Auditory deviance detection occurs around 150 ms after the onset of a deviant sound. Recent studies in animals and humans have described change-related processes occurring during the first 50 ms after sound onset. However, it still remains an open question whether these early and late processes of deviance detection are organized hierarchically in the human auditory cortex. We applied a beamforming source reconstruction approach in order to estimate brain sources associated with 2 temporally distinct markers of deviance detection. Results showed that rare frequency changes elicit an enhancement of the Nbm component of the middle latency response (MLR) peaking at 43 ms, in addition to the magnetic mismatch negativity (MMNm) peaking at 115 ms. Sources of MMNm, located in the right superior temporal gyrus, were lateral and posterior to the deviance-related MLR activity being generated in the right primary auditory cortex. Source reconstruction analyses revealed that detection of changes in the acoustic environment is a process accomplished in 2 different time ranges, by spatially separated auditory regions. Paralleling animal studies, our findings suggest that primary and secondary areas are involved in successive stages of deviance detection and support the existence of a hierarchical network devoted to auditory change detection.
The rapid discrimination of novel sounds in complex acoustic environments enables us to reallocate our attentional resources to rare and potentially relevant events (Escera et al. 1998; Escera and Corral 2007). In humans, the processing of such rarely occurring sounds is indexed by an automatic event-related brain potential/event-related field (ERF) peaking at 150–250 ms: The mismatch negativity (MMN; Näätänen et al. 2007) or its magnetic counterpart (MMNm). However, recent findings by our group have shown that violations of an auditory stimulus feature trace are also reflected at much earlier latencies in the auditory brain as indexed by amplitude modulations of the auditory middle latency responses (MLRs: Na, Pa, Nb, and Pb), in the initial 50 ms after change onset (Slabu et al. 2010; Althen et al. 2011; Grimm et al. 2011, 2012; Leung et al. 2012). Specifically, enhancements of Pa (Slabu et al. 2010), Nb (Grimm et al. 2011), and Na components of the MLR (Grimm et al. 2012), in addition to the MMN in long-latency responses (LLRs), were elicited as “genuine” deviance-related responses, that is, controlling for confounding frequency-specific adaptation effects. Such results suggest that change detection is hierarchically structured in different processing stages along the auditory pathway (Grimm and Escera 2011; Slabu et al. 2012), thus paralleling other properties of the auditory system such as spectral (Wessinger et al. 2001; Kumar et al. 2007), temporal-scale (Kiebel et al. 2008; Lerner et al. 2011), or speech processing (Scott and Johnsrude 2003). A hierarchically organized deviance detection network would be in line with the “predictive coding” hypothesis (Friston 2005), stating that the brain predicts the nature of forthcoming events based on hierarchical message passing among cortical areas (Garrido et al. 2009). Indeed, such hypothesis has been recently implemented in a neuronal model accounting for several MMN findings (Wacongne et al. 2012).
Additional evidence in favor of a hierarchical auditory deviance detection system comes from animal studies showing that stimulus-specific adaptation (SSA), that is, the reduction of the spiking rate to standard stimuli while keeping robust responses to deviant stimuli (Ulanovsky et al. 2003) is a widespread property of the auditory system, including cortical (Ulanovsky et al. 2003, 2004; von der Behrens et al. 2009) and subcortical structures (Pérez-González et al. 2005; Malmierca et al. 2009; Antunes et al. 2010).
In humans, source localization studies have yielded generators of MMN and MMNm confined to bilateral secondary auditory areas like the anterior Heschl's gyrus (HG) and superior temporal gyrus (STG; Opitz et al. 2005; Schönwiesner et al. 2007). On the other hand, MLRs are known to follow a medio-lateral and postero-anterior propagation starting in medial portions of the HG (Liegeois-Chauvel et al. 1994; Pantev et al. 1995; Yvert et al. 2005) with Pa/Pb complex thought to reflect a processing stream from primary auditory cortex (PAC) to STG (Howard et al. 2000). To our knowledge, no studies assessing the anatomical generators of change detection in the time range of MLR have been conducted yet.
We aim to show that hierarchically distinct areas of the auditory cortex are involved in early and late auditory change detection, thus supporting that detection of deviant sounds is functionally organized in a hierarchical fashion. We hypothesize that generators of deviance processing in the time range of MLR will be located in or close to PAC, whereas those accounting for MMNm will be located in secondary areas such as STG.
