in combination with the node p-value of < 0.005 that we applied. After 10,000 Monte Carlo iterations, globular clusters of 10 nodes appeared with a cluster-level likelihood of p < 0.001, and a corresponding node-level p-value of p < 0.00001. Our significance criterion may therefore be considered conservative.

The ERP curves for all groups and conditions are shown in Figure 3 for the central electrodes Fz and Cz. Color-coded bars represent significant Bonferroni corrected contrasts over time between grand-average ERPs of Non-musicians, Amateurs and Experts, with a minimum time constraint of 28 consecutive milliseconds for all 3 transgression conditions for both electrodes. Differences occurred between approximately 200 and 600 ms after stimulus onset but most often between 300 and 500 ms. Experts systematically displayed more positive voltages, but in response to T^app, Amateurs' waveforms overlapped those of the Experts at Cz. These positive waveforms are reminiscent of extended P300/P3b ERPs. Interestingly, these results resemble the behavioral d-prime values, showing the largest differences between Experts vs. Non-musicians and Amateurs for T^sub and between Experts and Amateurs vs Non-musicians for T^app. The exact time periods of significant differences are provided in Supplementary Table 1.

A k-means cluster analysis yielded 7 ERP microstates that together explained 92.75% of the ERP data variance for all groups and conditions.

Figure 4A depicts the results of the first stage of the microstate analyses. The occurrence of the microstates over time is represented by color coded segments below the 9 global field power (GFP) curves, one for each condition and group. GFP provides a single, positive, reference free value of neural response strength all over the brain by taking into account the absolute voltages of all electrodes (Murray et al., 2008). On the right, Figure 4B shows the scalp potential topographies of the 7 microstates, framed by color codes corresponding to panel A.

Between 295 and 480 ms, strong differences occurred between microstates as a function of expertise level and transgression level (cf. Figure 4A). These observed differences were statistically tested in the second stage, comparing the duration (in milliseconds) of Microstates 4, 5, 6, and 7 between all 59 individuals as a function of group and condition. The results of this statistical fitting procedure can be found in Figure 5.

For each microstate, we performed Kruskal-Wallis one-way ANOVAs on ranks to evaluate whether significant differences occurred between the three expertise groups, which were run separately for each of the three conditions (R, T^sub, and T^app). In the case of significant ANOVA results only, we performed Bonferroni corrected Mann-Whitney U-tests to systematically explore differences between expertise levels. We present the results of these tests in the panels of Figure 5 in a hierarchical way (Experts vs. Amateurs vs. Non-musicians), corresponding to our hypothesis on gradual change with expertise level.

Analyses of durations of Microstates 4 and 7 yielded significant differences for all 3 transgression conditions (R, T^sub, and T^app). All of these differences were gradual changes as a function of expertise level, with systematic significant or marginal differences for Experts vs. Non-musicians and Amateurs vs. Non-musicians. Moreover, for Microstate 4, a highly significant difference between Experts and Amateurs occurred for T^sub. Microstate 4 clearly represents non-musician and amateur processing, showing the longest durations for Non-musicians, intermediate durations for Amateurs, and minimal durations for Experts. Microstate 7 shows the opposite, although Amateurs and Experts showed similar durations for T^app. Microstate 5, which yielded duration differences between expertise levels for R and T^sub reflects expertise level, with the longest average durations for Experts. Although mean duration for T^app increased gradually with expertise level for Microstate 5, this difference was not statistically significant due to large variance. Durations for Microstate 6, which seems to be an amateur-specific processing stage based on the fact that Amateurs had the highest values in response to R and T^sub endings, did not yield any significant differences.

The overall conclusion is that for maps 4, 5, and 7, gradual differences manifested as a function of expertise, with Non-musicians showing the highest values for Microstate 4, and Experts showing the highest values for Microstates 5 and 7. These “expert” microstates are centro-parietal positivities with extended P300/P3b-like configurations.

Within the time period of significant difference determined in stage 2 of the microstate analyses (295–480 ms), we performed an ANOVA EX (3, inter) × T (3, intra) on CD for all nodes of the 3D grid, which were labeled according to the Talairach atlas. The results are provided in Table 3. For the statistical thresholds used, we refer to the methods section. The F-values and Talairach coordinates in the table correspond to the local maxima of each listed area.

