Mismatch negativity as a marker of music perception in individual cochlear implant users: A spike density component analysis study

(cid:1) MMN measurements predict individual cochlear implant (CI) users’ behavioral music perception. (cid:1) MMN is detected in fewer CI users when sound deviants are of smaller magnitude. (cid:1) New spike density component analysis method enables more accurate diagnostics than preceding state-of-the-art.


Introduction
A main goal for 30-37 % of cochlear implant (CI) users is to restore abilities to listen to music with the CI device (Kohlberg et al., 2014, Migirov et al., 2009). However, for CI users, music perception and enjoyment from music is more challenged and varied than in normally hearing individuals, despite music is a factor that often heightens the perceived quality of life, even with only residual hearing abilities (Dritsakis et al., 2017, Fuller et al., 2022, Lassaletta et al., 2007. Indeed, a study showed that 90 % of CI users are interested in rehabilitative programs for improving their music experience (Gfeller et al., 2019). Also, music stimuli can be applied to test general aspects of listeners' sound discrimination ability by using stimuli that most people find enjoyable (Dritsakis et al., 2017, Lassaletta et al., 2007. Therefore, it is pertinent to identify accurate diagnostic tools for testing the individual variation in CI users' music perception (Hahne et al., 2016, Koelsch et al., 2004, Petersen et al., 2020, Petersen et al., 2015, Sandmann et al., 2010, Timm et al., 2014, Torppa et al., 2012 Electrophysiological tests on auditory brain function provide relatively fast, objective, and non-invasive measures to complement behavioral tests and questionnaires. They are especially advantageous when verbal or written tests cannot be adopted such as for infants and young children (Cone-Wesson, 2003, Gibson et al., 2009, Gilley et al., 2008, Golding et al., 2007, Näätänen et al., 2012, Näätänen et al., 2017, Nelson et al., 2008, Ponton et al., 2000, Sharma et al., 2002, Silva et al., 2017, Skoe and Kraus, 2010, Wang et al., 2015, Zhang et al., 2011. Among those tests, the Wave V component of the auditory brainstem response (ABR) can be applied to predict the pure tone hearing threshold at specific frequencies (Gibson et al., 2009, Skoe and Kraus, 2010, Wang et al., 2015. However, when ABR tests are conducted with a hearing aid or CI device switched on, the sound stimulus is limited to a duration of only a few milliseconds, to avoid overlap between the electrical stimulation artifact from the hearing device and the measured wave V occurring at a latency of approximately 5 milliseconds.
Cortical auditory evoked potentials (CAEPs), such as the auditory thalamo-cortical P1 component, are becoming popular measures for testing hearing thresholds for temporally-spectrally complex sounds, which are relevant in everyday language learning and understanding (Gilley et al., 2008, Mehta et al., 2019, Sharma et al., 2002, Silva et al., 2017. The electrical stimulation artifacts from the hearing device can be supressed from the measures of the CAEPs, e.g., by means of independent component analysis (ICA) (Gilley et al., 2006). A limitation of the P1 test is that it is currently uncertain whether it can be applied to test listeners' ability to distinguish between sounds, which is an important prerequisite for understanding their meaning. Indeed, everyday sounds in music and language consist of continuous acoustical feature transitions, which the listener must be able to hear, discriminate, and learn (Iverson et al., 2006, Loizou et al., 1998, Shestakova et al., 2004.
For testing auditory discrimination in a clinical context, various CAEP responses have been considered, such as the thalamo-cortical middle latency Nb and Pb responses (Althen et al., 2016) and the cortical long latency P1 (Gilley et al., 2005, Munivrana andMildner, 2013), N1 (Finke et al., 2016, Munivrana and Mildner, 2013, Zhang et al., 2011, P2 (Han et al., 2016, Munivrana and Mildner, 2013, Zhang et al., 2011, mismatch negativity (MMN) (Näätänen et al., 2012, Näätänen et al., 2017, Ponton et al., 2000, P300 (Alniacik and Akdas, 2019, Finke et al., 2015, Groenen et al., 2001, Munivrana and Mildner, 2013, Van Yper et al., 2020, and N400 responses (Kallioinen et al., 2016). Among these, the MMN is a promising candidate, because the auditory MMN generally reflects discrimination ability for soundseither verbal, non-verbal, or musical -in health, pathology, and across the lifespan (Näätänen et al., 2017). Also, the MMN response does not require the listener to attend to the auditory stimuli or to perform a demanding task, which makes MMN suitable also for infants and young children in addition to adults. The auditory MMN is an automatic brain response to an infrequent and rare sound after a repeated and more expected standard one (Näätänen et al., 2017). The acoustical difference between the standard and deviant sound must be above the just-noticeably difference threshold to evoke an MMN (Näätänen et al., 2017, Rahne et al., 2014. For CI users, the MMN has been recorded to study even residual discrimination of spoken language (Kraus et al., 1993, Lonka et al., 2004, Singh et al., 2004, and musical sounds (Hahne et al., 2016, Koelsch et al., 2004, Lonka et al., 2013, Petersen et al., 2020, Petersen et al., 2015, Rahne et al., 2014, Sandmann et al., 2010, Timm et al., 2014, Torppa et al., 2018, Torppa et al., 2014, Torppa et al., 2012, Wable et al., 2000, Zhang et al., 2013. The musical multi-feature (MuMuFe) paradigm (Vuust et al., 2011) was devel-oped for investigating MMN responses to sound differences in spectral-temporal features presented in a musical context. With the latest development of the MuMuFe CI paradigm, statistically significant MMN responses were successfully obtained in a group of CI users for all tested features (intensity, pitch, timbre, rhythm) on at least one out of four deviant magnitude levels (Petersen et al., 2020). These findings support that MMN is a promising marker of music discrimination in CI users.
The last 30 years of research on MMN responses in individual listeners has, however, not reached sufficient accuracy for clinical application (Bishop and Hardiman, 2010, Dalebout and Fox, 2001, Escera and Grau, 1996, McGee et al., 1997, Näätänen et al., 2017, Paukkunen et al., 2011, Pekkonen et al., 1995, Uwer and von Suchodoletz, 2000 despite of the implementation of paradigms aimed at clinical practice such as the multi-feature versions. For instance, the least accurate MMN detection performance has been observed with visual inspection by EEG experts, where confounding cortical sources, such as alpha waves, have been mistakenly identified as MMN in 63 % cases when no MMN was present (McGee et al., 1997), which means that more than every second patient would incorrectly pass MMN testing without existing sound discrimination problems being identified. Proposals have been made to improve the accuracy of the individual MMN analysis (Bishop and Hardiman, 2010, Dalebout and Fox, 2001, Kalyakin et al., 2009, McGee et al., 1997, Näätänen et al., 2017, Ponton et al., 1997, Rahne et al., 2008, Torppa et al., 2012. The most accurate results have been achieved with individual-level statistics, with 88.2 % correct detections of MMN absence for indiscriminable sounds and 82.4 % correct detections of MMN presence for discriminable sounds (Bishop and Hardiman, 2010). A main challenge with measurement of individual evoked responses is the confounding cortical sources (Scharf et al., 2022), such as alpha waves, which also interfere with the individual MMN measurements .
In this study, we test whether the issues specified above, which all limit the accuracy of the individual MMN and its predictive value in assessing the individual music discrimination skills in CI users, might be overcome by our very recently method that allows decomposing the overlapping large-scale cortical activity with spike density component analysis (SCA) . Empirical and simulation studies have consistently confirmed that SCA is accurate in separating largescale bioelectromagnetic cortical activity into sparse temporal shapes with distinct scalp topographies from both adults  and children (Bruzzone et al., 2021). Previous studies on normally hearing young adults and computersimulated EEG  have shown that SCA is more accurate than spatial principal component analysis (PCA) and ICA in isolating individual MMN from confounding cortical interfering sources, such as alpha waves.
In sum, the aim of the present study is to obtain accurate individual markers of central auditory function in CI users that could predict their individual behavioral music discrimination thresholds. We hypothesized that statistical detection of the individual MMN response would predict individual behavioral music discrimination ability. This was followed up with supplementary assessments of the validity and precision of the hypothesized MMNbehavior relationship. Also, since smaller deviant magnitudes should extend below the discrimination thresholds of individual listeners, we tested whether MMN was detected in fewer individuals when the deviant magnitudes were smaller. Moreover, we assumed that the common dissimilarity between CI users and NH groups might be observed for pitch that is most technically challenging in CI listening. All these findings were obtained with a new SCA statistics procedure, therefore, it was relevant to prove that each option in this new procedure was optimal in terms of accuracy and replicability (see the Methods section and Appendices for further details).