Materials and Methods
Thirteen healthy, normal-hearing subjects (7 females) aged 22–34 years (27 years, mean age, standard deviation [SD] = 3.7) took part in the experiment. Hearing level was assessed binaurally with a pure tone audiometry for 5 frequencies (250, 500, 1000, 3000, and 8000 Hz) before or just after the task. The experimental protocol was approved by the Ethical Committee of the University of Barcelona and was in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki). Participants gave written informed consent before the experiment.
Stimuli and Procedure
Auditory stimuli consisted of 50-ms pure sine wave sounds (5 ms rise, 20 ms fall) delivered at 60-dB sound pressure level. Sounds were binaurally delivered by an Etymotic ER-30 system (Etymotic Research, Inc. United States of America) via plastic earpieces. The 18-ms delay in the transmission of sound was compensated for by an appropriate shift of the trigger signal.
The stimulus-onset asynchrony was randomly jittered between 150 and 350 ms. The experimental design consisted of an oddball block including 800-Hz frequency deviant tones occurring at a probability of 0.20- and 1040-Hz repetitive standard tones occurring at a probability of 0.80. In a reversed oddball block, the roles of deviant and standard stimuli were switched in order to allow the comparison between physically identical stimuli. Only standard stimuli from the reversed block and deviants from the oddball block were used in further analyses. For the sake of clarity, we will refer to “reversed standard” as “standard” condition from now on. A control block, comparable with the one first introduced by Schröger and Wolff (1996) was used, in which stimuli of 5 different frequencies (800, 1040, 1280, 1664, and 2040 Hz) were randomly presented, each at a probability of 0.20. This was done to preclude refractoriness confounds. The 3 blocks were split into 6 blocks, which were presented randomly. In sum, a total amount of 1200 stimuli of each condition (deviants from the oddball block, standards from the reversed block, and controls) were used in the subsequent analyses.
During the recording session, subjects were required to lie as still as possible on a bed with their head inside the helmet-like device for approximately 1 h. Participants were instructed to relax, to ignore the auditory stimulation, and to attend to a silent movie with subtitles. A whole-head magnetoencephalography (MEG) system (148 biomagnetometers, 4D Neuroimaging Magnes 2500WH, San Diego, CA, United States of America) recorded the magnetic currents at 1017 Hz sampling rate. Data were on-line high-pass filtered at 1 Hz and stored for off-line analysis. A bipolar electrooculogram was recorded to identify eye movements. Five small sensor position coils were attached to the forehead and to the periauricular points in order to determine the position of the head and to track any head movement occurring during the recording. Data sets in which the relative position of the head changed by >0.7 cm throughout the recording session were discarded from further analyses. For each subject, the headshape including the forehead, the nose, and the location of the sensor position coils were digitized using a digitizer wand (Polhemus Fastrak, Polhemus Inc., Colchester, VT, United States of America). Additionally, T1-weighted, 3-dimensional (3D) spoiled gradient-echo, magnetic resonance images (MRIs) of each individual brain were acquired using a 1.5-T Signa CV (General Electric, Milwaukee, WI, United States of America) to allow superimposition of MEG and MRI data.
Data analysis was performed using Matlab 7.10 (The MathWorks) and FieldTrip (Oostenveld et al. 2011; www.ru.nl/neuroimaging/fieldtrip, 20101006 release). The signal was digitally low-pass filtered at 150 Hz. The strongest components corresponding to cardiac and ocular artifacts were projected out of the MEG signal using independent component analysis (ICA; “runica” algorithm implemented in FieldTrip/EEGLAB, http://sccn.ucsd.edu/eeglab/). Eight hundred samples (1 s length) including data from each condition were used for ICA. On average, 3.4 components per subject were identified as blinks, saccades, or cardiac artifacts (maximum of 5 components in subject 7) on the basis of their scalp topography and continuous activity (Jung et al. 2000). Additionally, a rejection threshold of 4pT was applied to remove high amplitude artifacts. Finally, LLR data were epoched in 400-ms time windows (including 100-ms pre-stimulus baseline) and low-pass filtered at 30 Hz (2-pass Butterworth filter; filter order of 4). For MLR analysis, epoch length was 200 ms (including 50-ms pre-stimulus baseline) and a high-pass filter of 15 Hz was applied (2-pass Butterworth filter; filter order of 4). Artifact-free trials were averaged separately for each condition (deviant, standard, and control).