Figure 6 displays the brain clusters showing significant Expertise × Transgression interaction (Figure 6A), main effects of Expertise (Figure 6B), and main effects of Transgression (Figure 6C), with bar graphs representing the associated CD for the peakvoxel of each Talairach label. Because the main effect of Transgression yielded widely distributed clusters, we only displayed bar graphs for a subset of representative areas, after noting that all brain areas showed a similar CD pattern as a function of transgression condition, i.e., stronger CD values for T^app as compared to R and T^sub. In order to explore the directionality of the effects, we computed post-hoc analyses using Fisher's least significant difference (LSD) correction for multiple comparisons. Fishers' LSD restricts family-wise error rate to alpha in the special case of 3 groups (Howell, 2014; chapter 8). Significant results of these contrasts are shown by asterisks in the bar graphs in Figure 6.

The Expertise × Transgression interaction yielded differences exclusively in a cluster within the right posterior middle temporal gyrus, close to the lateral occipital complex. The peakvoxel occurred in Brodmann area (BA) 37 (rMTG, Talairach 43, −67, 17, see Figure 6A). No significant differences between the 3 groups occurred for R endings (although a gradual increase of mean CD can be clearly perceived). The post hoc tests indicated stronger CD in this brain area in Experts than in Non-musicians and Amateurs in response to T^sub. In contrast, in response to T^app, the post hoc tests evidenced stronger CD in both Experts and Amateurs as compared to Non-musicians. In other words, T^sub yielded stronger CD in experts than in all non-experts, and T^app induced stronger CD in all musicians than in non-musicians.

The main effect of Expertise indicated differences within clusters in the bilateral anterior cingulate gyrus, medial frontal gyrus and right parahippocampal gyrus. In all these brain areas, the post-hoc tests displayed stronger CD for Amateurs and Experts than for Non-musicians. In the ACC cluster identified by our analyses, mainly rostral but also dorsal parts of the ACC are present (see Figure 6Bi). The cluster in the right parahippocampal gyrus (see Figure 6Bii) also contained solution points in the amygdala and in the hippocampus proper.

The main effect of Transgression yielded differences in widely distributed brain clusters, including, in order of decreasing F-values, the bilateral anterior cingulate gyrus and medial frontal gyrus (see Figures 6Cii,iii,viii,ix), inferior frontal gyrus (see Figures 6Ci,vii), and parahippocampal gyrus (see Figures 6Cv,vi). In these brain areas and all others, the post hoc tests indicated stronger CD for T^app than for both R and T^sub. The enhanced activations in response to transgression salience in the bilateral inferior frontal cortex clusters (BA 47, see Figures 6Ci,vii) are situated in the vicinity of the Broca area (BA 44 & 45) and its right hemisphere homolog.

The peakvoxel of interaction between expertise and transgression levels occurred in the right posterior middle temporal gyrus (rMTG) in BA 37. However, as observed for d-prime values, ERP and microstate results, Experts manifested higher (CD) than Non-musicians and Amateurs in response to T^sub, and Experts and Amateurs both displayed higher CD than Non-musicians in response to T^app. This secondary auditory area is most likely involved in “sound-to-meaning-mapping” in language processing, operated via a connection from the posterior left MTG (lMTG) to the left inferior frontal areas via the ventral pathway (Hickok and Poeppel, 2004). The posterior MTG seems to function as an interface between the sound representations in the superior temporal gyrus and more abstract representations, which are largely distributed in the brain (Hickok and Poeppel, 2004). The lMTG can thus be considered the “gateway” to this processing chain in language processing. We recently showed the existence of a right hemisphere homologue of this white matter pathway for music processing in the same group of participants. The morphology of this pathway became gradually more concise as a function of expertise level, and this increased precision or “Tract Consistency” of the pathway correlated with higher d-prime values for T^sub and T^app (Figure 3E in Oechslin et al., 2017). Such an analysis of the meaning of sound patterns seems most pertinent in the context of transgressed musical closure, and may be progressively enhanced following musical training. On the functional level, this connection between the right middle temporal gyrus and right inferior frontal areas via the ventral pathway, specifically implicated in the analysis of auditorily presented stimuli that transgress musical structure, has also been observed (Bianco et al., 2016).

