- Email: [email protected]
The impact of musical experience on neural sound encoding performance
Accepted Manuscript Title: THE IMPACT OF MUSICAL EXPERIENCE ON NEURAL SOUND ENCODING PERFORMANCE ¨ Author: Serap Aydin Cagdas Gud ¨ uc ¨ u¨ Firat Kutluk Adile Oniz ¨ Murat Ozg¨oren PII: DOI: Reference:
S0304-3940(18)30803-6 https://doi.org/doi:10.1016/j.neulet.2018.11.034 NSL 33952
To appear in:
Neuroscience Letters
Received date: Revised date: Accepted date:
16 June 2018 26 October 2018 13 November 2018
Please cite this article as: Serap Aydin, C¸a˘gdas¸ Gddotudddotucddotu, Firat Kutluk, ¨ ¨ Adile Oniz, Murat Ozgddotoren, THE IMPACT OF MUSICAL EXPERIENCE ON NEURAL SOUND ENCODING PERFORMANCE, (2018), https://doi.org/10.1016/j.neulet.2018.11.034 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
• The experience on sound perception causes the more complexity in the brain.
ip t
• Auditory stimuli generate pattern variations in single trials, while the same stimuli can not affect the spectral properties of these trials.
cr
• The relatively higher auditory neuronal complexity is observed at right fronto-temporal regions.
Ac ce p
te
d
M
an
us
• Musicians have more capability in aspects of the neural encoding of atonal chords in accordance with excellent long-term musical training.
1
Page 1 of 22
ip t cr us an M d te Ac ce p Preprint submitted to Neuroscience Letters
June 17, 2018
Page 2 of 22
ip t
THE IMPACT OF MUSICAL EXPERIENCE ON NEURAL SOUND ENCODING PERFORMANCE ¨ b , Murat Ozg¨ ¨ orenb Serap Aydına , C ¸ a˘gda¸s G¨ ud¨ uc¨ ub , Fırat Kutlukc , Adile Oniz a Department
of Biophysics, Faculty of Medicine, University of Hacettepe, Ankara, Turkey of Biophysics, Faculty of Medicine, University of Dokuz Eyl¨ ul, Izmir, Turkey c Department of Musicology, Faculty of Fine Arts, University of Dokuz Eyl¨ ul, Izmir, Turkey
us
cr
b Department
Abstract
an
In this study, 64-channel single trial auditory brain oscillations (STABO) have been firstly analyzed by using complexity metrics to observe the effect of musical experience on brain functions. Experimental data was recorded from eyes-
M
opened volunteers during listening the musical chords by piano. Complexity estimation methods were compared to each other for classification of groups (professional musicians and non-musicians) by using both classifiers (Support Vec-
d
tor Machine (SVM), Naive Bayes (NB)) and statistical tests (one-way ANOVA)
te
with respect to electrode locations. Permutation Entropy (PermEn) is found to be the best metric (p<<0.0001, 98.37% and 98.41% accuracies for tonal and
Ac ce p
atonal ensembles) at fronto-temporal regions which are responsible for cognitive task evaluation and perception of sound. PermEn also provides the meaningful results at the whole cortex (p<<0.0001, 99.81% accuracy for both tonal and atonal ensembles) through SVM with Radial Basis kernels superior to Gaussians. Almost the similar performance is also obtained for temporal features. Although, performance improvements are observed for spectral methods with NB, the considerable better results are obtained with SVM. The results indicate that musical stimuli cause pattern variations instead of spectral variations on STABO due to relatively higher neuronal activities around auditory cortex. In conclusion, temporal regions produce response to auditory stimuli, while frontal area integrates the auditory task at the same time. As well, the parietal cortex produces neural information according to the degree of attention generated by
Preprint submitted to Neuroscience Letters
October 26, 2018
Page 3 of 22
environmental changes such as atonal stimuli. It can be clearly stated that mu-
mostly fronto-temporal regions.
cr
Keywords: Brain, Music, Entropy, Auditory, Tonality
ip t
sical experience enhances the neural encoding performance of sound tonality at
1. Introduction
us
Investigation of functional and anatomical development of the brain is an attractive research field important in neuroplasticity [1]. Musical chords have
an
been frequently used as auditory stimuli to evoke emotions for observation of cognitive brain functions [2, 3]. Several EEG based studies show that musical stimuli enhance the neural populations [4]. It has been shown that musical
M
experience can combine several skills such as auditory perception, kinesthetic control, visual perception, and memory encoding depending on time period of training [5]. Moreover, this combination induces anatomical differences between
d
musicians’ and non-musicians’ [6]. In particular, adult instrumental musicians are found to have more gray matter in somatosensory, premotor, superior pari-
te
etal, and inferior temporal regions [7], and to have larger cerebellar volume [8]. In addition, functional differences have also been presented between professional
Ac ce p
pianists and non-musicians with respect to motor activations during a bimanual sequential finger movement task, however, there is no significant differences in activation levels between parallel and mirror bimanual finger movements [9]. Regarding as neural structure or cortical functions in auditory and sensori-
motor areas, functional differences between musicians and non-musicians have been presented in both neuro-imaging modalities [10] and quantitative EEG analysis [11, 12]. However, uncommon findings are still questioned. Auditory stimulus travels along with multiple sub-cortical pathways extended from temporal to frontal lobes and then, reaches the thalamus (primary component of frontal cortical basal ganglia-thalamic circuits). Therefore, temporal and prefrontal regions have commonly prolonged maturation in comparison to both occipital and parietal lobes in auditory paradigms [13]. This study includes a
2
Page 4 of 22
pre-work to propose a new concept based on entropy estimation of single trial non-averaged auditory oscillations to highlight the relativity of functional brain
ip t
activations to musical experience.
In EEG research, complexity analysis has been frequently examined by using
cr
entropy methods for detection of several disorders such as mild cognitive impair-
ment [14], schizophrenia [15, 16], autism spectrum disorder [17], Alzheimer dis-
us
order [18, 19] and epilepsy [20]. Besides, embedding entropy estimation methods have been examined to track changes in EEG series in anesthesia states, and to detect burst suppression [21]. From 2nd Law of Thermodynamics, low entropy
an
refers the organized states in a regular system, while high entropy refers the disorganized states in a complex system. In EEG analysis, entropy values refer the degree of local neuronal complexity. Embedding entropies, spectral entropy
M
and, wavelet entropy are sensitive to amplitude variations, spectral variations and 2nd order statistical variations, respectively. The other complexity metric so called Lemple Ziv Complexity (LZC) is sensitive to spectral bandwidth as
d
well as amplitude distribution during both cognitive tasks [22] and night sleep
te
[23]. In addition, LZC has also been applied to spontaneous EEG series for detection of neurological disorders such as epileptic seizure [24] and Alzheimer
Ac ce p
Disease [25]. Another EEG complexity metric is Auto-Regressive (AR) model order such that short EEG epoches are assumed to be a linear filter driven by a white noise with small number of samples. AR model describes a current value of short EEG segment using weighted information from p past samples. Then
p is called as the model order. The AR model order estimation has been examined for estimation of EEG complexity level relativity in various states such as awake [26], sleep [27] and epileptic seizure [28, 29]. Singular Spectrum Analysis (SSA) can be considered as the last complexity metric which is a nonlinear filtering approach based on principle components of EEG segments. SSA has been firstly proposed to analyze quasi-periodic time signals contaminated with noise [30]. Later, it has been adapted for emotion recognition [31] and, sleep state detection [32]. The brain can be considered as a chaotic system. Cortical neuron activities 3
Page 5 of 22
produce asynchronous and synchronous discharges by means of spontaneous EEG activities. Then, EEG complexity is associated with neuronal levels from
ip t
system to single ion (cortical neural network, neurons, ionic channels, permeable chemical ions) in terms of statistical, probabilistic, spectral and pattern
cr
variations depending on the type and severity and neurophysiological functions of interest. Therefore, LZC, ARMO, SSA and entropy estimation methods are
us
compared to each other for classification of professional musicians (PMs) and non-musicians (NMs) with respect to recording channels for tonal and atonal
an
musical chords in the present study.
2. Methods
M
2.1. Data Collection and preprocessing
Eyes-opened participants comfortably seated on a chair during auditory tasks in the recording room equipped with electromagnetic insulation. Tonal
d
and atonal sounds were delivered to them through a headset (Sennheiser HD 380 Pro, Germany). Synchronization of auditory stimulus delivering and EEG
te
recording was provided by using EMISU stimulation system [34]. A high quality sound system (Zed 14, Allen and Heath, United Kingdom) was used to adjust
Ac ce p
the sound intensity of auditory stimuli to 60 dB. Surface brain oscillations were
recorded from volunteers by using Neuroscan Quick Cap (USA) with 64-channel digital system (Neuroscan, Synamps 64, USA) with respect to 10-10 electrode placement system [33]. The earlobes were reference locations. Notch filter and band-pass filter (0.5–48 Hz cut-off) were applied to each noisy single trial of
2 ms (1000 ms pre-stimulus and 1000 ms post-stimulus). An expert technician excluded the epoches including physiological artifacts originated from eye blinks or body/head movements. During experiments, two bipolar leads for both vertical and horizontal EOG (VEOG, HEOG) were also used with respect to international 10-20 electrode placement system.