For the source localization analysis, we first coregistered the MEG and the anatomical MRI coordinate systems by using a semiautomatic procedure. Three landmarks (nasion, and the 2 periauricular points) were localized in each individual MRI and used for a first alignment with the MEG coordinate system. We then performed an automatic fit between the digitized headshape and the scalp surface extracted from the MRIs based on an iterative procedure. In each iteration, we applied a modified version of the iterative closest point algorithm (Besl and Mckay 1992; icp2©: http://www.csse.uwa.edu.au/~ajmal/code/icp2.m) to a different initial position of the digitized headshape. The location of the headshape relative to the scalp surface was updated, in each iteration, to the one providing the minimum distance error between them. Based on the segmentation of the brain surface of each individual's MRI, we obtained a semirealistic single-shell head model for each participant. Subsequently, a standard 3D grid (6 mm spacing, 12773 voxels inside the head) derived from the Montreal Neurological Institute (MNI) T1 template brain was adapted to each individual's brain volume by means of an inverse-normalization procedure based on a linear affine transformation (SPM2, Wellcome Trust Center for Neuroimaging, London, United Kingdom; http://www.fil.ion.ucl.ac.uk/). The use of the standard MNI grid would thus allow for group averaging and statistical comparison of results. Leadfields were computed for each grid voxel on the basis of a quasistatic approximation of the brain surface as a single shell (Nolte 2003). The weakest orthogonal component at each voxel of the leadfield matrix was excluded. Neuronal sources of interest were identified using a time-domain minimum-variance spatial filter: The linearly constrained minimum variance beamformer, designed to detect a signal corresponding to a specific location and attenuate signals from all other locations (Van Veen et al. 1997). A single covariance matrix was computed for each subject from the combined datasets from all 3 conditions (deviant, standard, and control), both for MLR (from −50 to 100 ms) and for LLR (from −100 to 200 ms). The covariance matrix and the leadfield matrix were used to compute common spatial filter weights with regularization set to 10% of the mean power. By using common filter weights, we ensured that differences in source activity across conditions were not due to differences between filters. Subsequently, we projected the sensor-level signal of each condition and trial into each voxel of source space through the common spatial filter corresponding to a dipole at this location with fixed optimal orientation. Finally, single-trial data were subsequently averaged separately for each condition. This procedure thus provided an averaged time course of activity in source space for every subject and condition. To reduce spatial filter biases toward the center of the head, voxel activity was normalized using the neural activity index where the estimated power at each grid point is divided by an estimate of the noise (Van Veen et al. 1997).
Individual mean amplitude pseudo-Z maps of the ERF components in each condition were computed by averaging pseudo-Z values within a time window of 6 and 20 ms around the peak latencies of the MLR (Nam, Pam, Nbm, and Pbm) and LLR (N1m/MMNm) components, respectively. Peak latencies for the mentioned components were derived from the root-mean-squared grand-average computation for deviant, standard, and control stimuli. Evoked auditory activity of different conditions was statistically compared using nonparametric cluster-based permutation t-tests (Singh et al. 2003; Maris and Oostenveld 2007). This allowed us to determine which voxels showed statistically significant activity by comparing the grand-mean pseudo-Z value of a given voxel to a distribution of permuted pseudo-Z values. Permutation methods have been widely used in studies applying beamforming techniques on event-related data since no explicit parametric distribution of the population is required (Herdman et al. 2003, 2007; Chau et al. 2004; Cheyne et al. 2006).
Significant sources of MMNm and deviance-related effects for MLR were computed by pairwise comparisons of ERF pseudo-Z maps against conditions (deviant vs. standard; deviant vs. control) for all components of interest in the above-mentioned time intervals (Nam, Pam, Nbm, and Pbm for MLR; MMNm for LLR) using a nonparametric cluster-based procedure that effectively corrects for multiple comparisons (see Maris and Oostenveld 2007, for details on the method). This type of test defines the clusters of interest based on the actual distribution of the data and tests the statistical differences between condition waveforms in each particular voxel using a Monte-Carlo randomization method. Clusters were defined as spatially adjacent voxels where a dependent samples t-test with respect to the pseudo-Z values in 2 conditions exceeded an a priori threshold (P < 0.001 for LLR and P < 0.01 for MLR). A lower significance level for MLR was chosen in order to obtain cluster volumes (for MLR and LLR) of comparable size. For each cluster, a statistical analysis was calculated by taking the sum of all individual voxel t-statistics within a cluster. The Type I error rate for the complete set of 12773 voxels was controlled by evaluating the cluster-level test statistic under the randomization null distribution of the maximum cluster-level test statistic. Specifically, the null distribution was obtained by randomly permuting the data between the 2 experimental conditions within each participant 5000 times.