Participants
This study is based on a re-analysis of 312 cases of averaged MMN waveforms from repeatedly measured cochlear implant (CI) users and 424 cases from normally hearing controls, previously introduced in Petersen et al. (2020). A group of eleven experienced older CI users with an average of 7.0 years CI experience (range 1-14 years) and average age of 56.1 years (range 34-77 years) (nine women) participated in the study. Nine CI users had a high CI outcome with reported ability to speak on the phone (for individual details, see Table 1). The average duration of deafness prior to receiving the CI was 23.6 years (range 0 to 56), nine had one CI implant in an ear (unilateral implant) and two in both ears (bilateral), and six used the left ear for testing whereas five used the right ear. Four CI users in addition used a hearing aid in the ear opposite to the ear with the implant. As a control group participated fourteen older participants with normal hearing and an average age of 63.4 years (range 56-77), similar to the age of the CI group. None of the participants had an academic background in music, and they reported not to be professional or amateur musicians and to have less than 5 years formal singing or instrument training. The study was conducted in accordance with the Helsinki declaration and approved by the Research Ethics Committee of the Central Denmark Region. All participants provided written informed consent.

Stimuli
The cochlear implant musical multi-feature (CI MuMuFe) mismatch negativity (MMN) paradigm (Petersen et al., 2020) was based on a four-tone Alberti bass melody played in four different keys (C, Eb, Gb, and A) (listen to an excerpt of the stimuli in the Supplementary Material audio file). It contained tones in the middle register range from Ab3 (208 Hz) to E5 (659 Hz). All tones had a duration of 200 ms (except rhythm deviants), a rise and fall time of 18 ms, and were followed by a silent interstimulus interval of 5 ms. MMN responses to tones were tested on four fundamental features in music: intensity, pitch, timbre, and rhythm. Deviant tones evoking MMN for each sound feature were inserted at the third tone in the Alberti Bass pattern with one out of four deviant magnitudes: extra-large (XL), large (L), medium (M), or small (S). MMN to intensity deviants were created by decreasing the intensity of the tone by 12 (XL), 9 (L), 6 (M), or 3 (S) decibels (dB). Pitch deviants were inserted by lowering the pitch by eight semitones (-37 % Hz) (XL), three semitones (-16 % Hz) (L), two semitones (-11 % Hz) (M), or one semitone (-6% Hz) (S). Timbre deviants were created with an exchange of a regular piano sound to a guitar (XL), trumpet (L), blues piano (M), or bright piano (S) sound. Finally, rhythm deviants were inserted as earlier onsets of the third note by 155 ms (XL), 103 ms (L), 52 ms (M), or 26 ms (S), (while at the same time shortening the second note to avoid sound overlap and lengthening the third note accordingly to avoid a silent gap).
The sixteen deviants occurred with equal probability (6 %) on the third tone in the repeated melody. All deviant tones and changes of key occurred in a pseudo-randomized order, where the same deviant or key was not repeated consecutively. (The change of key occurred after three repetitions of each of the 16 deviant conditions, and the first tone at each key change was excluded from subsequent analysis.) To reduce the required EEG recording time by 50 %, the so-called ''no-standard" paradigm uses a common and well-tested procedure of applying the repeated tones 1, 2, and 4 in the Alberti Bass pattern as standard tones, as an alternative to standard tones on the third tone in every second tone-pattern, with either option resulting in similar MMN difference waveforms (e.g., Bonetti et al., 2017, Kliuchko et al., 2019, Kliuchko et al., 2016, Pakarinen et al., 2010. In total there was 2304 stimulus trains (i.e., groupings of four tones) and 48 key changes.

EEG procedure
The experiment was conducted at the EEG lab facilities of Aarhus University Hospital. EEG was recorded at a sampling rate of 1000 Hz in an electrically and acoustically shielded room with a BrainAmp amplifier system (Brain Products, Gilching, Germany). Electrodes were placed in a 32-electrode cap according to the international 10/20 system. It was ensured that electrode impedances were < 25 kX. An electrooculogram (EOG) was recorded with electrodes beside and above the left eye. FCz was applied as an initial reference electrode. The auditory stimuli were played in mono at a 44.1 kHz sampling rate. During EEG recordings participants were instructed to ignore the auditory stimuli and focus on a movie in which the audio was muted. For all participants, the sound level was individually adjusted to a comfortable level from a defined starting point of 65 dB SPL. The total EEG recording time was approximately 35 min.