4
Page 6 of 22
2.2. Stimulus Parameters
ip t
Identical duration of the chord sounds was 2000 msec. Tonal chords satisfied the harmony rules in Western Art Music (WAM), while atonal chords did not. From academic musical point of view, tonal chords were F-II, G-II, C, C-I, E-
cr
II, Am, Am-I, Em-II, Fm-II and Cm, while atonal chords were generated by combination of two intervals, K2-B2 and K2-K2. Both of them were introduced
and atonal chords was minimized during tests.
us
to participants as an adaptation session. The sound difference between tonal
an
2.3. Inclusion and Exclusion Criteria in Volunteers
18 healthy right-handed PMs and matched NMs having no any musical education took part in experiments. Females (F) were pianists graduated from
M
conservatory, while males (M) were from the Department of audio technologies at Faculty of Fine Art in PMs. Any neurological, psychiatric, and chronic medical illness were exclusion criteria. The inclusion criteria was the threshold level
d
lower than 10% regarding the grand average potential. Edinburg hand pref-
te
erence test was applied to volunteers with respect to laterality index to select right-handed volunteers. Lastly, audiometric hearing test was used as exclusion
Ac ce p
criteria such that the individual who did not sense the difference between two different sounds (125 Hz and 8 kHz) was excluded in tests. 2.4. Feature Extraction from Musical EEG Single-channel noisy non-averaged EEG epoches were separately analyzed by
using ten complexity metrics with respect to individual, electrode placement and tonality: The first metric, LZC is a non-parametric complexity approach which measures the rate of distinct substrings’ recurrence along the finite length EEG data [1]. A symbolisation procedure was applied to raw data in modified version of LZC (MLZC) [35]. It was used for detection of brain states in reference [36]. The secondary metric, entropy measures the amount of statistical information carried by a random sequence. Entropy can be estimated by using large variety of methods as follows: Wavelet entropies based on nearby trajectory point 5
Page 7 of 22
Table 1: Demographic information of the volunteers
ip t
PMs gender
F
F
F
F
F
F
F
F
age
19
20
21
23
25
38
40
44
L.Q.
100
70
90
100
100
100
100
60
gender
M
M
M
M
M
M
M
age
45
22
25
26
26
26
27
L.Q.
100
80
100
100
90
100
100
70
90
F
F
F
F
F
22
28
28
40
45
80
80
100
100
100
M
M
M
M
M
26
28
30
32
40
40
100
100
80
100
100
90
F
F
F
age
19
20
21
22
L.Q.
90
60
100
90
gender
M
M
M
M
age
21
24
25
L.Q.
90
100
90
cr
100
M
M
28
44
us
F
d
M
gender
an
NMs
F
te
probabilities (Shannon E., Logarithmic Energy E.), embedding entropies based on amplitude variations (Approximate E., Sample E., Permutation E., Fuzzy
Ac ce p
E.), spectral entropy based on power spectral density (Spectral E.). Third metric is SSA based on nonlinear whitening filtering concept for quasi-periodic time series. The last metric, ARMO is based on parametric spectral estimation [28, 37, 38].
2.4.1. Classification of Musical Features Two-class classification problems (PMs versus NMs with respect to atonal
and tonal complexity features) were solved by using SVM with both radial basis function (RBF) and gaussian kernels. In addition, Naive Bayes (NB) classifier based on Bayes’ naive independence assumption between the features was also performed for the same groups. For two-class classification problem, SVM algorithm achieves a good generalisation by maximising the minimal margin between these classes. Its per6
Page 8 of 22
formance is generally measured by using cross-validation testing strategies [39]. In case of low sample size data, high dimensionality might lead to bias the
ip t
estimated classification performance as side-effects [40]. In biomedical applica-
tions, SVM has been frequently used for 10-fold cross-validated data for large
cr
data [41, 42, 43, 44]. Regarding these applications, 10-fold cross validation is
performed in classification steps in this study due to considerably large data
us
size.
The feature array is assigned to a class with the highest probability based on Bayes rule in NB classification [45, 46]. This class of such feature array is esti-
an
mated through Maximum A Posteriori rule such that: A maximum probability algorithm is used to determine the class of earlier probabilities in NB model. Then, a feature’s probability distribution is estimated from a training set with
M
a maximized posteriori decision tree to assign the class. Assuming a different normal distribution of training data, leads to quadratic decision boundaries. Therefore, NB classifiers have been successfully examined for motor imagery
d
[47, 48] and mental task classification [49, 50]. The reference work showed that
te
the classification capability of NB is uncorrelated with the degree of feature dependencies measured in terms of class-conditional mutual information between
Ac ce p
two feature sets [51]. Thus, NB model is assumed in our experimental tests to obtain better classification performance. In tests, the feature set (n) was randomly broken into 10 equal sized sub-
sets, then the 9n/10 part of the feature set is trained to test the remaining features (n/10) ten times. Then, mean accuracy is obtained as reliable classi-
fication performance. Radial Basis Function (RBF) and gaussian were used as kernels for SVM models. The model based parameters of SVM was optimized by using Statistics and Machine Learning Toolbox in MatlabR2018b. Multivariate multinomial data distribution is assumed in NB classifiers for multinomial models and, the width was determined automatically for the feature in the class. Since, auditory stimuli are expected to produce the relatively higher neuronal activations at temporal locations, three main feature sets were determined in tests: The first set (FS-I) consists of the complexity levels for 64-channels, the 7
Page 9 of 22
second set (FS-II) consists of the complexity levels for 13 electrode placements corresponding to the outer line of the head (P7, TP7, T7, FT7, F7, AF7, FPz,
ip t
AF8, F8, FT8, T8, TP8, P8) and, the third set (FS-III) consists of complexity levels for 6 temporal channels (TP7, T7, FT7, FT8, T8, TP8)). In these sets,
cr
the number of features (# of epoches x # of channels x # of subjects) were
us
1.255.680, 255.060, 117.720 for FS-I, FS-II and FS-III, respectively.