Similarly, we estimated the sources of the transient N1m fields in the temporal lobe in order to identify those cortical regions showing distinctive activity to identical stimuli embedded in the different conditions (deviant, standard, and control). Pairwise comparisons of the baseline maps against the N1m maps were carried out. Baseline maps were obtained by averaging pseudo-Z values in each condition between −100 and 0 ms. N1m pseudo-Z maps were obtained by averaging each condition between 105 and 140 ms. By defining a broad interval, we included the N1m component in each condition. Clusters were defined as spatially adjacent voxels where dependent samples t-test with respect to the pseudo-Z values in the 2 time intervals exceeded an a priori threshold (P < 0.0002 for deviant, P < 0.0002 for control, and P < 0.005 for standard). The null distribution was obtained by randomly permuting the data between the baseline and the N1m response within each participant 5000 times.
Determination of across-subject differences in the localization of deviance-related auditory areas between MLR and LLR was performed by pairwise comparisons of the location of those individual voxels showing the largest pseudo-Z values, “peak-voxels,” in the subtracted activation maps. Individual pseudo-Z maps for MMNm and MLR components were computed by averaging the pseudo-Z values within a time window of 6 and 20 ms around the individual peak latencies of the MLR (Nam, Pam, Nbm, and Pbm) and LLR (N1m/MMNm) components, respectively. Peak-voxel coordinates in the 3 axes (x, y, and z) were exclusively extracted from previously computed clusters in order to ensure that all locations were inside an area showing statistically significant deviance-related effects. Results were considered significant when Student's t-tests yielded P-values (2-tailed) <0.05. Likewise, localization differences of N1m peak-voxels for the 3 experimental conditions were estimated across subjects. Peak-voxel coordinates in the 3 axes (x, y, and z) were extracted solely from previously computed clusters showing a significant effect for N1m. For each axis, a repeated-measures analysis of variance (ANOVA) including the factor condition (deviant, standard, and control) was calculated. Results were considered significant when P < 0.05 using a 2-tailed analysis. Bonferroni correction was used for multiple pairwise contrasts.
We delivered low-probability sounds interspersed in a context of frequently repeated tones to show that auditory change detection can be traced by both LLR and MLR. The event-related beamforming approach allowed us to estimate the contribution of different neural generators involved in deviance detection. Figure 1 shows the ERF and topographies of the grand-averaged magnetic fields corresponding to LLR and MLR for the 3 different conditions. N1m component peaked at 115 and 118 ms after sound onset in the deviant and control condition, respectively. The N1m field in the standard condition showed a clear adaptation as a result of repetitive stimulation. Unexpectedly, N1m in response to control tones elicited a higher magnetic response than deviant tones. Middle latency evoked fields were approximately 3 times smaller in amplitude when compared with LLR fields. On average, Nam component peaked at 23 ms; Pam peaked at 32 ms; Nbm peaked at 43 ms; and Pbm peaked at 55 ms. As shown in Figure 1, the field distribution of both MLR (Nbm) and LLR (N1m) responses showed a clear dipolar distribution over sensors located in the Sylvian fissure.
For LLR, deviant versus standard comparison resulted in one cluster where pseudo-Z values were significantly higher for deviant when compared with standard tones (P < 0.001, corrected for multiple comparisons) in the latency range of N1m/MMNm. Figure 2 shows the grand-mean source estimate of deviant versus standard (MMNm) activity exceeding the statistical threshold in the long-latency range. This cluster was clearly located in the right temporal lobe, specifically, in both medial and lateral portions of STG, HG, and middle temporal gyrus. The peak-voxel in the thresholded grand-mean source estimate of MMNm was located between the right middle temporal gyrus (MTG) and the right STG. This area, posterior to Heschl's sulcus, is usually defined as planum temporale (PT; MNI coordinates: 54, −18, −2). No clusters over left auditory cortex areas exceeded the imposed statistical threshold for MMNm. As can be inferred from the virtual channel waveform obtained from the peak-voxel in Figure 3, MMNm peaked in the grand-average waveform at 115 ms. Deviant versus control comparisons, intended to provide a clearer picture of net memory-based comparison processes than the “classic MMN”, resulted in no statistically significant voxels. Tones in the control condition elicited stronger activity than deviant tones; however, this difference in activity did not reach statistical significance.