EEG preprocessing
The EEG data were preprocessed as in Petersen et al. (2020) by using the FieldTrip Toolbox for Matlab (Oostenveld et al., 2011) for 1 Hz high-pass, 25 Hz lowpass filtering, independent component analysis (ICA) for suppression of eye movement and cochlear  (Näätänen et al., 2017). We followed a common procedure of analysing difference waveforms at the individual level for automatized MMN detection (Bishop andHardiman, 2010, McGee et al., 1997) by subtracting the mean standard waveform across trials from each deviant trial to isolate the individual MMN at the single trial-level.

Behavioral auditory discrimination test
In addition to the individual MMN testing, all participants completed a three-alternative forced choice test, where the task was to identify which one out of sets of three melodic patterns contained a sound deviant (Petersen et al., 2020). This test resulted in scores of 0 %, 17 %, 33 %, 50 %, 67 %, 83 %, or 100 % correct answers (with chance level at 33.3 %), where higher scores reflect more reliable sound discrimination. This behavioral test reflects the sound discrimination ability for the same music parameters and levels of magnitude as presented in the MMN paradigm measured with EEG.

Individual-level EEG statistics
For clinical applications, it would be important to statistically determine whether the individual MMN can be interpreted as absent (i.e., either absence or presence interpretation in, e.g., Carter et al., 2010, Golding et al., 2007, because this can indicate a hearing impairment and a need to refer a patient to further testing or treatment (cf. Kuki et al., 2013, Schmidt et al., 2001. If the individual MMN is interpreted as absent for a sound deviance that would trigger a measureable MMN in a normally hearing population (e.g., MMN to a pitch deviant at a level of 8 semitones difference), then the individual MMN could be applied as a diagnostic tool indicating an auditory discrimination disorder (e.g., impaired pitch discrimination). Statistical detection of individual MMN presence would be less clinically important because it means that a patient would pass MMN diagnostics, indicating an ability to distinguish between specific sounds. Therefore, we define the test sensitivity as the percentage correct interpretation of individual MMN absence for indistinguishable sounds (cf. Kuki et al., 2013, Schmidt et al., 2001. We define the test specificity as the percentage correct detection of individual MMN presence for distinguishable sounds (cf. Kuki et al., 2013, Schmidt et al., 2001. Whether the individual MMN was interpreted as either absent or present was automatically determined by an SCA statistics algorithm described below.

Spike density component analysis (SCA) decomposition
As a first step, in order to suppress confounding alpha waves disturbing the MMN waveform, the individual averaged difference waveforms were decomposed into spike density component analysis (SCA) components   (Fig. 1 A) (FieldTrip compatible Matlab functions for SCA decomposition and SCA statistics are available at https://github.com/nielsthaumann/sca). SCA is a type of fast, automatic, non-supervised machine learning method, which optimizes a model predicting the individual bioelectromagnetic activity of large-scale cortical neural assemblies from individual EEG or MEG measurements. SCA implements a constraint based on the assumption that the bioelectromagnetic source of a ''spike density", or neural current, generated in a large-scale cortical neural assembly ($10,000 neurons) approximates a Gaussian shape over time (the term ''spike density" is derived from the ''spike train" concept for describing the peristimulus time histograms measured on single neurons). It was verified that the Gaussian temporal shapes explain most variance in scalprecorded EEG and MEG, whereas sine, gamma, and asymmetrical Gaussian halves shapes explained significantly less variance, and it was confirmed that the Gaussian shapes specifically modeled most of the neural activity originating within the scalp and could not model non-neural external signals, such as eye movement artifacts . It should also be noted that the SCA method assumes that the signal of interest (i.e., the MMN) shows one or more peaks in time at a signal-to-interference-and-noise ratio larger than 1 (i.e., in at least one electrode at one time point the MMN should have larger amplitude compared to confounding alpha waves) . These assumptions seem realistic in the present study for the separation of MMN from confounding alpha waves in pre-processed averaged EEG waveforms, because the EEG is time-locked at tone stimulus onsets and thereby confounding alpha waves are partially distributed in phase and suppressed in amplitude.

Region of interest constraints
To ensure that the individual pre-attentive MMN was detected, and not other clearly differing responses such as following P3(a,b) responses, and also to ensure fast computational processing of the statistical analyses, we applied region of interest (ROI) constraints (Bruzzone et al., 2021). Only SCA components with a negative peak in the frontal half of the EEG channels within a relevant MMN latency range were tested statistically (Fig. 1B). Since the MMN latency is more varied in patients, older listeners, and children than in normally hearing young adults, it is common to apply a relatively broad individual MMN latency time-window ranging approximately between 0-450 ms in clinical individual MMN literature (Dalebout and Fox, 2001, Escera and Grau, 1996, McGee et al., 1997, Näätänen et al., 2017, Ponton et al., 1997, Taylor et al., 2017. Here we constrained the MMN peak latency to between 75-300 ms.

Trial-level spatiotemporal SCA filter
The inter-trial variance originating from interfering SCA components and EEG channel noise was reduced by using the SCA component as a spatiotemporal filter (cf. Li et al., 2008) (Fig. 1C). The spatiotemporal filter was applied to each trial by multiplying the trial by the tested SCA component to obtain a weighted average, which reduced signals unrelated to the tested SCA component: (1)where trial f is the filtered multichannel trial waveform, trial is the multichannel trial waveform, W i is the component channel weights, and x i is the com- The formula ensured that the tested SCA component amplitude was always positive-valued. (Also, filtering an ideal trial waveform without cross-trial variance, trial = W i x i , results in the same (positive-valued) output as the input, trial f =| trial|.) Since the SCA component amplitude is always positivevalued after the spatiotemporal filtering, an area measure was derived for each trial simply as the summed amplitude over all channels and time samples (for more details, see Appendix B). The distribution of area measures across trials was then submitted to individual-level SCA statistics.