3. Results
In tests, SVM classifiers with RBF and gaussian kernels and NB classi-
an
fiers were run for three separate feature sets in order to highlight the corticofunctional differences between PMs and NMs. The performance of these applica-
M
tions was measured by conventional classification accuracy (CA) percent. Table 2 and 3 shows mean CAs over 10-fold cross-validation in SVM classifiers driven by RBF and gaussian kernels, respectively. By comparing two tables, complex-
d
ity metrics produced higher performance for RBF kernels, however, PermEn provided the most successful classification for both kernels for both atonal and
te
tonal ensembles. In particular, useful CAs higher than 96% were obtained by using MLZC and embedding entropy methods for each feature set. As well,
Ac ce p
LogEn, SSA and ARMO did not distinguish the groups in any set for RBF kernels. In addition to these three methods, the performance of SpecEn was low and non-useful for gaussian kernels. Table 4 shows mean CAs over 10-fold cross-validation in NB classifiers. Al-
though, the performance of the methods (LogEn, SpecEn, SSA, ARMO), which were found to be unsuccessful in SVM applications, produced relatively better results, the method (PermEn), which was found to be the best method in SVM applications, produced the relatively lower results in NB classification. However, the useful CAs, produced by MLZC and two embedding entropy approaches were around 80% in table. Since, EEG data included the auditory brain responses, the features, obtained from both right and left recordings locations close to auditory cortex,
8
Page 10 of 22
cause the high classification permormance in PermEn estimations. So, ANOVA test was used to observe the statistically meaningful differences between two
ip t
groups at these electrode placements for both atonal and tonal ensembles. Re-
garding Table 5, the most clear meaningful differences (p << 0.0001) were
cr
observed for atonal ensembles at six electrode placements (TP7, FT7, F7, FPz,
AF8, F8). Then, statistical distribution (mean ∓ std) of PermEn estimations
us
were shown for both atonal and tonal ensembles in Figure 1 and Figure 2 for PMs and NMs, respectively. By comparing these figures, the higher entropy values were produced by PMs at mostly right temporal location (T8), while statistical
an
std intervals were relatively larger at left fronto-temporal region (FT7) in NMs. Table 2: Mean CAs obtained for Complexity Metrics through SVM with RBF kernels
FS-II
FS-III
Atonal
Tonal
Atonal
Tonal
Atonal
MLZC
96.86
97.37
96.76
97.37
96.86
96.47
ApEn
98.89
98.73
98.53
98.73
99.08
98.28
SampEn
97.43
PermEn
99.81
97.91
86.05
96.56
98.07
97.91
te
d
Tonal
M
FS-I
99.81
99.81
99.81
99.81
99.72
83.92
80.10
80.00
80.10
79.50
79.38
92.47
93.57
92.47
93.57
93.11
93.48
LogEn
49.54
49.72
49.54
49.72
49.54
49.72
SpecEn
74.31
74.59
71.74
74.59
71.55
74.95
SSA
55.96
56.38
57.88
56.38
57.88
57.23
ARMO
51.00
51.17
50.91
51.17
51.10
50.90
FuzyEn
Ac ce p
ShanEn
4. Conclusion
The results support that PMs provide the higher complexity levels due to the more neural activations at right fronto-temporal lobes. In conclusion, this finding can be explained by increasing processing capabilities in PMs. In other
9
Page 11 of 22
FS-I
FS-II
FS-III
ip t
Table 3: Mean CAs obtained for Complexity Metrics through SVM with Gaussian kernels
Atonal
Tonal
Atonal
Tonal
Atonal
MLZC
78.82
75.11
86.07
75.11
84.60
76.54
ApEn
93.66
95.57
92.56
95.57
93.30
95.20
SampEn
92.93
94.66
73.48
94.12
93.76
94.21
PermEn
99.72
99.54
99.72
99.54
FuzyEn
79.21
60.61
69.80
60.61
ShanEn
76.88
78.84
77.06
78.84
LogEn
49.54
49.72
49.54
SpecEn
56.97
55.27
SSA
49.54
50.59
ARMO
49.54
49.82
us
cr
Tonal
99.45
72.84
67.00
76.88
78.93
49.72
49.54
49.72
49.90
55.27
51.46
56.27
49.54
50.59
49.54
50.59
49.54
49.82
49.54
49.91
te
d
M
an
99.72
Tonal
Atonal
Tonal
Atonal
Tonal
Atonal
MLZC
80.78
79.82
79.21
79.82
80.29
79.91
ApEn
83.48
83.00
87.33
83.00
87.88
85.54
SampEn
82.56
83.00
67.24
84.36
84.95
85.99
PermEn
72.75
80.56
74.49
80.56
71.55
79.74
FuzyEn
80.29
80.92
78.23
80.92
83.23
82.64
ShanEn
73.85
75.23
75.87
75.23
75.50
75.96
LogEn
77.33
80.55
76.78
80.55
77.15
82.45
SpecEn
64.67
68.43
66.14
68.43
65.68
70.14
SSA
73.48
71.04
70.82
71.04
71.65
69.50
ARMO
56.78
63.54
57.43
63.54
57.52
62.28
Table 4: Mean CAs obtained for Complexity Metrics through NB classifier
Ac ce p
FS-I
FS-II
FS-III
10
Page 12 of 22
ip t
Table 5: Anova Test Results (p-values) for FS-II (** refers p<<0.0001, * refers p<0.0001)
P7
TP7
T7
FT7
F7
AF7
FPz
Tonal
**
0.3
*
0.1
0.7
**
0.01
Atonal
0.5
**
0.1
**
**
0.03
**
Right side
AF8
F8
FT8
T8
TP8
P8
Tonal
0.8
0.002
0.0002
0.01
**
0.005
Atonal
*
**
**
0.04
*
0.8
Ac ce p
te
d
M
an
us
cr
Left side
Figure 1: Statistical distribution of PermEn estimations in PMs
11
Page 13 of 22
ip t cr us an
M
Figure 2: Statistical distribution of PermEn estimations in NMs
words, both experience and knowledge about the acoustic stimuli cause the high
d
neuronal integration within neural ensembles together with at auditory cortex.
te
Tonal and atonal musical forms require discussion in regard to pleasure, primed memory, experience, fitness for music scales. Therefore, we can state the electrophysiological followings: Music processing is lateralized at mostly
Ac ce p
right hemisphere. A complex neural processing network is required in high cognitive musical functions such as musical attention, and the tracking of harmonic structure in time. Our current findings are compatible with the previous biological findings indicating that certain neurotransmitters (dopamine) is produced as a result of functions in amygdala affected by listening pleasure music [52]. The reference states that successfully tracking of tonal chords activates mostly prefrontal regions numbered by 44, 45, and 47 in Brodmann areas in addition to both anterior and posterior cingulate gyrus [53]. As well, musicians are capable of neural encoding of sounds due to excellent long-term musical training [54, 55]. PermEn is found to be best complexity method to observe superior cognitive functions of musicians due to the fact that auditory stimuli induce amplitude variations in single trials. PermEn was proposed as useful method for estimation 12
Page 14 of 22
of neural dysfunctions such as epilepsy [56], seizure [57] and OCD [58]. As well, the computational complexity of PermEn is lower. In this study, PermEn
ip t
provided the highest and successful results in every type of classification with
10-fold cross-validation. In nonparametric statistical problems, cross-validation
cr
methodologies have been used as a means of selecting tuning parameters. In
future work, the method, proposed to improve the reliability of cross-validation
us
[59], will be implemented in classification steps.
SVM classifiers have been frequently used to classify power spectra of EEG data including mental task [60] and seizure events [61] when the training set
an
is large, since the optimal decision function can be obtained through statistical learning theory. However, kernel function specification affect the classification performance. In this study, RBF was found to be more useful for entropy es-
M
timations, even if the gaussian kernels provide the relative improvements in classification performance for parametric and spectral estimations. However, NB classifiers provided the relatively much more improvements for the same
d
parametric and spectral estimations in comparison to SVM with gaussian ker-
te
nels. In reference, RBF network shows better generalization performance and computationally faster than SVM with Gaussian kernel, specially for large train-
Ac ce p
ing data sets [62]. In conclusion, both distribution of feature values and kernel types can be considered as performance parameters in EEG applications. The reference discuss the close relation between kernel function and classification accuracy for continuous time stationary process [63]. Feature set characteristics highly affect the performance of NB classifier. Baye
References
[1] Munte T.F. (2002) The musician brain as a model of neuroplasticity, Nature Review Neuroscience, 3:473-478. [2] Andrade P.E., (2003) Brain tuned to music, J. Roy. Soc. Med. 96, 284-287. [3] Peretz I., Hebert S. (2000) Toward a biological account of music experience, Brain Cognition, 42, 131-134. 13
Page 15 of 22
[4] Nozaradan S., Zerouali Y., et al. (2015) Capturing with EEG the neural entrainment and coupling underlying sensorimotor synchronization to the
ip t
beat, Cereb. Cortex 25, 736–747.
[5] Herholz S.C., Zatorre R.J. (2012) Musical training as a framework for brain
cr
plasticity: behavior, function, and structure, Neuron. 76(3), 486-502.
[6] Strait D.L., Kraus N. (2014) Biological impact of auditory expertise across
us
the life span: musicians as a model of auditory learning, Hear Res., 308, 109-21.
an
[7] Gaser C., Schlaug G. (2003), Brain structures differ between musicians and non-musicians, J. Neurosci. 23, 9240–9245
Cereb. Cortex, 13, 943–949.
M
[8] Hutchinson S., Lee L.H.L., et al. (2003), Cerebellar volume of musicians,
[9] Haslinger B., Erhard, P., et al. (2004) Reduced recruitment of motor asso-
te
Mapp. 22, 206–15.
d
ciation areas during bimanual coordination in concert pianists, Hum Brain
[10] J¨ancke L., (2009) Music drives brain plasticity, Biol. Rep., p.78, 10.3410,b1-
Ac ce p
78
[11] Brattico E., Tervaniemi H., et al., (2006) Musical scale properties are automatically processed in the human auditory cortex, Brain Res., 1117(1), 162-174.
[12] Fujioka T., Trainor L.J., et al. (2005) Automatic encoding of polyphonic melodies in musicians and nonmusicians, J. Cogn. Neurosci., 17(10), 15781592,
[13] Sowell E.R., Thompson P.M., et al., (2004) Longitudinal mapping of cortical thickness and brain growth in normal children, J. Neurosci. 24, 8223–8231.
14
Page 16 of 22
[14] Zhu B., Chai C., et al. (2017) Analysis of EEG complexity in patients with
ip t
mild cognitive impairment, J. Neurol Disord 5:354. [15] Thilakvathi B., Shenbaga Devi S., et al. (2017) EEG signal complexity anal-
ysis for schizophrenia during rest and mental activity, Biomedical Research,
cr
28(1), 1-9.
[16] Sabeti M., Katebi S., Boostani R. (2009) Entropy and complexity measures
tificial Intelligence in Medicine, 47(3), 263-274
us
for EEG signal classification of schizophrenic and control participants, Ar-
an
[17] Bosl W., Tierney A., et al. (2011) EEG complexity as a biomarker for autism spectrum disorder risk, BMC Medicine, 9-18.