In the middle latency time range, the same procedure was repeated for the main MLR components. Deviant versus standard comparisons for the Nam, Pam, and Pbm components showed no active voxels exceeding the mentioned a priori threshold. Only deviant versus standard pairwise comparison for the Nbm component resulted in a significant cluster of activity (P < 0.01). As shown in Figure 4, this cluster was localized over the right temporal lobe, specifically, in areas including HG, a small aspect of the medial STG, and extending anteriorly to the insula. The peak-voxel in the thresholded grand-mean source estimate of Nbm (deviant vs. standard) was located in the anterior part of the medial HG, next to the long insular gyri (MNI coordinates: 42, −18, 4). Figure 5 shows the virtual channel obtained from the peak-voxel in Nbm (deviant vs. standard) peaking at 43 ms. Cluster-based t-test for paired differences between deviant and control tones yielded no statistically significant clusters for any of the components. Unlike LLR, Nbm deviant condition elicited numerically larger activity than control tones; however, the difference did not reach statistical significance.
To further investigate location differences in the generators of MMNm and Nbm differential activities across subjects, individual peak-voxels from the deviant minus standard subtraction waveforms were extracted. To ensure that chosen peak-voxels' activity represented a significant deviance enhancement, individual peaks of maximal activity were only extracted from those areas showing statistically significant effects in the grand-averaged activity maps. Accordingly, MMNm and Nbm (deviant–standard) individual peak-voxel coordinates (x, y, and z) were pairwise compared among our sample of participants. Figure 6 shows the location of the individual peak-voxels for LLR (MMNm) and MLR (Nbm) located inside a cluster reflecting significant deviance detection activity. Results derived from repeated measures t-test showed that there was a significant difference in the y-axis (antero-posterior; t = 3.288, P = 0.006) between the peak-voxels for MMNm (mean = −16.77, SD = 7.17) and Nbm differential activity (mean = −5.54, SD = 11.26). In the x-axis (medio-lateral; t = −2.334, P = 0.038), Nbm difference peak-voxel positions (mean = 46.92, SD = 11.3) were located more medial than MMNm peak-voxels (mean = 56.3, SD = 11.84). No statistically significant differences were found in the z-axis (inferio-superior).
As a priori-defined region of interest (ROI) for each latency range could bias localization results, we decided to repeat the procedure but using only one cluster or ROI for both MLR and LLR subtracted responses. This cluster contained all possible voxels that appeared as statistically significant in both LLR and MLR cluster-based analyses. We found that a difference between MMNm (mean = 53.31, SD = 11.84) and Nbm (mean = 46.62, SD = 11.72) peak-voxels existed in the x-axis (t = −2.51, P = 0.027). Also, for the y-axis, MMNm peak-voxels (mean = −16.77, SD = 7.17) were posterior to deviance-related Nbm peak-voxels (mean = −7.54, SD = 12.65; t = 2.43, P = 0.032).
Additional analyses on the localization of the N1m response to the different conditions provided means to argue in favor for a genuine involvement of change detection. N1m in response to deviant and control tones resulted in one cluster on the right hemisphere where pseudo-Z values were significantly higher for N1m when compared with baseline levels (P < 0.002, corrected for multiple comparisons). For N1m to standard tones, the same procedure yielded 2 clusters located bilaterally on the supratemporal planes (P < 0.005, corrected for multiple comparisons). Results are displayed in Figure 7. These statistically significant clusters were subsequently used as spatial constrains to compute statistics across individual peak-voxels. One-way ANOVAs for each of the 3 axes (x, y, and z) revealed statistically significant differences between conditions. In the x-axis (F1,12 = 8.03, P < 0.003), pair-wise comparisons showed that deviant-related N1m sources were located lateral to standard-related N1m sources (t = 3.597; P < 0.012). Deviant–control contrast did not yield a statistically significant difference when Bonferroni correction was applied (t = 2.61; P < 0.07). In the y-axis (y-axis: F1,12 = 18.670, P < 0.001), both deviant and control conditions showed more anterior N1m source locations than the standard condition (t = 4.98, P < 0.002; t = 4.726, P < 0.002). Finally, in the z-axis (F1,12 = 21.956, P < 0.001), deviant and control conditions yielded N1m source locations being more superior than the standard condition (t = 6.236, P < 0.001; t = 5.162, P < 0.002).
In sum, both grand-average mean locations and intersubject peak-voxel results showed that the deviance-related enhancement of activity reflected in the Nbm component was generated by sources located medially and anteriorly to the ones of the deviance-related activity in the MMNm time range. In addition, analyses of transient responses in the long-latency range revealed spatially distinct source generators involved in the N1m response to the 3 different stimulus conditions.