Individual-level SCA statistics
In previous work, the SCA components were considered significant, if they correlated with a specific template, e.g., a grandaverage template . However, the template match method assumed that the individual latency or topography did not deviate significantly from the specific template, which is violated in heterogenous samples, particularly in developmental or clinical samples. These limitations were solved by proposing that each SCA component could be statistically tested based on the individual-level trial distribution (Bruzzone et al., 2021). For the present study, we solved two remaining issues. First, we introduce a simple and standardized significance p-threshold, instead of a first significance threshold for SCA component correlation with sub-averages, and a second significance threshold for determining SCA component consistency across sub-averages (Bruzzone et al., 2021). Second, we increase the computational speed by applying one statistical test on each SCA component, instead of previous 100 statistical tests for each SCA component's correlation with 100 sub-averages. For statistical inference, the distribution of the SCA component area measures across trials was tested with a one-sample t-test (Fig. 1D). If the t-test resulted in a p-value smaller than the p-threshold, then the null-hypothesis that the SCA component is part of the error distribution centred on the baseline at 0 lV was rejected (Fig. 1E), otherwise, the null-hypothesis was retained.
To verify the effects of choosing the proposed optimal SCA statistics on the diagnostic accuracy and the control of the replicability of the individual-level results, the tailoring of the individuallevel SCA statistics was compared to common nonoptimal options. Choosing common but nonoptimal options led to significantly lower diagnostic accuracy and lower control of the replicability of the individual-level results (all methodological comparisons and results are explained in detail in Appendix C).
Finally, the state-of-the-art ''t-cluster" method was tested. For valid comparison with SCA statistics, the same ROI constraints (section 2.6.2) were applied for t-clusters as for SCA statistics. At each time sample, one-sample t-tests were conducted across trials, and the t-cluster p-values were obtained as the maximum p-value over 32 ms consecutive time samples (8 at 250 Hz) (Bishop and Hardiman, 2010). If the p-value remained smaller than the pthreshold across 32 ms consecutive time samples, the t-cluster was interpreted as significant, otherwise non-significant (Bishop and Hardiman, 2010).

Statistical evaluation of the prediction of behavioral auditory discrimination based on the individual MMN
All subsequent statistical analyses were conducted with the IBM SPSS v27 software package (IBM, Armonk, New York, USA). The CI and the NH groups were investigated separately. As is common in classification studies where the EEG is applied to predict a certain behavioral outcome, the repeated measures for the 16 stimulus conditions from each participant were treated as independent observations, where each observation depends on noise in the EEG trials and in the behavioral tests (e.g., Amin et al., 2017, Golz et al., 2016, Perez-Valero et al., 2021. The advantage of this procedure is that the evaluations reflect the performance of the individual-level MMN test on more deviant features and magnitudes, as well as on measures from different participants, which is relevant for clinical applications with individual CI users where consistent individual-level results are important. 2.7.1. MMN as a marker of auditory discrimination thresholds in individual CI users 2.7.1.1. Diagnostic accuracy. The diagnostic accuracy was analyzed with a receiver operating characteristic (ROC) curve based on varying the p-threshold on 100 equal logarithmic steps from 10 -34 to 1. The test sensitivity in the ROC analysis was based on the ''false MMN" with indistinguishable sounds. The sensitivity was defined as the percentage correct retainment of the null-hypothesis, meaning that there was no individual MMN, and indicating inability to discriminate between the sounds. Also, an individual should be able to pass the MMN testing with an outcome that does not motivate for treatment. This is indicated by the test specificity, which was estimated as the percentage correct individual MMN detections for the cases where the individual study participants scored above chance level on the behavioral sound discrimination test. The balanced accuracy was calculated as the sum of the sensitivity and specificity divided by two. The area under the curve for the sensitivity and specificity was compared to the random chance level diagnostic outcome with the Wilcoxon statistic as implemented in SPSS (Coelho and Braga, 2015). The ROC curves were visualized with bias-corrected and accelerated 95 % confidence intervals with the Matlab function fitglm (using the binomial distribution logistic regression option) and the Matlab function perfcurve (applying 1000 bootstrap permutations).
2.7.1.2. Validity. Supplementary to the main hypothesis on diagnostic accuracy, the validity and precision of the individual MMN-behavior relationships were further assessed. Initial inspection suggested that the assumption of a continuous (linear or nonlinear) relationship between the individual MMN peak amplitude measured within the region of interest (defined above) and the behavioral hit rate was violated. Therefore, it was considered inappropriate to model a unit-wise change in the MMN amplitude in lV per change in the behavioral score in percentage hit rate.
Instead, we tested whether the individual MMN amplitude was lower when the behavioral discrimination was not detected (at chance level or lower) compared to detected (above chance level). Due to unequal sample-sizes, this comparison was tested with the Welch's test. A comparison was also made with the Welch's test to assess whether the individual MMN peak latency measured within the region of interest (defined above) was longer for absent than detected behavioral discrimination. Further, the behavioral hit rate might be lower when the individual MMN was not detected compared to detected, also assessed with the Welch's test.
The above assessments pointed towards the common observation that the individual MMN can be detected when behavioral discrimination is absent (Näätänen et al., 2017), and behavioral detection can also be present when individual MMN is not detected (Bishop and Hardiman, 2010). It was investigated whether the common occasional MMN-behavior dissociations might indicate different discrimination thresholds between the MMN and behavioral tests. This assumption was tested with Pearson's chi-squared tests comparing the percentage of individual MMN detections against behavioral detections for each group and each stimulus condition.

Precision.
To further clarify the reason of the occasional MMN-behavior dissociations, logistic regression was applied to test a measurement error hypothesis: that the percentage missing individual MMN or behavioral detections would increase when the deviant magnitudes decrease. I.e., when the deviant magnitude approaches the individual discrimination threshold more measurement errors would be expected, which would be indicated by more frequent incongruent individual MMN compared with behavioral test outcomes (where either individual MMN or behavioral discrimination is detected but not both). Finally, it was evaluated whether the MMN and behavioral tests were equally precise. This was assessed with the Mann-Whitney U test for inferring whether the number of missing individual MMN in comparison to behavioral detections were equally distributed over the deviant magnitudes.
2.7.1.4. Correlation with music appreciation and clinical factors. Relationships between the individual MMN amplitude and latency at the Fz electrode and music appreciation and clinical factors were explored with Pearson's product moment correlations. Statistical inferences were derived for the following expected relational directions: higher individual MMN amplitude and faster individual MMN latency for more rated enjoyment of music, hours of music listening, and knowledge of music, higher rated quality of musical sounds with the CI, younger age, fewer years of deafness prior to the CI implantation, and more years of CI experience. Each of the 16 MMNs (four features at four magnitudes) were tested for each individual CI users. The results were not corrected for multiple comparisons and only regarded as non-conclusive inspiration for future studies.