M
[18] Tylovia L., Kukal J., et al. (2018) Unbiased estimation of permutation entropy in EEG analysis for Alzheimer’s disease classification, Biomedical Signal Proc. and Control, 39, 424-430
d
[19] Simons S., Espino P., Ab´asolo D. (2018) Fuzzy entropy analysis of the
te
electroencephalogram in patients with Alzheimer’s disease: Is the method superior to sample entropy?, Entropy, 20, 21.
Ac ce p
[20] Acharya U.R., Molinari F., Sree S.V. (2012) Automated diagnosis of epileptic EEG using entropies, Biomedical Signal Proc. and Control, 7(4), 401-408
[21] Liang Z., Wang Y., et al. (2015) EEG entropy measures in anesthesia, Front. Comput. Neurosci.
[22] Silva A.C.S., Arce A.I.C. et al. (2010) Quantifying electrode position effects in EEG data with Lempel-Ziv complexity, 32nd Annual Int. Conf of the IEEE EMBS, Argentina.
[23] Ab´asolo D., Simons S. et al. (2015) Lempel-Ziv complexity of cortical activity during sleep and waking in rats, Journal of Neurophysiology, 113(7), 2742-2752.
15
Page 17 of 22
[24] Hu J., Gao J, Principe J.C. (2006) Analysis of Biomedical Signals by the Lempel-Ziv Complexity: the Effect of Finite Data Size, IEEE Trans on
ip t
BME, 53(12), 2606-2609.
[25] Ab`asolo D., Hornero R., et al. (2006) Analysis of EEG background activity
cr
in Alzheimer’s disease patients with Lempel–Ziv complexity and central tendency measure, Medical engineering & physics, 28(4), 315-322
us
[26] Mohammadi G., Shoushtari P. et al., (2006) Person identification by using AR model for EEG signals, Proceed of World Academy of Science, Eng
an
and Techn, 11, 281–285.
[27] Wang T.W., Guohui W., Huanqing F., (2004) Autoregressive model order
M
property for sleep EEG, J. of Biomedical Eng, 21(3), 394-6. [28] Aydın S. (2009) Comparison of power spectrum predictors in computing coherence functions for intracortical EEG signals, Annals of BME, 37(1),
d
192–200.
te
[29] Faust O., Acharya R.U., et al. (2008) Analysis of EEG signals during epileptic and alcoholic states using AR modeling techniques, IRBM, 29(1), 44-52.
Ac ce p
[30] Vautard R., Yiou P., Ghil M. (1992) Singular-spectrum analysis: A toolkit for short, noisy chaotic signals, Physica D, 58, 95–126.
[31] Aydın S., Demirta¸s S. et al. (2016) Emotion recognition with eigen features of freauency band activities embedded in induced brain oscillations mediated by affective pictures, Int. J. of Neural Systems, 26(3), 1650013.
[32] Kouchaki S., Sanei S., et al. (2015) Tensor based singular spectrum analysis for automatic scoring of sleep EEG, IEEE Trans on Neural Systems and Rehabilitation Eng., 23(1) [33] Jasper H. (1958) The ten-twenty electrode system of the international federation, EEG and Clinical Neurophysiology, 10, 371–375.
16
Page 18 of 22
¨ oren M., Erdogan U., et al. (2009) Brain asymmetry measurement using [34] Ozg¨ EMISU in applied brain biophysics, Computers in Biology and Medicine,
ip t
39(10), 879–888.
[35] Aboy M., Hornero R., Ab´asolo D., Alvarez D. (2006) Interpretation of the
cr
Lempel-Ziv complexity measure in the context of biomedical signal analysis, IEEE Trans on BME, 532282–2288.
us
[36] Tosun P.D., Ab´asolo D., Stenson G., et al. (2017) Characterisation of the effects of sleep deprivation on the electroencephalogram using permuta-
an
tion lempel–ziv complexity, a non-linear analysis tool, Entropy, 19, 673, doi:10.3390/e19120673.
[37] Aydın S. (2010) Determination of autoregressive model orders for seizure
M
detection, T. Journal of Elec Eng & Comp Sci, 18(1), doi:10.3906/elk-090683
d
[38] Andrzejak R.G., Lehnertz K., et al (2001) Indications of nonlinear deter-
te
ministic and finite dimensional structures in time seires of brain electrical activity: Dependence on recording region and brain state. Physical Review
Ac ce p
E, 64, 061907.
[39] Vapnik V.N. (2000) The Nature of Statistical Learning Theory. New York, USA Springer.
[40] Webb A.R., Copsey K.D. (2011) Statistical pattern recognition, 3rd. Hoboken, NJ, USA, John Wiley & Sons,
[41] Lin L.C., Ouyang C.S. et al. (2014) Early prediction of medication refractoriness in children with idiopathic epilepsy based on scalp EEG analysis, Int. J. Neural Syst. 24(7):1450023. [42] Wang Y., Zhou W. et al. (2013) Comparison of fractal features of ictal and interictal EEGs, Int. J. Neural Syst. 23(6):1350028.
17
Page 19 of 22
[43] Jumutc V., Zayakin P. et al. (2011) Ranking based kernels in applied biomedical diagnostics using a support vector machine, Int. J. Neural Syst.
ip t
21(6):459–473.
[44] Amin H.U., Mumtaz W. et al. (2017) Classification of EEG signals based
cr
on pattern recognition approach, Frontiers in Comput Neurosci, 11: 103.
[45] R. O. Duda, P. E. Hart, and D. G. Stork. (2001) Pattern Recognition,
us
second edition. Wiley Inter Science.
[46] K. Fukunaga. (1990) Statistical Pattern Recognition, seconde edition. Aca-
an
demic Pres.
[47] S. Lemm, C. Schafer, and G. Curio. (2004) Bci competition 2003–data set
M
iii: probabilistic modeling of sensorimotor mu rhythms for classification of imaginary hand movements. IEEE Trans on BME, 51(6):1077–1080. [48] S. Solhjoo and M. H. Moradi. (2004) Mental task recognition: A comparison
d
between some of classification methods. In BIOSIGNAL 2004 International
te
EURASIP Conference.
[49] Z. A. Keirn and J. I. Aunon. (1990) A new mode of communication between
Ac ce p
man and his surroundings, IEEE Trans on BME, 37(12).
[50] G. A. Barreto, R. A. Frota, and F. N. S. de Medeiros. (2004) On the classification of mental tasks: a performance comparison of neural and statistical approaches. In Proceedings of the IEEE Workshop on Machine Learning for Signal Processing.
[51] Rish I. (2001) An empirical study of the naive Bayes classifier, IJCAI 2001 workshop on empirical methods in artificial intelligence, 3, page 41–46. IBM New York [52] Levitin D.J. (2012) What Does It Mean to Be Musical ?, Neuron, 73(4):633637.
18
Page 20 of 22
[53] Huron D. (2006) Sweet Anticipation: Music and the Psychology of Expec-
ip t
tation, MIT Press, Cambridge, MA. [54] Thatcher R.W., Biver C.J., North D. (2007) Spatial-temporal current source correlations and cortical connectivity, Clin. EEG Neurosci. 38,
cr
35–48.
[55] Fair D.A., Cohen A.L., et al. (2009) Functional brain networks develop from
us
a “local to distributed” organization. PLoS Comput. Biol. 5, e1000381.
[56] Veisi I., Pariz N., Karimpour A. (2007) Fast and robust detection of epilepsy
an
in noisy EEG signals using permutation entropy, IEEE 7th Int. Symposium on BioInformatics and BioEngineering, 200-203.
M
[57] Mammone N., Lay-Ekuakille A., et al. (2011) Analysis of absence seizure EEG via permutation entropy spatio-temporal clustering, IEEE Int. Symposium on Medical Measurements and Applications, 532-535.
d
[58] Aydın S., et al. (2015) Classification of obsessive compulsive disorder by
te
EEG complexity and hemispheric dependency measurements, Int. J. of Neural Systems, 25, 1550010.
Ac ce p
[59] Wang Q., Lindsay B.G. (2015) Improving cross-validated bandwidth selection using subsampling-extrapolation techniques, Computational Statistics & Data Analysis, 89, 51-71.
[60] Garrett D., Peterson D.A., Anderson C.W., et al. (2003). Comparison of linear, nonlinear, and feature selection methods for EEG signal classification, IEEE Trans on Neural Systems and Rehabilitation Eng. 11(2), 141-144.
[61] Park Y., Luo L., Parhi K.K., et al. (2011) Seizure prediction with spectral power of EEG using cost-sensitive support vector machines, Epilepsia, 52(10):1761–1770. [62] Debnath R., Takahashi H. (2002) Learning Capability: Classical RBF Network vs. SVM with Gaussian Kernel. In: Hendtlass T., Ali M. (eds) Devel19
Page 21 of 22
opments in Applied Artificial Intelligence. IEA/AIE 2002. Lecture Notes
ip t
in Computer Science, vol 2358. Springer, Berlin, Heidelberg [63] Heda K.E., Louani D. (2018) Optimal bandwidth selection in kernel density estimation for continuous time dependent processes, Statistics & Probabil-
Ac ce p
te
d
M
an
us
cr
ity Letters, 138, 9-19.