Results from this study have shown that areas involved in auditory deviance detection in the human brain exist in 2 separated spatial and temporal domains. Regions showing larger activity for deviant events in the time range of MLR were located in the right hemisphere, anterior and medial to those of the MMNm, and overlapping anterior aspects of HG, temporoinsular areas, and antero-medial portions of the right STG. Consistent with recent electroencephalography (EEG) findings (Grimm et al. 2011, Leung et al. 2012), unexpected pure tones elicited larger Nbm responses than physically identical tones occurring in a repetitive fashion. Source estimates for deviance-related Nbm overlapped the anterior rim of HG, including its medial and central aspects, a region equivalent to the cytoarchitechtonically koniocortical Te1 area (Brodmann area 41). Te1.0 subdivisions have been described in cytoarchitechtonic and probabilistic maps of the human auditory cortex as the PAC (Morosan et al. 2001; Rademacher et al. 2001). The grand-averaged peak-voxel was located medially in a site overlapping the medial aspect of HG. These results are in line with previous electrophysiological studies pointing at different regions along the medio-lateral and posterior–anterior axes of HG as the neuronal generators of the Pa/Pb components of the MLR. Intracranial recording studies (Liegeois-Chauvel et al. 1991, 1994) localized generators of auditory MLR between 30 and 50 ms in medial and central parts of HG, respectively. These findings were further confirmed by means of MEG recordings, showing an origin of MLR at 20 ms in medial HG (Scherg et al. 1989; Scherg and von Cramon 1990; Lütkenhöner et al. 2003). Similarly, MEG studies (Yoshiura et al. 1995; Gutschalk et al. 1999; Inui et al. 2006) showed 2 different sources accounting for the early and late part of the MLR, located in the medial and lateral parts of HG, respectively. Their findings, suggestive of a serial activation, are consistent with anatomical findings in monkeys showing strong connections between core and adjacent belt regions (Merzenich and Brugge 1973; Galaburda and Pandya 1983). Source generators of transient Nbm component (peaking around 40 ms after sound onset) have been located over lateral areas of the STG (Yvert et al. 2001; Inui et al. 2006). Source localization discrepancies might stem from differences in the time intervals chosen to delineate event-related activity in this latency, different reconstruction techniques, and experimental design. Still, the localization of the deviance-related activity found here is highly consistent with localization of MLR components shown by the above-mentioned MEG studies and intracranial recordings (Liegeois-Chauvel et al. 2001). To our knowledge, this is the first time that the sources of deviance-related MLR have been localized in a region overlapping medial and lateral HG.
Source estimates of MMNm (deviant minus standard) spanned right hemisphere regions of STG, MTG, superior temporal sulcus (STS), and posterior portions of HG. Our results are in agreement with the bulk of MMN localization studies, where, independently of the technique or stimulation employed, STG is the most consistent finding (Opitz et al. 1999, 2002; Tervaniemi et al. 2000, 2006; Doeller et al. 2003; Schall et al. 2003; Rinne et al. 2005; Schönwiesner et al. 2007; but see Maess et al. 2007). Specifically, the grand-average peak-voxel for MMNm was localized in STG, bordering posteriorly with HG. Source estimates as a whole overlapped cyto- and receptoarchitectonic Te2.2 and Te3 subdivisions of the auditory cortex (corresponding to BA 42 and BA 22; Morosan et al. 2005), thought to correspond to secondary auditory areas. Also, activity overlapping HG was found, consistent with previous findings (Opitz et al. 2005; Maess et al. 2007; Schönwiesner et al. 2007). Schönwiesner et al. (2007) reported activation in STG and in both medial and lateral HG, suggesting contributions of PAC areas to deviance detection. Similarly, Opitz et al. (2005) disentangled the existence of a cognitive mechanism (deviant vs. control) in secondary areas, and a sensory mechanism (standard vs. control) generated in a primary auditory area of HG. The right lateralization of the MMNm observed in our study is consistent with previous non-linguistic studies showing larger MMN amplitude in the right hemisphere, irrespective of the stimulated ear (Scherg et al. 1989; Giard et al. 1990; Paavilainen et al. 1991; Grimm et al. 2006). Similarly, sources of Nbm were lateralized to the right hemisphere. This might point to the fact that the MLR change-related enhancement is an upstream index of contextual deviance processing preceding MMNm. Nevertheless, a complementary explanation based either on the greater spectral variation sensitivity of the right hemisphere for pure-tone processing (Liegeois-Chauvel et al. 2001; Zatorre and Belin 2001; Jamison et al. 2006) or methodological differences cannot be ruled out.