Music discrimination thresholds for individual experienced CI users
To further substantiate the main hypothesis on the relationship between individual MMN and behavioral discrimination, we verified that the individual music discrimination threshold was passed when the deviant magnitudes decreased. For all features, logistic regression was applied to test whether the percentage of individuals showing individual MMN decreased as expected when the deviant magnitude decreased. A significant result indicated that the deviant magnitude crossed from above to below the individual's discrimination threshold.
Additionally, we attempted to replicate the common grouplevel findings of similar intensity, timbre, and rhythm, but not pitch discrimination, between individual CI users and normally hearing. The percentage individuals showing MMN at each deviant magnitude was compared between the CI and normally hearing groups. Since the expected cell count tended to be less than 5 in the contingency tables, the Fisher's exact test was applied for the group comparisons instead of the Pearson's chi-squared test. Bonferroni correction at p =.05/4 = 0.0125 was applied to interpret the significance of the planned group comparisons with tests on each of the four deviant magnitudes.

SCA statistics compared with preceding state-of-the-art
Finally, it was assessed whether SCA statistics showed higher diagnostic accuracy compared to the preceding state-of-the-art. For signal-to-interference-and-noise ratios (SNIRs) in the range of approximately 1-10, the SCA method is known to show higher gains in accuracy at lower SNIRs . Therefore, for each individual MMN test, a SNIR measure was estimated based on the highest valued t-statistic from the t-tests applied in the SCA statistics procedure. This SNIR estimate is a unitless quantity reflecting the mean amplitude divided by the standard error of the mean amplitude across trials. With the Welch' test it was investigated whether the SNIR was lower in the CI users compared to the NH. For the CI and NH groups, by applying the Wilcoxon test on the ROC areas under the curves it was tested whether the diagnostic accuracy was higher with the proposed SCA statistics method compared to the state-of-the-art ''t-cluster" method (Bishop and Hardiman, 2010).

MMN as a marker of auditory discrimination thresholds in individual cochlear implant users
The behavioral discrimination thresholds were predicted with high accuracy by the individual MMN detections, and the discrimination thresholds estimated with the pre-attentive individual MMN and attentive behavioral tests were similar (Fig. 2). Occasional differences between the individual MMN and behavioral results did not indicate functional differences and were mainly explainable by measurement errors (Fig. 2B-C). For the adult experienced cochlear implant (CI) users and normally hearing (NH) controls the occasional measurements errors were similar between the individual MMN and the behavioral tests. This suggests that the individual MMN is an accurate marker of auditory discrimination thresholds.

Diagnostic accuracy
The individual CI users' ability (or inability) to discriminate between the tones was predicted with high accuracy by the statistically significant (or nonsignificant) individual MMN. The ROC analysis showed that the automatic individual MMN detection is a highly accurate marker of individual CI users' auditory discrimination ability ( Fig. 2A), AUC = 0.96 (95 % CI: [0.94, 0.98]), and is significantly more accurate than the 50 % chance level (AUC = 0.50), p <.001. Based on the ROC analysis, we recommend applying a theoretically expected confidence level of 99 % for the sensitivity (see the Appendix D, Figure D1

Validity
Relationships between MMN detection and behavioral detection: When the behavioral test indicated ability to discriminate between the tones compared to no discrimination ability, the group-average MMN amplitude was significantly higher at the Fz electrode in the  (Fig. 2B). When the individual MMN was detected compared to interpreted as absent, the group-average behavioral hit rate showed a tendency of higher scores for the CI users, M diff =+11.2 %, 95 %CI: [-1.0, 23.4], t(38.5) = 1.9, p =.072, d = 0.4, and was significantly higher for the NH controls, M diff =+18.4 %, 95 %CI:[6.5, 30.2], t(28.8) = 3.2, p =.004, d = 0.8 (Fig. 2B). These results further substantiates that the individual MMN is a valid marker of behavioral auditory discrimination.
Similar MMN and behavioral discrimination thresholds: A relevant question was whether the typical, occasional double dissociations between the individual MMN and behavioral test outcomes (Fig. 2B) were caused by functional differences in the underlying auditory discrimination ability. When comparing the total number of discrimination detections achieved with the MMN to the behav-ioral tests (Table 2) for each group, deviant feature, and deviant magnitude, no significant differences were found between the MMN and behavioral tests (Appendix D, Table D2). These results suggest that the individual MMN and behavioral tests reflect similar auditory discrimination thresholds.

Precision
Behavioral detections were occasionally missing when individual MMN was detected above the ''false MMN" detection rate (Fig. 2B). Vice versa, occasionally (on average 14.7 % (20/136) cases for CI users and 10.0 % (20/200) cases for NH controls), individual MMN was missing when behavioral discrimination was detected above chance level (Fig. 2B). As expected from the measurement error hypothesis, lower deviant magnitudes were significantly related to a higher percentage of missing behavioral or MMN detections (Fig. 2C) Table D1). The deviant magnitudes where detections were generally missing were not significantly different between MMN and behavioral tests (Fig. 2C) for the CI users, U = 348.0, n = 52, p =.580, r = 0.08, and the NH controls, U = 124.5, n = 39, p =.065, r = 0.31, which suggests that the MMN and behavioral tests were similarly precise.

Exploratory correlations with music appreciation and clinical factors
With regards to music appreciation, CI users providing higher ratings of enjoyment of listening to music tended to show faster individual MMN latency for the S (1-semitone) pitch deviant (scatter plots are shown in Appendix D, Figure D2), r(6) = -0.91, p =.011. Also, CI users providing higher ratings of sound quality tended to show higher individual MMN amplitudes (i.e., more negativity at Fz) for the S (bright piano), r(11) = -0.77, p =.006, and the XL (guitar), r(11) = -0.69, p =.019, timbre deviants (Appendix D, Figure D2).
With respect to clinical factors, fewer years of deafness prior the CI implantation tended to be related to higher individual MMN amplitude, r(11) = 0.68, p =.020, and faster individual MMN latency, r(11) = 0.65, p =.031, for the L rhythm deviant (Appendix D, Figure D3).
No additional significant correlations were observed between individual MMN amplitude or latency and music appreciation and clinical factors. The p-values concerning music appreciation and clinical factors were not corrected for multiple comparisons, and the observed significant tendencies should be interpreted with caution.