20
Page 22 of 22
S0304-3940(18)30803-6 https://doi.org/doi:10.1016/j.neulet.2018.11.034 NSL 33952
To appear in:
Neuroscience Letters
Received date: Revised date: Accepted date:
16 June 2018 26 October 2018 13 November 2018
Please cite this article as: Serap Aydin, C¸a˘gdas¸ Gddotudddotucddotu, Firat Kutluk, ¨ ¨ Adile Oniz, Murat Ozgddotoren, THE IMPACT OF MUSICAL EXPERIENCE ON NEURAL SOUND ENCODING PERFORMANCE, (2018), https://doi.org/10.1016/j.neulet.2018.11.034 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
• The experience on sound perception causes the more complexity in the brain.
ip t
• Auditory stimuli generate pattern variations in single trials, while the same stimuli can not affect the spectral properties of these trials.
cr
• The relatively higher auditory neuronal complexity is observed at right fronto-temporal regions.
Ac ce p
te
d
M
an
us
• Musicians have more capability in aspects of the neural encoding of atonal chords in accordance with excellent long-term musical training.
1
Page 1 of 22
ip t cr us an M d te Ac ce p Preprint submitted to Neuroscience Letters
June 17, 2018
Page 2 of 22
ip t
THE IMPACT OF MUSICAL EXPERIENCE ON NEURAL SOUND ENCODING PERFORMANCE ¨ b , Murat Ozg¨ ¨ orenb Serap Aydına , C ¸ a˘gda¸s G¨ ud¨ uc¨ ub , Fırat Kutlukc , Adile Oniz a Department
of Biophysics, Faculty of Medicine, University of Hacettepe, Ankara, Turkey of Biophysics, Faculty of Medicine, University of Dokuz Eyl¨ ul, Izmir, Turkey c Department of Musicology, Faculty of Fine Arts, University of Dokuz Eyl¨ ul, Izmir, Turkey
us
cr
b Department
Abstract
an
In this study, 64-channel single trial auditory brain oscillations (STABO) have been firstly analyzed by using complexity metrics to observe the effect of musical experience on brain functions. Experimental data was recorded from eyes-
M
opened volunteers during listening the musical chords by piano. Complexity estimation methods were compared to each other for classification of groups (professional musicians and non-musicians) by using both classifiers (Support Vec-
d
tor Machine (SVM), Naive Bayes (NB)) and statistical tests (one-way ANOVA)
te
with respect to electrode locations. Permutation Entropy (PermEn) is found to be the best metric (p<<0.0001, 98.37% and 98.41% accuracies for tonal and
Ac ce p
atonal ensembles) at fronto-temporal regions which are responsible for cognitive task evaluation and perception of sound. PermEn also provides the meaningful results at the whole cortex (p<<0.0001, 99.81% accuracy for both tonal and atonal ensembles) through SVM with Radial Basis kernels superior to Gaussians. Almost the similar performance is also obtained for temporal features. Although, performance improvements are observed for spectral methods with NB, the considerable better results are obtained with SVM. The results indicate that musical stimuli cause pattern variations instead of spectral variations on STABO due to relatively higher neuronal activities around auditory cortex. In conclusion, temporal regions produce response to auditory stimuli, while frontal area integrates the auditory task at the same time. As well, the parietal cortex produces neural information according to the degree of attention generated by
Preprint submitted to Neuroscience Letters
October 26, 2018
Page 3 of 22
environmental changes such as atonal stimuli. It can be clearly stated that mu-
mostly fronto-temporal regions.
cr
Keywords: Brain, Music, Entropy, Auditory, Tonality
ip t
sical experience enhances the neural encoding performance of sound tonality at
1. Introduction
us
Investigation of functional and anatomical development of the brain is an attractive research field important in neuroplasticity [1]. Musical chords have
an
been frequently used as auditory stimuli to evoke emotions for observation of cognitive brain functions [2, 3]. Several EEG based studies show that musical stimuli enhance the neural populations [4]. It has been shown that musical
M
experience can combine several skills such as auditory perception, kinesthetic control, visual perception, and memory encoding depending on time period of training [5]. Moreover, this combination induces anatomical differences between
d
musicians’ and non-musicians’ [6]. In particular, adult instrumental musicians are found to have more gray matter in somatosensory, premotor, superior pari-
te
etal, and inferior temporal regions [7], and to have larger cerebellar volume [8]. In addition, functional differences have also been presented between professional
Ac ce p
pianists and non-musicians with respect to motor activations during a bimanual sequential finger movement task, however, there is no significant differences in activation levels between parallel and mirror bimanual finger movements [9]. Regarding as neural structure or cortical functions in auditory and sensori-
motor areas, functional differences between musicians and non-musicians have been presented in both neuro-imaging modalities [10] and quantitative EEG analysis [11, 12]. However, uncommon findings are still questioned. Auditory stimulus travels along with multiple sub-cortical pathways extended from temporal to frontal lobes and then, reaches the thalamus (primary component of frontal cortical basal ganglia-thalamic circuits). Therefore, temporal and prefrontal regions have commonly prolonged maturation in comparison to both occipital and parietal lobes in auditory paradigms [13]. This study includes a
2
Page 4 of 22
pre-work to propose a new concept based on entropy estimation of single trial non-averaged auditory oscillations to highlight the relativity of functional brain
ip t
activations to musical experience.
In EEG research, complexity analysis has been frequently examined by using
cr
entropy methods for detection of several disorders such as mild cognitive impair-
ment [14], schizophrenia [15, 16], autism spectrum disorder [17], Alzheimer dis-
us
order [18, 19] and epilepsy [20]. Besides, embedding entropy estimation methods have been examined to track changes in EEG series in anesthesia states, and to detect burst suppression [21]. From 2nd Law of Thermodynamics, low entropy
an
refers the organized states in a regular system, while high entropy refers the disorganized states in a complex system. In EEG analysis, entropy values refer the degree of local neuronal complexity. Embedding entropies, spectral entropy
M
and, wavelet entropy are sensitive to amplitude variations, spectral variations and 2nd order statistical variations, respectively. The other complexity metric so called Lemple Ziv Complexity (LZC) is sensitive to spectral bandwidth as
d
well as amplitude distribution during both cognitive tasks [22] and night sleep
te
[23]. In addition, LZC has also been applied to spontaneous EEG series for detection of neurological disorders such as epileptic seizure [24] and Alzheimer
Ac ce p
Disease [25]. Another EEG complexity metric is Auto-Regressive (AR) model order such that short EEG epoches are assumed to be a linear filter driven by a white noise with small number of samples. AR model describes a current value of short EEG segment using weighted information from p past samples. Then
p is called as the model order. The AR model order estimation has been examined for estimation of EEG complexity level relativity in various states such as awake [26], sleep [27] and epileptic seizure [28, 29]. Singular Spectrum Analysis (SSA) can be considered as the last complexity metric which is a nonlinear filtering approach based on principle components of EEG segments. SSA has been firstly proposed to analyze quasi-periodic time signals contaminated with noise [30]. Later, it has been adapted for emotion recognition [31] and, sleep state detection [32]. The brain can be considered as a chaotic system. Cortical neuron activities 3
Page 5 of 22
produce asynchronous and synchronous discharges by means of spontaneous EEG activities. Then, EEG complexity is associated with neuronal levels from
ip t
system to single ion (cortical neural network, neurons, ionic channels, permeable chemical ions) in terms of statistical, probabilistic, spectral and pattern
cr
variations depending on the type and severity and neurophysiological functions of interest. Therefore, LZC, ARMO, SSA and entropy estimation methods are
us
compared to each other for classification of professional musicians (PMs) and non-musicians (NMs) with respect to recording channels for tonal and atonal
an
musical chords in the present study.
2. Methods
M
2.1. Data Collection and preprocessing
Eyes-opened participants comfortably seated on a chair during auditory tasks in the recording room equipped with electromagnetic insulation. Tonal
d
and atonal sounds were delivered to them through a headset (Sennheiser HD 380 Pro, Germany). Synchronization of auditory stimulus delivering and EEG
te
recording was provided by using EMISU stimulation system [34]. A high quality sound system (Zed 14, Allen and Heath, United Kingdom) was used to adjust
Ac ce p
the sound intensity of auditory stimuli to 60 dB. Surface brain oscillations were
recorded from volunteers by using Neuroscan Quick Cap (USA) with 64-channel digital system (Neuroscan, Synamps 64, USA) with respect to 10-10 electrode placement system [33]. The earlobes were reference locations. Notch filter and band-pass filter (0.5–48 Hz cut-off) were applied to each noisy single trial of
2 ms (1000 ms pre-stimulus and 1000 ms post-stimulus). An expert technician excluded the epoches including physiological artifacts originated from eye blinks or body/head movements. During experiments, two bipolar leads for both vertical and horizontal EOG (VEOG, HEOG) were also used with respect to international 10-20 electrode placement system.