The arrangement of the distinct auditory regions found here is comparable with the structures described in previous studies using intracranial recording in humans (Howard et al. 2000; Brugge et al. 2003), showing a functional differentiation between the mesial-HG and the postero-lateral STG. In functional magnetic resonance imaging, a source location of the MMN posterior to the source of the P50 component has been shown (Mathiak et al. 2002). In accordance, our results suggest that generators of MMNm are partially located in secondary areas, whereas deviance-related Nbm arises from primary auditory areas (Borgmann et al. 2001).
Our source localization results extend the notion of a multistage organization of deviance detection (Boutros et al. 1995; Sonnadara et al. 2006; Slabu et al. 2010, 2012; Althen et al. 2011; Grimm and Escera 2011; Grimm et al. 2011, 2012; Leung et al. 2012) by showing differences in the anatomical domain. Our results suggest that the existence of a sequential and redundant transition of change-related activity occurs between ∼40 and ∼100 ms, spreading from anterior areas of HG to lateral and posterior areas of the auditory cortex. It is a widely held view that redundant information in the brain is efficiently reduced as stimuli are successively processed in different stations (Barlow 1961; Friston 2005). Deviance detection, as shown in the present study, fits well with the idea of hierarchically organized areas devoted to increasing levels of auditory regularity and acoustic violation processing (Grimm and Escera 2011). Only further studies using more complex levels of acoustic regularities will allow for disentangling a differentiation between the functional role of early and late change detection. Modulations of early auditory activity during oddball situations have been broadly interpreted as early indexes of deviance detection (Boutros et al. 1995; Sonnadara et al. 2006; Slabu et al. 2010, 2012). However, according to the lack of results delineating pure memory-based deviance detection processes, alternative interpretations might be taken into account. MLR amplitude enhancements could be indexing stimulus change per se, irrespective of the previously encoded regularity. Under this view, the early mechanism indexed by modulations of Nbm in primary areas would act as a stimulus detector signaling to higher order mechanisms devoted to memory-based change detection, the MMNm. This interpretation is in line with a study by Schönwiesner et al. (2007) which observed the sources of the MMN to duration changes located in lateral and medial portions of HG, PT, and along the STG and STS, whereas only STG and PT showed a deviance magnitude modulation. Based on that finding, the authors suggest that changes are initially detected in primary auditory areas, whereas STG and PT might be enrolled in a detailed analysis of acoustic changes. Similarly, Leung et al. (2012) suggested that, with the exception of frequency, deviance detection at early latencies is feature unspecific, whereas MMN indexed feature specific changes. Altogether, interpretations based on the different functional role of MMN and deviance-related MLR support the notion of deviance detection as a hierarchically organized network in the human brain Although the idea of an early noncomparator mechanism processing stimulus change per se is feasible, previous results showing enhanced responses to frequency deviants when compared with control sounds would not support this view (Slabu et al. 2010; Grimm et al. 2011; 2012). Similarly, the numerically largest deviant response found in the MLR time range of the present study argues in favor of a sequential activation of a change detection network.
Even though a direct functional relationship between change detection in the MLR and MMN time range cannot be drawn from the present study, the existence of 2 brain mechanisms of change detection operating at different spatial and temporal scales leads to the hypothesis that deviance processing might be framed under the same hierarchic model described for the auditory cortex of the macaque monkey (Kaas and Hackett 1998; Rauschecker 1998; Kaas et al. 1999). Single-cell and multiunit recordings in animals studying SSA have described an ubiquitous presence of neurons responding to deviant events along the auditory pathway, from anatomically low levels such as the inferior colliculi (Malmierca et al. 2009), the auditory thalamus (Antunes et al. 2010), and extending up to the PAC (Ulanovsky et al. 2003, 2004; von der Behrens et al. 2009). Although SSA has been regarded as the single-neuron correlate of the MMNm (Nelken and Ulanovsky 2007), remarkable differences contradict the assumption that the former directly accounts for the latter (von der Behrens et al. 2009). Instead, the most parsimonious explanation is that SSA in the PAC is reflecting a process “upstream” of MMN generation, that is to say, PAC first detects changes that subsequently would elicit MMN in higher areas. Since the sources of the deviance-elicited Nbm component described here (peaking at about 40 ms from stimulus onset) were estimated in, or near, the vicinity of PAC, our findings are not only suggestive of an upstream deviance detection mechanism, but provide anatomical evidence to support the hypothesis that deviance-related MLR may be the human analog of single-cell firing and local-field potential patterns found in PAC of animals (Ulanovsky et al. 2003, 2004; von der Behrens et al. 2009). This notion is further supported by the fact that both scalp-recorded MLR and single-neuron spikes share a common firing time interval, responding to deviant stimuli at about 20 ms from stimulus onset (Pérez-González et al. 2005; von der Behrens et al. 2009).