Music discrimination thresholds for individual experienced CI users
For the experienced CI users, the tested intensity and pitch deviant magnitudes were crossing 50 %-64 % of the participants' discrimination thresholds, whereas the tested small timbre and rhythm deviants remained above most (91 %-100 %) of the experienced CI participants' discrimination thresholds ( Table 2). As Table 2 Percentage individuals with MMN detected and behavioral discrimination. Percentage individual MMN detections in the CI user group and the normally hearing (NH) group is shown for each deviant feature and deviant magnitude, indicated by statistically significant spike density component analysis (SCA) components within the ROI at p <.01 (with Benjamini-Yekutieli False Discovery Rate correction for multiple SCA component testing). ns. means no significant SCA components within the region of interest. Also is shown the percentage individuals with behavioral discrimination, based on individual scores above chance level (>33 % correct).  expected, the experienced CI users showed similar music discrimination thresholds compared to NH controls, except for the pitch discrimination thresholds that were higher for the experienced CI users in comparison to the NH (Table 2).

Intensity
All experienced CI users showed significant individual MMN responses to the extra-large (XL) intensity deviant (Table 2) Comparisons between the experienced CI and NH groups showed that the percentage individual MMN detection was the same for the XL intensity deviant and not significantly different at the large (L), medium (M), and small (S) deviant magnitudes (  Table E1).

Pitch
Among the experienced CI users, 91 % showed significant individual MMN responses to the XL pitch deviant (Table 2) Comparisons between the experienced CI and NH control groups showed that the percentage of individual MMN detection was not significantly different at the XL, L, and M deviant magnitudes, though, at the S deviant magnitude, significantly fewer CI users (55 %) than NH controls (100 %) showed individual MMN (Table 3) (Bonferroni corrected significance level for the four planned comparisons: p =.05/4 = 0.0125). This means that, according to the individual MMN results, the experienced CI users' pitch discrimination thresholds were often between 1-2 semitones for pitch decrease (between 6 % and 11 % decrease in Hz), whereas determining the discrimination thresholds for the NH required smaller intervals than the smallest tested 1 semitone pitch decrease. The behavioral discrimination test showed similar tendencies, which, however, did not reach statistical significance (Appendix E, Table E1).

Timbre
All experienced CI users showed significant individual MMN responses to the XL timbre deviant (  Table E1).

Rhythm
Among the experienced CI users, 91 % showed significant individual MMN responses to the XL rhythm deviant (Table 2) Comparisons between the experienced CI and NH control groups showed that the percentage of individual MMN detection was not significantly different at the XL, the same at the L deviant,  and not significantly different at the M and S deviant magnitudes (  Table E1).

Discussion
In this study, we investigated whether the statistical detection of individual MMN responses is an accurate neurobiological marker of behavioral individual music discrimination thresholds of cochlear implant (CI) users. Automatic MMN detections for individual CI users showed a high diagnostic balanced accuracy (BAC) of 89.2 % and a similarly high BAC of 90.5 % for normally hearing (NH) controls. For each deviant magnitude and feature, the number of participants showing no MMN detection and no behavioral discrimination above chance level did not differ significantly within the CI and the NH groups, suggesting that the sound discrimination thresholds estimated with the individual MMN and the behavioral tests were similar. Also, there was lower congruency between the MMN and behavioral test results for smaller deviant magnitudes. Though, for each deviant magnitude, auditory feature, and participant group, there were no significant differences between the number of participants with missing MMN (when behavioral discrimination was detected) compared to the number of participants with missing behavioral discrimination (when MMN was detected), suggesting that the MMN and behavioral tests were equally precise for estimating the individual sound discrimination thresholds. These findings are particularly promising, e.g., for applying the individual MMN to test the development of individual pediatric CI users, who can complete MMN tests (Torppa et al., 2018, Torppa et al., 2014, Torppa et al., 2012 but commonly experience difficulties with completing behavioral tests (e.g., Kuki et al., 2013, Kumari et al., 2016. All these findings were partly achieved with an optimized, semi-automatic, and easy to inspect (by suppressing interfering cortical signals from the EEG images) spike density component analysis (SCA) statistics method, which showed higher diagnostic accuracy (BAC = 89.2 %) compared to the preceding state-of-the-art on individual-level EEG statistics methods (BAC = 70.8 %-85.3 %) (Bishop andHardiman, 2010, McGee et al., 1997).
Altogether, the improved accuracy in measuring individual MMN responses opens new avenues for testing the progress of the individual CI user with rehabilitation interventions, such as the effects of music training even suggesting improved speech processing in pediatric CI users related to music training (Petersen et al., 2015, Torppa et al., 2018. The behavioral testing of sound discrimination is based on slower attentive responses compared to the individual MMN, and contextual information for sounds both preceding and succeeding the deviant sound can influence the behavioral response. By contrast, the individual MMN response is a faster, automatic, preattentive, neural response compared to the behavioral response, and the individual MMN follows immediately after the onset of the deviant sound (Näätänen et al., 2017). Despite these differences in processing stages, the present results did not suggest functional differences between the individual MMN detection and the behavioral discrimination. While evidence for a correspondence between the MMN and behavioral auditory discrimination exists for groups of CI listeners (e.g., Rahne et al., 2014), we are not aware of previous studies showing a direct relationship between MMN and behavioral sound discrimination ability at the individual level in both normally hearing listeners and CI users. Moreover, the individual discrimination thresholds were similar, regardless of whether they were estimated with the MMN or a behavioral test. These findings mutually support that the individual MMN is an accurate marker of auditory discrimination thresholds in individual CI users. The findings also suggest that incongruent MMN and behavioral outcomes are not always explainable by functional differences, also, measurement errors near the discrimination thresholds and suprathreshold ceiling effects on the behavioral hit rates should be taken into consideration.