4
Page 6 of 22
2.2. Stimulus Parameters
ip t
Identical duration of the chord sounds was 2000 msec. Tonal chords satisfied the harmony rules in Western Art Music (WAM), while atonal chords did not. From academic musical point of view, tonal chords were F-II, G-II, C, C-I, E-
cr
II, Am, Am-I, Em-II, Fm-II and Cm, while atonal chords were generated by combination of two intervals, K2-B2 and K2-K2. Both of them were introduced
and atonal chords was minimized during tests.
us
to participants as an adaptation session. The sound difference between tonal
an
2.3. Inclusion and Exclusion Criteria in Volunteers
18 healthy right-handed PMs and matched NMs having no any musical education took part in experiments. Females (F) were pianists graduated from
M
conservatory, while males (M) were from the Department of audio technologies at Faculty of Fine Art in PMs. Any neurological, psychiatric, and chronic medical illness were exclusion criteria. The inclusion criteria was the threshold level
d
lower than 10% regarding the grand average potential. Edinburg hand pref-
te
erence test was applied to volunteers with respect to laterality index to select right-handed volunteers. Lastly, audiometric hearing test was used as exclusion
Ac ce p
criteria such that the individual who did not sense the difference between two different sounds (125 Hz and 8 kHz) was excluded in tests. 2.4. Feature Extraction from Musical EEG Single-channel noisy non-averaged EEG epoches were separately analyzed by
using ten complexity metrics with respect to individual, electrode placement and tonality: The first metric, LZC is a non-parametric complexity approach which measures the rate of distinct substrings’ recurrence along the finite length EEG data [1]. A symbolisation procedure was applied to raw data in modified version of LZC (MLZC) [35]. It was used for detection of brain states in reference [36]. The secondary metric, entropy measures the amount of statistical information carried by a random sequence. Entropy can be estimated by using large variety of methods as follows: Wavelet entropies based on nearby trajectory point 5
Page 7 of 22
Table 1: Demographic information of the volunteers
ip t
PMs gender
F
F
F
F
F
F
F
F
age
19
20
21
23
25
38
40
44
L.Q.
100
70
90
100
100
100
100
60
gender
M
M
M
M
M
M
M
age
45
22
25
26
26
26
27
L.Q.
100
80
100
100
90
100
100
70
90
F
F
F
F
F
22
28
28
40
45
80
80
100
100
100
M
M
M
M
M
26
28
30
32
40
40
100
100
80
100
100
90
F
F
F
age
19
20
21
22
L.Q.
90
60
100
90
gender
M
M
M
M
age
21
24
25
L.Q.
90
100
90
cr
100
M
M
28
44
us
F
d
M
gender
an
NMs
F
te
probabilities (Shannon E., Logarithmic Energy E.), embedding entropies based on amplitude variations (Approximate E., Sample E., Permutation E., Fuzzy
Ac ce p
E.), spectral entropy based on power spectral density (Spectral E.). Third metric is SSA based on nonlinear whitening filtering concept for quasi-periodic time series. The last metric, ARMO is based on parametric spectral estimation [28, 37, 38].
2.4.1. Classification of Musical Features Two-class classification problems (PMs versus NMs with respect to atonal
and tonal complexity features) were solved by using SVM with both radial basis function (RBF) and gaussian kernels. In addition, Naive Bayes (NB) classifier based on Bayes’ naive independence assumption between the features was also performed for the same groups. For two-class classification problem, SVM algorithm achieves a good generalisation by maximising the minimal margin between these classes. Its per6
Page 8 of 22
formance is generally measured by using cross-validation testing strategies [39]. In case of low sample size data, high dimensionality might lead to bias the
ip t
estimated classification performance as side-effects [40]. In biomedical applica-
tions, SVM has been frequently used for 10-fold cross-validated data for large
cr
data [41, 42, 43, 44]. Regarding these applications, 10-fold cross validation is
performed in classification steps in this study due to considerably large data
us
size.
The feature array is assigned to a class with the highest probability based on Bayes rule in NB classification [45, 46]. This class of such feature array is esti-
an
mated through Maximum A Posteriori rule such that: A maximum probability algorithm is used to determine the class of earlier probabilities in NB model. Then, a feature’s probability distribution is estimated from a training set with
M
a maximized posteriori decision tree to assign the class. Assuming a different normal distribution of training data, leads to quadratic decision boundaries. Therefore, NB classifiers have been successfully examined for motor imagery
d
[47, 48] and mental task classification [49, 50]. The reference work showed that
te
the classification capability of NB is uncorrelated with the degree of feature dependencies measured in terms of class-conditional mutual information between
Ac ce p
two feature sets [51]. Thus, NB model is assumed in our experimental tests to obtain better classification performance. In tests, the feature set (n) was randomly broken into 10 equal sized sub-
sets, then the 9n/10 part of the feature set is trained to test the remaining features (n/10) ten times. Then, mean accuracy is obtained as reliable classi-
fication performance. Radial Basis Function (RBF) and gaussian were used as kernels for SVM models. The model based parameters of SVM was optimized by using Statistics and Machine Learning Toolbox in MatlabR2018b. Multivariate multinomial data distribution is assumed in NB classifiers for multinomial models and, the width was determined automatically for the feature in the class. Since, auditory stimuli are expected to produce the relatively higher neuronal activations at temporal locations, three main feature sets were determined in tests: The first set (FS-I) consists of the complexity levels for 64-channels, the 7
Page 9 of 22
second set (FS-II) consists of the complexity levels for 13 electrode placements corresponding to the outer line of the head (P7, TP7, T7, FT7, F7, AF7, FPz,
ip t
AF8, F8, FT8, T8, TP8, P8) and, the third set (FS-III) consists of complexity levels for 6 temporal channels (TP7, T7, FT7, FT8, T8, TP8)). In these sets,
cr
the number of features (# of epoches x # of channels x # of subjects) were
us
1.255.680, 255.060, 117.720 for FS-I, FS-II and FS-III, respectively.
3. Results
In tests, SVM classifiers with RBF and gaussian kernels and NB classi-
an
fiers were run for three separate feature sets in order to highlight the corticofunctional differences between PMs and NMs. The performance of these applica-
M
tions was measured by conventional classification accuracy (CA) percent. Table 2 and 3 shows mean CAs over 10-fold cross-validation in SVM classifiers driven by RBF and gaussian kernels, respectively. By comparing two tables, complex-
d
ity metrics produced higher performance for RBF kernels, however, PermEn provided the most successful classification for both kernels for both atonal and
te
tonal ensembles. In particular, useful CAs higher than 96% were obtained by using MLZC and embedding entropy methods for each feature set. As well,
Ac ce p
LogEn, SSA and ARMO did not distinguish the groups in any set for RBF kernels. In addition to these three methods, the performance of SpecEn was low and non-useful for gaussian kernels. Table 4 shows mean CAs over 10-fold cross-validation in NB classifiers. Al-
though, the performance of the methods (LogEn, SpecEn, SSA, ARMO), which were found to be unsuccessful in SVM applications, produced relatively better results, the method (PermEn), which was found to be the best method in SVM applications, produced the relatively lower results in NB classification. However, the useful CAs, produced by MLZC and two embedding entropy approaches were around 80% in table. Since, EEG data included the auditory brain responses, the features, obtained from both right and left recordings locations close to auditory cortex,
8
Page 10 of 22
cause the high classification permormance in PermEn estimations. So, ANOVA test was used to observe the statistically meaningful differences between two
ip t
groups at these electrode placements for both atonal and tonal ensembles. Re-
garding Table 5, the most clear meaningful differences (p << 0.0001) were
cr
observed for atonal ensembles at six electrode placements (TP7, FT7, F7, FPz,
AF8, F8). Then, statistical distribution (mean ∓ std) of PermEn estimations
us
were shown for both atonal and tonal ensembles in Figure 1 and Figure 2 for PMs and NMs, respectively. By comparing these figures, the higher entropy values were produced by PMs at mostly right temporal location (T8), while statistical