Our preceding fMRI study on the same population (Oechslin et al., 2013b) also gave rise to an Expertise x Transgression interaction within the rMTG, although more anteriorly located. Even though the results cannot be directly compared, as we performed fMRI contrasts that subtracted brain regions that responded to R endings from areas responding to T^sub and T^app, we nevertheless consider this compatible finding interesting. Indeed, identical groups and stimuli gave rise to an interaction effect with the peakvoxel in the rMTG in both fMRI and EEG settings, albeit at different locations. The peakvoxel found in the current experiment is situated in the vicinity of the angular gyrus and the temporal–parietal–occipital junction, and represents a good candidate for a sound-meaning interface (Hickok and Poeppel, 2000). Damage to the posterior MTG has been associated to pitch amusia in non-recovered lesion patients when occurring in the right hemisphere (Sihvonen et al., 2016), and to alexia and agraphia when occurring in the left hemisphere (Sakurai et al., 2008).

Music-syntactic processing, which relies on fine-grained pitch processing, is known to take place mostly in the right hemisphere (Zatorre, 2001; James et al., 2008; Cha et al., 2016; Oechslin et al., 2017).

Microstates 5 and 7, which are characterized by expert processing and context-based memory updating, may rely partially on the right MTG in this musical syntax processing paradigm. The P300/P3b components appear to have temporal-parietal origins as well as anterior cingulate and hippocampal sources (Polich and Criado, 2006; Polich, 2007).

A main effect of expertise manifested in three different brain areas, systematically separating responses from musicians (both Experts and Amateurs) from Non-musicians. Musicians always showed substantially higher CD in these areas than Non-musicians. The fact that the three transgression conditions set intrinsically different requirements to the participants may explain why no significant differences occurred between Experts and Amateurs.

Detecting T^app implies perceiving an explicit syntactic transgression of correct closure, as it presents a degree IV (Subdominant) instead of a degree I (Tonic). In contrast, T^sub provides the proper degree; i.e., all the notes of the appropriate tonic chord are present, only with the third in the bass instead of the root. Only experts recognized this very subtle transgression consistently. As observed for the behavioral, ERP, and microstate results, T^sub responses most strongly dissociated Experts from Non-musicians and Amateurs, and T^app responses distinguished Experts & Amateurs from Non-musicians. It is therefore plausible that gradual differences between Amateurs and Experts did not appear for the main effect of expertise when all transgression conditions were collapsed. The observed increased CD rather shows a general effect of musical training and proficiency (for all levels of transgression), annihilating differences between Experts and Amateurs, which were primarily expressed for T^sub.

The bilateral anterior cingulate cortex (ACC; see Figure 6B and Table 3), that showed stronger CD in musicians, plays a role in executive functioning related to conflict monitoring and error detection (Bush et al., 2000; Iannaccone et al., 2015). It seems logical that trained pianists, who cope constantly with error detection and prevention during musical training, would develop a stronger conflict monitoring function. However, dorsal and rostral parts of the ACC appear differently involved the respective cognitive and emotional processes accompanying conflict monitoring and error detection, although a functional continuum between cognitive and affective processing is suggested rather than fully segregated operations (Mohanty et al., 2007). As mainly the rostral but also the dorsal parts of the ACC are present in the ACC cluster identified by our analyses, both cognitive and affective aspects may play a role in the processing of correctness of musical closure. Interestingly, the dorsal ACC appears to connect to prefrontal areas, whereas the rostral ACC connects to limbic zones (Mohanty et al., 2007), corresponding to the other two clusters (in the medial frontal cortex and parahippocampal gyrus) that manifested stronger CD in musicians. This supports our hypothesis on the double nature (emotional and cognitive) of closure congruity processing. After all, music represents an emotional stimulus, and deceptive cadences as used here induce -as the terminology suggests- a kind of deception, especially in trained musicians. But such non-conclusive endings also challenge cognitive processing of music. Magno et al. (2006) suggested an evaluative function of the ACC and found the highest ACC activity for error prevention.

The adjacent medial frontal gyrus may participate in this function of error monitoring and evaluation on the cognitive and affective levels (Ridderinkhof et al., 2007; Simons, 2010). Moreover, this area seems specifically associated to tonality processing, representing an interface of cognitive, affective, and mnemonic processing (Janata et al., 2002a).