The influence of measurement errors on MMN and behavioral tests
We observed that the individual MMN and behavioral test results were more disturbed by measurement errors at the smaller deviant magnitudes near the discrimination thresholds compared to the larger deviant magnitudes well above the discrimination thresholds. For the here tested adult CI user and NH groups we found that the measurement errors were similar, regardless of whether the MMN or the behavioral tests were applied. Though, the sources of the measurement errors are known to differ between the MMN and the behavioral test. While the individual MMN is mainly masked by neural interferences from ongoing background brain processes (e.g., Haumann et al., 2020), the behavioral test is typically confounded by attentive, cognitive, or motoric disabilities (e.g., Norrix, 2015). Infants and children are known to elicit measurable MMN responses to deviant sounds (Fellman andHuotilainen, 2006, Torppa et al., 2012), whereas measurement errors seem to be more limiting behavioral testing of infants and children (Kuki et al., 2013, Norrix, 2015 compared to, e.g., the adult groups in the present study. A study found that behavioral observation audiometry testing for infants reached a sensitivity of 94.2 % but only a specificity of 67.7 % (with auditory brainstem response applied as the correct reference test). This means that approximately every third infant was incorrectly diag-nosed as having a hearing disorder, due to measurement errors in the behavioral test (Kuki et al., 2013). One advantage of the individual MMN is that it is measurable without requiring the attentive, cognitive, or behavioral abilities needed for behavioral testing (Näätänen et al., 2017). Thereby, the individual MMN could potentially facilitate better auditory discrimination tests for CI users with attentional, cognitive, or motor impairments. Moreover, the individual MMN might be considered more ecologically valid than behavioral testing, because the MMN indicates that sound discrimination ability has been established at an auditory processing level where attentive and cognitive effort is minimal, which is an important aspect for diagnostics of hearing performance in realistic listening situations, in which minimization of required attention and cognitive effort is desirable Pisoni, 2013, Perreau et al., 2017). Another potential advantage might be a possibly higher sensitivity of the individual MMN to listening effort than behavioral measures. For instance, if the fast early-stage MMN is missing for a deviant stimulus, suggesting that the auditory discrimination has not been established pre-attentively, behavioral discrimination might still be detected based on slower later-stage auditory processes requiring more listening effort.

Suprathreshold ceiling effects on behavioral hit rates but not on MMN amplitudes
In the present study we investigated binary diagnostic outcomes, where the individual MMN was either detected or not, and we found that the individual MMN detection was significantly related to sound discrimination being present or not. An outstanding question is whether CI users perceive larger deviant magnitudes as larger compared to smaller deviant magnitudes, and whether these perceptual magnitude differences might be reflected by the individual MMN amplitudes. For this purpose, the traditional application of behavioral hit rate measures in MMN studies (Bishop and Hardiman, 2010, Petersen et al., 2015, Ponton et al., 2000, Rahne et al., 2014, Sandmann et al., 2010, Timm et al., 2014, Torppa et al., 2018 might maintain a validity issue, because the hit rate indicates the probability of perceiving a difference, and it has a noticeable ceiling effect for suprathreshold stimuli (due to the ''S"-shaped relationship between the deviant magnitude and the discrimination probability) (Grondin, 2016). By contrast, the MMN amplitude increases continuously for deviant magnitudes above the discrimination threshold (Näätänen et al., 2017). In a group-level analysis on the same dataset analysed in the present study (Petersen et al., 2020), suprathreshold ceiling effects were evident in behavioral hit rates, whereas the MMN amplitudes continued increasing for deviant magnitudes above the discrimination threshold (especially for the NH pitch, timbre, and rhythm results). Also, the suprathreshold ceiling effect appears to be present in the individual-level results (compare Fig. 3 to Appendix Table D1). In future studies, a behavioral measure of perceptual magnitude, e.g., a Likert or visual analogue scale, might be more comparable to the MMN amplitudes than hit rates. As such, the individual MMN amplitude might be applied as a more linear objective estimate of perceived magnitude of sound differences compared to behavioral hit rates.

Exploratory correlations with music appreciation and clinical factors
The rated enjoyment of music and quality of life of CI users have been found to be related to the rated sound quality achieved with the CI (Dritsakis et al., 2017, Fuller et al., 2022, Lassaletta et al., 2007. Our exploratory findings at the neural level suggested that CI users with higher ratings of enjoyment of listening to music might be predicted by faster individual MMN latencies for the 1semitone pitch deviant. This is in line with previous findings that moods or emotions in music are often conveyed by 1-semitone dif-ferences, and inefficient auditory processing of the 1-semitone differences limits CI users ability to perceive musical moods or emotions, which are part of the enjoyment of music listening (Caldwell et al., 2015). Moreover, we observed that individual CI users with higher MMN amplitude for the bright piano and guitar timbre deviants might predict higher ratings of the perceived quality of musical sounds. Furthermore, our findings suggested that earlier CI implantation after the severe hearing loss might lead to higher individual MMN amplitude and faster individual MMN latency for the large -103 ms rhythm deviant (that occurs maximally out of phase in between the 205 ms tone stimulus onset asynchrony and shows the average highest rhythm MMN amplitude in the experienced CI users (Petersen et al., 2020)). It should be mentioned that these relationships were not observed in previous group analysis on the same EEG data when the new SCA statistics method was not applied (Petersen et al., 2020). The replicability of these uncorrected exploratory relationships between individual MMN and music appreciation and clinical factors would need to be further tested in future studies before any conclusions can be drawn.

Music discrimination thresholds for individual experienced cochlear implant users
For the first time, we estimated music discrimination thresholds for experienced individual cochlear implant (CI) users based on individual MMN detections.

Intensity
As expected, MMN was detected in fewer individual CI users and normally hearing (NH) participants when the intensity deviant magnitudes were smaller, which supports the validity of the individual MMN as a marker of music discrimination thresholds. The neural discrimination thresholds for the intensity differences were similar between the individual experienced CI users and the NH controls. This suggests that, despite the dynamical range is limited with CI (Petersen et al., 2020), the individual experienced CI users detected the dynamical changes between -3 to -12 dB in the music stimuli equally well compared to the NH controls.

Pitch
For the pitch differences, also, MMN was detected in fewer individual CI users when the deviant magnitudes were smaller, which again supports the validity of the individual MMN as a marker of music discrimination thresholds in individual CI users. The individual CI users showed neural ability to detect the XL, L, and M pitch deviants comparable to the NH. However, only 55 % (6/11) of the experienced CI users showed individual significant MMN for the small pitch deviant of 1 semitone decrease (-6% in Hz), whereas significant individual MMN was detected in all (14/14) NH for this small pitch deviant. This is in line with previous research showing that pitch perception with CI is limited and varied across individuals compared to NH , Oxenham, 2008, and this finding further supports the validity of the individual MMN as a marker of music discrimination thresholds in CI listening.

Timbre
For the timbre feature, all individual experienced CI users and NH showed comparable ability to detect the differences in tone color, or sound quality. These findings are more promising than a recent behavioral study showing overall lower music instrument timbre discrimination in experienced CI users than NH , which might, though, be partly explained by the relative simplicity in detecting whether the sound differed in timbre (compared to the more complex task of identifying the correct image of a playing instrument among 16 possible music instruments).