an
std intervals were relatively larger at left fronto-temporal region (FT7) in NMs. Table 2: Mean CAs obtained for Complexity Metrics through SVM with RBF kernels
FS-II
FS-III
Atonal
Tonal
Atonal
Tonal
Atonal
MLZC
96.86
97.37
96.76
97.37
96.86
96.47
ApEn
98.89
98.73
98.53
98.73
99.08
98.28
SampEn
97.43
PermEn
99.81
97.91
86.05
96.56
98.07
97.91
te
d
Tonal
M
FS-I
99.81
99.81
99.81
99.81
99.72
83.92
80.10
80.00
80.10
79.50
79.38
92.47
93.57
92.47
93.57
93.11
93.48
LogEn
49.54
49.72
49.54
49.72
49.54
49.72
SpecEn
74.31
74.59
71.74
74.59
71.55
74.95
SSA
55.96
56.38
57.88
56.38
57.88
57.23
ARMO
51.00
51.17
50.91
51.17
51.10
50.90
FuzyEn
Ac ce p
ShanEn
4. Conclusion
The results support that PMs provide the higher complexity levels due to the more neural activations at right fronto-temporal lobes. In conclusion, this finding can be explained by increasing processing capabilities in PMs. In other
9
Page 11 of 22
FS-I
FS-II
FS-III
ip t
Table 3: Mean CAs obtained for Complexity Metrics through SVM with Gaussian kernels
Atonal
Tonal
Atonal
Tonal
Atonal
MLZC
78.82
75.11
86.07
75.11
84.60
76.54
ApEn
93.66
95.57
92.56
95.57
93.30
95.20
SampEn
92.93
94.66
73.48
94.12
93.76
94.21
PermEn
99.72
99.54
99.72
99.54
FuzyEn
79.21
60.61
69.80
60.61
ShanEn
76.88
78.84
77.06
78.84
LogEn
49.54
49.72
49.54
SpecEn
56.97
55.27
SSA
49.54
50.59
ARMO
49.54
49.82
us
cr
Tonal
99.45
72.84
67.00
76.88
78.93
49.72
49.54
49.72
49.90
55.27
51.46
56.27
49.54
50.59
49.54
50.59
49.54
49.82
49.54
49.91
te
d
M
an
99.72
Tonal
Atonal
Tonal
Atonal
Tonal
Atonal
MLZC
80.78
79.82
79.21
79.82
80.29
79.91
ApEn
83.48
83.00
87.33
83.00
87.88
85.54
SampEn
82.56
83.00
67.24
84.36
84.95
85.99
PermEn
72.75
80.56
74.49
80.56
71.55
79.74
FuzyEn
80.29
80.92
78.23
80.92
83.23
82.64
ShanEn
73.85
75.23
75.87
75.23
75.50
75.96
LogEn
77.33
80.55
76.78
80.55
77.15
82.45
SpecEn
64.67
68.43
66.14
68.43
65.68
70.14
SSA
73.48
71.04
70.82
71.04
71.65
69.50
ARMO
56.78
63.54
57.43
63.54
57.52
62.28
Table 4: Mean CAs obtained for Complexity Metrics through NB classifier
Ac ce p
FS-I
FS-II
FS-III
10
Page 12 of 22
ip t
Table 5: Anova Test Results (p-values) for FS-II (** refers p<<0.0001, * refers p<0.0001)
P7
TP7
T7
FT7
F7
AF7
FPz
Tonal
**
0.3
*
0.1
0.7
**
0.01
Atonal
0.5
**
0.1
**
**
0.03
**
Right side
AF8
F8
FT8
T8
TP8
P8
Tonal
0.8
0.002
0.0002
0.01
**
0.005
Atonal
*
**
**
0.04
*
0.8
Ac ce p
te
d
M
an
us
cr
Left side
Figure 1: Statistical distribution of PermEn estimations in PMs
11
Page 13 of 22
ip t cr us an
M
Figure 2: Statistical distribution of PermEn estimations in NMs
words, both experience and knowledge about the acoustic stimuli cause the high
d
neuronal integration within neural ensembles together with at auditory cortex.
te
Tonal and atonal musical forms require discussion in regard to pleasure, primed memory, experience, fitness for music scales. Therefore, we can state the electrophysiological followings: Music processing is lateralized at mostly
Ac ce p
right hemisphere. A complex neural processing network is required in high cognitive musical functions such as musical attention, and the tracking of harmonic structure in time. Our current findings are compatible with the previous biological findings indicating that certain neurotransmitters (dopamine) is produced as a result of functions in amygdala affected by listening pleasure music [52]. The reference states that successfully tracking of tonal chords activates mostly prefrontal regions numbered by 44, 45, and 47 in Brodmann areas in addition to both anterior and posterior cingulate gyrus [53]. As well, musicians are capable of neural encoding of sounds due to excellent long-term musical training [54, 55]. PermEn is found to be best complexity method to observe superior cognitive functions of musicians due to the fact that auditory stimuli induce amplitude variations in single trials. PermEn was proposed as useful method for estimation 12
Page 14 of 22
of neural dysfunctions such as epilepsy [56], seizure [57] and OCD [58]. As well, the computational complexity of PermEn is lower. In this study, PermEn
ip t
provided the highest and successful results in every type of classification with
10-fold cross-validation. In nonparametric statistical problems, cross-validation
cr
methodologies have been used as a means of selecting tuning parameters. In
future work, the method, proposed to improve the reliability of cross-validation
us
[59], will be implemented in classification steps.
SVM classifiers have been frequently used to classify power spectra of EEG data including mental task [60] and seizure events [61] when the training set
an
is large, since the optimal decision function can be obtained through statistical learning theory. However, kernel function specification affect the classification performance. In this study, RBF was found to be more useful for entropy es-
M
timations, even if the gaussian kernels provide the relative improvements in classification performance for parametric and spectral estimations. However, NB classifiers provided the relatively much more improvements for the same
d
parametric and spectral estimations in comparison to SVM with gaussian ker-
te
nels. In reference, RBF network shows better generalization performance and computationally faster than SVM with Gaussian kernel, specially for large train-
Ac ce p
ing data sets [62]. In conclusion, both distribution of feature values and kernel types can be considered as performance parameters in EEG applications. The reference discuss the close relation between kernel function and classification accuracy for continuous time stationary process [63]. Feature set characteristics highly affect the performance of NB classifier. Baye
References
[1] Munte T.F. (2002) The musician brain as a model of neuroplasticity, Nature Review Neuroscience, 3:473-478. [2] Andrade P.E., (2003) Brain tuned to music, J. Roy. Soc. Med. 96, 284-287. [3] Peretz I., Hebert S. (2000) Toward a biological account of music experience, Brain Cognition, 42, 131-134. 13
Page 15 of 22
[4] Nozaradan S., Zerouali Y., et al. (2015) Capturing with EEG the neural entrainment and coupling underlying sensorimotor synchronization to the
ip t
beat, Cereb. Cortex 25, 736–747.
[5] Herholz S.C., Zatorre R.J. (2012) Musical training as a framework for brain
cr
plasticity: behavior, function, and structure, Neuron. 76(3), 486-502.
[6] Strait D.L., Kraus N. (2014) Biological impact of auditory expertise across
us
the life span: musicians as a model of auditory learning, Hear Res., 308, 109-21.
an
[7] Gaser C., Schlaug G. (2003), Brain structures differ between musicians and non-musicians, J. Neurosci. 23, 9240–9245
Cereb. Cortex, 13, 943–949.
M
[8] Hutchinson S., Lee L.H.L., et al. (2003), Cerebellar volume of musicians,
[9] Haslinger B., Erhard, P., et al. (2004) Reduced recruitment of motor asso-
te
Mapp. 22, 206–15.
d
ciation areas during bimanual coordination in concert pianists, Hum Brain
[10] J¨ancke L., (2009) Music drives brain plasticity, Biol. Rep., p.78, 10.3410,b1-
Ac ce p
78
[11] Brattico E., Tervaniemi H., et al., (2006) Musical scale properties are automatically processed in the human auditory cortex, Brain Res., 1117(1), 162-174.
[12] Fujioka T., Trainor L.J., et al. (2005) Automatic encoding of polyphonic melodies in musicians and nonmusicians, J. Cogn. Neurosci., 17(10), 15781592,
[13] Sowell E.R., Thompson P.M., et al., (2004) Longitudinal mapping of cortical thickness and brain growth in normal children, J. Neurosci. 24, 8223–8231.
14
Page 16 of 22
[14] Zhu B., Chai C., et al. (2017) Analysis of EEG complexity in patients with
ip t
mild cognitive impairment, J. Neurol Disord 5:354. [15] Thilakvathi B., Shenbaga Devi S., et al. (2017) EEG signal complexity anal-
ysis for schizophrenia during rest and mental activity, Biomedical Research,
cr
28(1), 1-9.
[16] Sabeti M., Katebi S., Boostani R. (2009) Entropy and complexity measures
tificial Intelligence in Medicine, 47(3), 263-274
us
for EEG signal classification of schizophrenic and control participants, Ar-
an
[17] Bosl W., Tierney A., et al. (2011) EEG complexity as a biomarker for autism spectrum disorder risk, BMC Medicine, 9-18.
M
[18] Tylovia L., Kukal J., et al. (2018) Unbiased estimation of permutation entropy in EEG analysis for Alzheimer’s disease classification, Biomedical Signal Proc. and Control, 39, 424-430
d
[19] Simons S., Espino P., Ab´asolo D. (2018) Fuzzy entropy analysis of the
te
electroencephalogram in patients with Alzheimer’s disease: Is the method superior to sample entropy?, Entropy, 20, 21.