The activations in the right parahippocampal gyrus, which also contains parts of the amygdala and the hippocampus proper, nicely match with earlier findings on expert pianists' responses to apparent in-key harmonic-syntactic incongruities in piano music (peaking between 350 and 400 ms after musical closure onset; James et al., 2008). Apart from the obvious “emotional” hypothesis, these right medial-temporal lobe structures are also involved in higher-order pitch and tonality processing. Indeed, patients with right-sided amygdala-hippocampectomy were severely hampered performing the Seashore musical aptitude test (Wieser, 2003; James et al., 2008), and healthy musically experienced persons processing tonality changes also activated their right hippocampus (Janata et al., 2002b). In the same vein, the amygdala has received the epithet “relevance detector” (Sander et al., 2003), because it seems implicated in the appraisal of relevant events. Now, musical congruity processing is highly relevant and a much-practiced skill for musically trained people.

If we compare the current results with the main effects of expertise in our fMRI study (Oechslin et al., 2013b), the peak value was also found in the right ACC, but in contrast to the current study, only right hemisphere expertise effects were found. However, as mentioned above, as we performed fMRI contrasts that subtracted brain regions that responded to R endings from areas responding to T^sub and T^app, the results cannot be directly compared, and left-sided ACC in musicians may have been cancelled out by the fMRI contrasts. As for the rMTG cluster identified in the Expertise × Transgression interaction, the ACC may also partially underlie Microstates 5 and 7, and may underpin error detection and conflict monitoring, which are closely linked to context-based memory updating. Both error detection and conflict monitoring are necessary to detect an error or deviance with respect to an internal learned model.

This experiment focused on gradual brain adaptations as a function of musical expertise and proficiency in the context of different levels of transgression saliency. Therefore, the main effect of Transgression is of the least interest in the current study, as all expertise levels were collapsed. Moreover, as the three transgression conditions set intrinsically different requirements for the participants, as explained in the second paragraph of the preceding subchapter Main Effect of Expertise, widely distributed brain clusters responded. Discussing all of them would lead us out of the scope of this study, which principally focuses on musical expertise.

However, the bilateral clusters in the ACC and medial frontal cortex as well as in the parahippocampal gyrus that contributed to the main effect of Transgression pertain to the same brain areas as those representing the main effect of Expertise and showed systematically higher CD in response to T^app as compared to R and T^sub. In all participants, responses to apparent thus more salient transgressed endings evoked stronger responses in these brain areas than regular and subtly transgressed closures, independent of expertise level. As both Non-musicians and Amateurs appraised T^sub with relatively low d-prime values, the accurate appraisal by Experts seems to be annihilated by the collapsing of the groups. The explanations we provided for the involvement of the ACC, medial frontal cortex, and parahippocampal gyrus for the factor Expertise also hold for the Transgression condition without taking Expertise into account: all participants -to varying degrees- were involved in error monitoring and evaluation on the cognitive and affective levels and performed higher-order pitch analyses. However, it seems that when all expertise levels are collapsed, more involvement of d-ACC and thus the cognitive component of the error monitoring manifests for the main effect of transgression discussed in this subchapter than for the main effect of expertise (see Figure 6Ciii). This may be explained by the fact that the task becomes progressively easier with increasing expertise, demanding less cognitive effort and likewise, requiring less error prevention.

The stronger activations in the infero-frontal regions, close to the Broca area, can be explained by their participation in musical syntax processing (Maess et al., 2001; Zatorre, 2001; Levitin and Menon, 2003; Tillmann et al., 2006; James et al., 2008; Koelsch et al., 2013) as well as in working memory function (Janata et al., 2002b; Oechslin et al., 2017). Both explanations are plausible in the context of detecting transgressions while appraising the endings of rich polyphonic musical pieces.

Comparing the current results with the main effects of transgression in our fMRI study (Oechslin et al., 2013b), no overlapping activations occurred. Again, apart from the different neural mechanisms involved, this may be due to the fMRI contrasts used, specifically for the effect of transgression, as we subtracted brain regions that responded to R endings from areas responding to T^sub and T^app.