Rhythm
For the rhythm feature, most experienced CI users (93 %, 10/11) showed significant individual MMN for even the smallest 26 ms rhythm deviant, which was comparable to the NH individual MMN detections (100 %, 14/14). This is consistent with previous findings suggesting that perception of relatively simple rhythmical structure is comparable between CI and NH (Innes-Brown et al., 2013, Jiam and.

Clinical feasibility
The musical stimuli of the CI MuMuFe paradigm might be considered relatively complex for clinical settings. Though, first, the MMN paradigm applies a task-free, passive listening procedure that is less challenging, even for paediatric patients (Torppa et al., 2018, Torppa et al., 2014, Torppa et al., 2012, compared to auditory judgment tests and questionnaires (e.g., Kuki et al., 2013, Kumari et al., 2016. Also, faster few-minutes shortened versions of the full paradigm could be adopted, e.g., by testing only particularly challenging deviants for an individual CI user, such as pitch deviants of specific magnitudes. Second, a majority (90 %) of CI users report they are interested in improving their music perception skills (Gfeller et al., 2019), suggesting that music is a particularly engaging type of stimulus for testing auditory perception, and objective tests are required to trace the development of the music perception skills. Third, cross-over relationships between music and language perception abilities have been reported (for reviews, see Besson et al., 2018, Jancke, 2012, Kraus and Slater, 2015, Patel, 2003, Torppa and Huotilainen, 2019, suggesting that the testing of music perception abilities can indicate more general traits of auditory perception shared with language perception. Fourth, the application of more complex but naturalistic stimuli has the benefit of increasing the ecological validity in the testing of the CI users' auditory perception of complex sounds, which can be more challenging than the perception of simpler sounds (Jiam and Limb, 2020). A fifth point is that music stimuli are by nature often repetitive (Huron, 2013), and repetitiveness is a wanted prerequisite for the measurement of mismatchnegativity (MMN) responses to auditory deviants (Näätänen et al., 2017).
Moreover, there are ongoing improvements in EEG recording with CI users, e.g., a recent feasibility study showed that cortical auditory evoked responses (CAEPs) can be recorded directly with electrodes built into the receiver and the cochlear array on the contralateral CI device to the ear of stimulation (resulting in CAEP signals of comparable quality to conventional scalp recordings with Fz and mastoid electrodes), thus making it unnecessary to prepare electrodes located on the scalp prior to the measurements (Attias et al., 2022).
Finally, developments are currently being made towards automatic suppression of CI artifacts in clinical EEG data (for a review, see Intartaglia et al., 2022). Despite the complexity of the individual-level statistics, the individual MMN can already be detected automatically as shown here and elsewhere (Bishop andHardiman, 2010, McGee et al., 1997).

Auditory diagnostics and screening
The results did not approach 100 % (>99 %) sensitivity, which suggests that the individual MMN is not recommendable for large-scale universal screening for general auditory deficits. Though, the approximately 90 % accuracy of the individual MMN is high and desirable for diagnostics on music discrimination abil-ity. We recommend that the automatic statistical diagnostics is followed up by visual confirmation. Since the amplitude range of spurious ''false MMN" detections was relatively low, tests resulting in low individual MMN amplitudes might be followed up by retesting.

N1 or MMN?
Subtracting the average standard response will include the difference in N1 between the deviant and standard responses. This ensures that the N1 will not directly be present in the difference wave, assuming the N1 does not differ between the standard and deviation tones. However, if there is a disinhibition of the N1 at the deviant tones, this would result in an N1 disinhibition response, which some considers also to be a subcomponent of the MMN (Gu et al., 2018, May andTiitinen, 2004). Regardless of these theoretical distinctions, the present findings suggested that in practice these early latency negative fronto-central evoked responses could accurately predict behavioral music discrimination ability.

Opposite diagnostical and statistical terminology for CAEP tests
In this study, and all other studies on cortical auditory evoked potential tests of sound detection and discrimination ability, there is an opposition between the diagnostical and statistical definitions of ''sensitivity" and ''specificity". To avoid contradiction with the common clinical terminology, we think it makes most sense to apply the diagnostical definition. Diagnostic ''sensitivity" means individual MMN absence or retaining the null-hypothesis, because ''sensitivity" indicates the abnormality or disease detection, which is the lack of discrimination indicated by the individual MMN absence or retainment of the null-hypothesis. Diagnostic ''specificity" means that individual MMN was present, or the rejection of the null-hypothesis, which means the test was passed without indication of abnormality. Since diagnostic ''sensitivity" is more important than ''specificity", it is important to keep the correct diagnostic definition to avoid confusion from the clinical perspective. This is, nevertheless, the opposite of the common statistical definition, where ''sensitivity" would mean the null-hypothesis was rejected or individual MMN detection, and ''specificity" would mean the null-hypothesis was retained or individual MMN absence.

Conclusion
We investigated the feasibility of applying the individual MMN as a marker of musical sound discrimination thresholds in individual cochlear implant (CI) users. We introduced an optimized spike density component analysis (SCA) statistics method to improve the accuracy of the individual MMN detections. Our findings showed that the automatically detected individual MMN correctly predicted the sound discrimination ability for CI users and normally hearing (NH) in approximately 9/10 cases. Moreover, the individual MMN findings suggested that adult CI users can reach similar music discrimination abilities for intensity, timbre, and rhythm features compared to NH after sustained CI experience (average 7 years). In conclusion, the individual MMN can be applied as an accurate diagnostic marker for assessing music discrimination in CI users, which is especially beneficial for individuals for whom MMN measurements are feasible but behavioral testing is limited, such as pediatric CI users.

Author contributions
NTH wrote the first manuscript draft, preprocessed the EEG data, designed and programmed the SCA statistics method, and conducted the individual-and group-level analyses. BP, EB, PV, and contributed to the conception and design of the study. BP and AA conceived the paradigm and the behavioral test and created the stimuli. AA organized and carried out the recruitment and tests. KFF contributed with participant recruitments and manuscript revisions. PV contributed with funding and manuscript revisions. All authors contributed to manuscript revisions and read and approved the manuscript.

Funding
NTHs contributions to the project were partly funded by the Oticon Medical, Denmark, and partly by the Danish National Research Foundation's Center for Music in the Brain (DNRF117).

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.