Differences between deviant minus control conditions yielded no significant effects, neither for MLR, nor for LLR. In this regard, it should be argued that we were not able to delineate pure memory-based deviance detection processes from contributions of a differential state of refractoriness (Jacobsen et al. 2003). Despite the differential brain response to deviant tones and their homologous control tones is a well established finding in the EEG literature pointing to the memory comparison nature of the MMN (Schröger and Wolff 1996; Jacobsen et al. 2003; Slabu et al. 2010; Grimm et al. 2011, 2012), it is also acknowledged that the genuine comparison underestimates the MMN due to a higher refractory state of neurons responding to deviants compared with controls (Schröger 2007). The unobserved genuine MMNm effect could be related to the tendency of the control condition to “overcontrol”, yielding less adaptation to control tones when compared with deviants (Kujala et al. 2007; Schröger 2007; Taaseh et al. 2011). It could also be related to the extremely fast presentation rate yielding amplitude-diminished transient responses (May and Tiitinen 2010). In agreement with the notion that the traditional MMN comparison yields a combination of memory comparison-based and memory comparison-unrelated deviance-related effects (Jacobsen and Schröger 2001; Taaseh et al. 2011), our analyses on the sources of N1m showed a clear distinction between the localization of pseudo-Z maps for the different conditions. Our results suggest that neural generators for infrequent tones deviating in frequency were located in different positions of each axis when compared with repeating tones. The difference between neural generators for N1m to deviant and standards in the anterior–posterior axis is consistent with previous MEG studies arguing for the involvement of separate change-specific neural populations underlying MMNm (Hari et al. 1992; Tiitinen et al. 1993; Korzyukov et al. 1999). The different location of N1m in response to deviant tones when compared with standard tones supports the existence of separate change-specific neural populations giving rise to the MMN/MMNm (Korzyukov et al. 1999; Näätänen et al. 2005). The lack of controlled results poses interpretative limitations regarding the genuine nature of memory-based deviance detection reported here, since hypothesis based on combined contributions of adaptation and memory comparison cannot be ruled out. In the time range of the MLR, lack of deviant versus control effects might point to the existence of a stimulus detection mechanism that is not related to memory comparison, as suggested previously. In any case, evidence points to the areas underlying the aforementioned components being acting in a sequential fashion during auditory deviance detection.
To our knowledge, this is the first study that applied spatial filter source reconstruction to both early and late deviance-related responses. Results in the present study nicely delineated the sources of auditory deviance detection on right hemispheric auditory areas. The use of spatial filters to estimate auditory activity has been, from a theoretical and computational point of view, limited by the fact that high correlations between sources deteriorate the performance of beamformers by introducing distortion in time courses, spatial blurring of sources, and reduction in reconstructed source intensities (Van Veen et al. 1997; Sekihara et al. 2002). Yet, it has been shown that standard beamformers are able to adequately reconstruct the magnitude of simulated correlations up to a relatively high level of source correlation (μ = 0.7–0.8) and low signal-to-noise ratios (Sekihara et al. 2002). Belardinelli et al. (2012) drew the same conclusion by assessing the impact of correlation on real data. Authors showed that phantom generated parallel dipoles in the 2 hemispheres (as might be the case in our data) separated by 3 cm were just slightly distorted in location and still recognizable when correlation levels reached 0.55, but clearly more blurred, with a high correlation of 0.95. The use of common filters and non-averaged data for the covariance matrix computation makes high correlation between hemispheres less likely to occur in our data (Brookes et al. 2010).
In summary, we provide evidence for 2 temporally and spatially separated neuronal sources indexing auditory deviance detection in the human auditory cortex. Our findings strongly support the presence of a neuronal assembly in the PAC located medially and anteriorly to the source generators of MMN. The existence of hierarchically different areas responding to auditory deviances is highly suggestive of a large distributed cortical network devoted to the processing of rare and unexpected auditory events.
This work was supported by the program Consolider-Ingenio 2010 (grant number CDS2007-00012), the National Program for Fundamental Research (reference number PSI2009-08063) of the Spanish Ministry of Science and Innovation, the 2009SGR11 grant of the Generalitat de Catalunya, the ICREA Academia Distinguished Professorship awarded to Carles Escera, and the FPI grant (BES-2010-030160) of the Spanish Ministry of Science and Innovation.
The authors thank Joachim Gross, Gavin Paterson, and Jan-Mathijs Schoffelen for their contribution in the development of the coregistration algorithm. Conflict of Interest: None declared.