Ac ce p
[20] Acharya U.R., Molinari F., Sree S.V. (2012) Automated diagnosis of epileptic EEG using entropies, Biomedical Signal Proc. and Control, 7(4), 401-408
[21] Liang Z., Wang Y., et al. (2015) EEG entropy measures in anesthesia, Front. Comput. Neurosci.
[22] Silva A.C.S., Arce A.I.C. et al. (2010) Quantifying electrode position effects in EEG data with Lempel-Ziv complexity, 32nd Annual Int. Conf of the IEEE EMBS, Argentina.
[23] Ab´asolo D., Simons S. et al. (2015) Lempel-Ziv complexity of cortical activity during sleep and waking in rats, Journal of Neurophysiology, 113(7), 2742-2752.
15
Page 17 of 22
[24] Hu J., Gao J, Principe J.C. (2006) Analysis of Biomedical Signals by the Lempel-Ziv Complexity: the Effect of Finite Data Size, IEEE Trans on
ip t
BME, 53(12), 2606-2609.
[25] Ab`asolo D., Hornero R., et al. (2006) Analysis of EEG background activity
cr
in Alzheimer’s disease patients with Lempel–Ziv complexity and central tendency measure, Medical engineering & physics, 28(4), 315-322
us
[26] Mohammadi G., Shoushtari P. et al., (2006) Person identification by using AR model for EEG signals, Proceed of World Academy of Science, Eng
an
and Techn, 11, 281–285.
[27] Wang T.W., Guohui W., Huanqing F., (2004) Autoregressive model order
M
property for sleep EEG, J. of Biomedical Eng, 21(3), 394-6. [28] Aydın S. (2009) Comparison of power spectrum predictors in computing coherence functions for intracortical EEG signals, Annals of BME, 37(1),
d
192–200.
te
[29] Faust O., Acharya R.U., et al. (2008) Analysis of EEG signals during epileptic and alcoholic states using AR modeling techniques, IRBM, 29(1), 44-52.
Ac ce p
[30] Vautard R., Yiou P., Ghil M. (1992) Singular-spectrum analysis: A toolkit for short, noisy chaotic signals, Physica D, 58, 95–126.
[31] Aydın S., Demirta¸s S. et al. (2016) Emotion recognition with eigen features of freauency band activities embedded in induced brain oscillations mediated by affective pictures, Int. J. of Neural Systems, 26(3), 1650013.
[32] Kouchaki S., Sanei S., et al. (2015) Tensor based singular spectrum analysis for automatic scoring of sleep EEG, IEEE Trans on Neural Systems and Rehabilitation Eng., 23(1) [33] Jasper H. (1958) The ten-twenty electrode system of the international federation, EEG and Clinical Neurophysiology, 10, 371–375.
16
Page 18 of 22
¨ oren M., Erdogan U., et al. (2009) Brain asymmetry measurement using [34] Ozg¨ EMISU in applied brain biophysics, Computers in Biology and Medicine,
ip t
39(10), 879–888.
[35] Aboy M., Hornero R., Ab´asolo D., Alvarez D. (2006) Interpretation of the
cr
Lempel-Ziv complexity measure in the context of biomedical signal analysis, IEEE Trans on BME, 532282–2288.
us
[36] Tosun P.D., Ab´asolo D., Stenson G., et al. (2017) Characterisation of the effects of sleep deprivation on the electroencephalogram using permuta-
an
tion lempel–ziv complexity, a non-linear analysis tool, Entropy, 19, 673, doi:10.3390/e19120673.
[37] Aydın S. (2010) Determination of autoregressive model orders for seizure
M
detection, T. Journal of Elec Eng & Comp Sci, 18(1), doi:10.3906/elk-090683
d
[38] Andrzejak R.G., Lehnertz K., et al (2001) Indications of nonlinear deter-
te
ministic and finite dimensional structures in time seires of brain electrical activity: Dependence on recording region and brain state. Physical Review
Ac ce p
E, 64, 061907.
[39] Vapnik V.N. (2000) The Nature of Statistical Learning Theory. New York, USA Springer.
[40] Webb A.R., Copsey K.D. (2011) Statistical pattern recognition, 3rd. Hoboken, NJ, USA, John Wiley & Sons,
[41] Lin L.C., Ouyang C.S. et al. (2014) Early prediction of medication refractoriness in children with idiopathic epilepsy based on scalp EEG analysis, Int. J. Neural Syst. 24(7):1450023. [42] Wang Y., Zhou W. et al. (2013) Comparison of fractal features of ictal and interictal EEGs, Int. J. Neural Syst. 23(6):1350028.
17
Page 19 of 22
[43] Jumutc V., Zayakin P. et al. (2011) Ranking based kernels in applied biomedical diagnostics using a support vector machine, Int. J. Neural Syst.
ip t
21(6):459–473.
[44] Amin H.U., Mumtaz W. et al. (2017) Classification of EEG signals based
cr
on pattern recognition approach, Frontiers in Comput Neurosci, 11: 103.
[45] R. O. Duda, P. E. Hart, and D. G. Stork. (2001) Pattern Recognition,
us
second edition. Wiley Inter Science.
[46] K. Fukunaga. (1990) Statistical Pattern Recognition, seconde edition. Aca-
an
demic Pres.
[47] S. Lemm, C. Schafer, and G. Curio. (2004) Bci competition 2003–data set
M
iii: probabilistic modeling of sensorimotor mu rhythms for classification of imaginary hand movements. IEEE Trans on BME, 51(6):1077–1080. [48] S. Solhjoo and M. H. Moradi. (2004) Mental task recognition: A comparison
d
between some of classification methods. In BIOSIGNAL 2004 International
te
EURASIP Conference.
[49] Z. A. Keirn and J. I. Aunon. (1990) A new mode of communication between
Ac ce p
man and his surroundings, IEEE Trans on BME, 37(12).
[50] G. A. Barreto, R. A. Frota, and F. N. S. de Medeiros. (2004) On the classification of mental tasks: a performance comparison of neural and statistical approaches. In Proceedings of the IEEE Workshop on Machine Learning for Signal Processing.
[51] Rish I. (2001) An empirical study of the naive Bayes classifier, IJCAI 2001 workshop on empirical methods in artificial intelligence, 3, page 41–46. IBM New York [52] Levitin D.J. (2012) What Does It Mean to Be Musical ?, Neuron, 73(4):633637.
18
Page 20 of 22
[53] Huron D. (2006) Sweet Anticipation: Music and the Psychology of Expec-
ip t
tation, MIT Press, Cambridge, MA. [54] Thatcher R.W., Biver C.J., North D. (2007) Spatial-temporal current source correlations and cortical connectivity, Clin. EEG Neurosci. 38,
cr
35–48.
[55] Fair D.A., Cohen A.L., et al. (2009) Functional brain networks develop from
us
a “local to distributed” organization. PLoS Comput. Biol. 5, e1000381.
[56] Veisi I., Pariz N., Karimpour A. (2007) Fast and robust detection of epilepsy
an
in noisy EEG signals using permutation entropy, IEEE 7th Int. Symposium on BioInformatics and BioEngineering, 200-203.
M
[57] Mammone N., Lay-Ekuakille A., et al. (2011) Analysis of absence seizure EEG via permutation entropy spatio-temporal clustering, IEEE Int. Symposium on Medical Measurements and Applications, 532-535.
d
[58] Aydın S., et al. (2015) Classification of obsessive compulsive disorder by
te
EEG complexity and hemispheric dependency measurements, Int. J. of Neural Systems, 25, 1550010.
Ac ce p
[59] Wang Q., Lindsay B.G. (2015) Improving cross-validated bandwidth selection using subsampling-extrapolation techniques, Computational Statistics & Data Analysis, 89, 51-71.
[60] Garrett D., Peterson D.A., Anderson C.W., et al. (2003). Comparison of linear, nonlinear, and feature selection methods for EEG signal classification, IEEE Trans on Neural Systems and Rehabilitation Eng. 11(2), 141-144.
[61] Park Y., Luo L., Parhi K.K., et al. (2011) Seizure prediction with spectral power of EEG using cost-sensitive support vector machines, Epilepsia, 52(10):1761–1770. [62] Debnath R., Takahashi H. (2002) Learning Capability: Classical RBF Network vs. SVM with Gaussian Kernel. In: Hendtlass T., Ali M. (eds) Devel19
Page 21 of 22
opments in Applied Artificial Intelligence. IEA/AIE 2002. Lecture Notes
ip t
in Computer Science, vol 2358. Springer, Berlin, Heidelberg [63] Heda K.E., Louani D. (2018) Optimal bandwidth selection in kernel density estimation for continuous time dependent processes, Statistics & Probabil-
Ac ce p
te
d
M
an
us
cr
ity Letters, 138, 9-19.
20
Page 22 of 22