Music's manifold emotions for mental well-being: a

topological approach
Joseph Darcourt (  [email protected] )
Olifeclass https://orcid.org/0009-0009-1820-2691

Biological Sciences - Article

Keywords:

Posted Date: November 28th, 2023

DOI: https://doi.org/10.21203/rs.3.rs-3602878/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Additional Declarations: There is NO Competing Interest.

Joseph Darcourt
Music, an intrinsic part of human culture, has long been recognized for its ability to evoke emotions,
memories and physiological responses. From the rhythmic beats that make us dance to the soothing
melodies that provide solace in times of distress, music's impact on the human psyche is undeniable.
Timely, in the wake of the COVID-19 pandemic, the world has witnessed a subsequent and equally
concerning mental health "pandemic.” As societies grapple with the repercussions of prolonged isolation,
economic downturns, and the collective trauma of a global health emergency, the search for therapeutic
interventions has never been more pressing. Among the myriad of potential solutions, one timeless human
experience stands out: the power of music. Our study aims to understand how music’s benefits for well-
being can be harnessed to combat the rising levels of neurodegenerative diseases and depression. By
exploring the neural manifolds and latent neural dynamics of music listening and emotional experience, we
aim to uncover the therapeutic potential of music through topology. Our findings not only shed light on
music’s role in mental well-being through emotional synchrony across subjects, but also offer hope in times
of crisis. Alongside shared neural dynamics across subjects, we also find one subject in which a major
difference in emotional experience leads to a drastically different manifold shape, clustering and coloring.
These results are the first to pain music listening, emotional experience and well-being in a topological
light. We see our results positioned as a backdrop for future research in computational neuroscience on
music and well-being, our study being the first to approach mental well-being and music listening with
topology. We also outline next steps that would extend our preliminary results immediately to
neurodegenerative diseases, depression and more. In a world where mental health issues are on the rise,
this topological approach to music and well-being can provide deeper understanding to music therapy’s
non-invasive, accessible and enjoyable avenue for improving the well-being of countless individuals; by
accounting for inter-subject variability with topologically-inspired computational neursocience, we can
finally put a name to how music helps heal.

The COVID-19 pandemic has had a wide-ranging times of distress, music's impact on the human psyche
negative impact on mental health globally1-12; is undeniable. But could there be more to this
depression, now often considered the world’s most relationship than meets the ear? Could the intricate
common mental disorder, has a strong connection to patterns of neural activity elicited by music hold the
emotional function, cognitive capacity and general key to understanding and bettering our mental well-
well-being; on the same token, risk for developing being?
neurodegenerative diseases has risen steadily, along We explore these questions here with a topological
with worsened conditions for those already with approach inspired by oodles of complexity theory.
neurodegenerative disease, with the pandemic13-16. At First, we address topology: In years recent, advances in
the same time, treatments have gradually started recording methods have enabled researchers to delve
including music-listening therapy more and more, into studying neural population dynamics, gradually
gaining traction in both well-being and professional moving beyond a single-neuron focus and towards the
rehabilitation circles17-27. population doctrine where neural circuitry and neural
Music, a physical, emotional and spiritual populations contribute to function and complex
phenomenon placed at the core of what it means to be behaviors, instead of the single neuron as the main
human, has long been recognized for its ability to functioning component focus29-33. This has led to rapid
evoke emotions, memories, and even physiological growth in neural manifold research, which posits the
responses28. From the rhythmic beats that make us following: high dimensional, “latent (hidden) neural
dance to the soothing melodies that provide solace in dynamics”34 can be represented by low-dimensional
manifolds/topologies; these topologies of neural emphasize that this is merely a choice of nomenclature
population activity patterns directly correspond with to extend our findings in latent neural dynamics of
and represent both the task condition of interest38, and music listening and emotional experience to the notion
function in and across cross-functioning sensory, of greater well-being, which is a complex, holistic
motor, cognitive and emotional brain regions30,34,38. phenomenon important to everyday functioning and
However, such manifold-based approaches in human, dealing with diseases, however, this naming is not
non-human primate and even rodent neuroscience explored much more in depth in our study besides
research have been limited in their generalizability as denoting the sum of each subjects' sum of emotional
they are often studied in the context of simple and states in their UMAP manifold spaces. A future study
stereotypical motor tasks34-36, rather than more could investigate other aspects of emotional experience
complex, naturalistic settings, let alone naturalistic like recall of emotions, or shared emotional experience
music listening. Likewise, no research has been done with those around one, but for now we stick to this
to relate neural manifolds to complex human emotional naming simply to account for a wide range of
experience, let alone music, with the music subfield of emotional states in a subject and his/her UMAP
neuroscience research instead focusing on stochastic embedding, and leave other possibilities for future
processes and simple “top-down,” unidirectional research.
predictive coding music-emotion processing models37, In this study, we aim to shed light on music (classical
rather than the population/network-level perspective music specifically) and its undeniable connection to
and insights on emotion and cognition that topology improving mental well-being. We explore the
can offer38. qualitative qualities—the geometry of the neural space,
Then, addressing complexity theory: By opting to use like structure, shape/clustering and colors/emotions of
topology and its neural population-level each subjects’ UMAP-embedded manifold space—of
techniques/algorithms to understand changing latent neural dynamics of music listening and
emotional states in listeners, we take a topological emotional experience (now dubbed “music listening-
approach that is inspired by complexity theory38. emotional experience latent neural dynamics” for
Starting off, in emotion research, valence and arousal simplicity), that is to say, latent neural dynamics
are consistently the most common measures for meaning the inherent properties of brain activity
measuring emotional states39-41, with valence values patterns that we cannot deterministically compute in
representing low to high positivity of emotions and quantitative form due to the overwhelming complexity
arousal values representing low to high excitement or of the brain, so we project this high dimensional
engagement of emotions41. However, the majority of brain/neural data and its activity patterns onto a low
emotion research has focused on classifying emotions dimensional neural manifold space48. Through these
into discrete bins/categories (e.g., referring to emotions layers, we distill an interdisciplinary approach that
over a large period time simply as high valence-low aims to further understand music’s implications in
arousal, without regard to the dynamic changing nature bettering well-being through population/network-level
of emotion), which would fail to take into account the techniques: a topological approach. By exploring
dynamic nature of music, music listening and its music listening-emotional experience latent neural
impact on continuous, non-static emotional states for dynamics, we can, with time, uncover the therapeutic
our case42-43. Therefore, going forward in this study we potential of music with a dynamical approach,
refer to each subject’s collective emotional states in extending impact to a wide range of modalities,
each respective UMAP embedding as his/her own ranging from everyday life to depression to
“complex emotional experience” (or simply “emotional neurodegenerative diseases. Our findings not only shed
experience” for short). This takes into account the light on music’s role in mental well-being through
temporally-varying “experience” (rather than emotional synchrony across subjects, but also offer
discretized time snippets of neural data and mere hope in times of crisis. In a world where mental health
snapshots of static neural-emotional responses) of issues are on the rise, this topological approach to the
emotion, as well as nomenclature and theories from music of well-being will provide a non-invasive,
psychology, philosophy and complexity theory accessible and enjoyable avenue for improving the
concerning subjective, neurophysiological and well-being of countless individuals.
behavioral responses over time emerging (from smaller
parts of functioning, complex neural systems) in the Results
form of individual emotional experiences as part of We analyzed the Openneuro dataset, “A dataset
greater consciousness44-47. Still, we do want to recording joint EEG-fMRI during affective music

2
listening,” to investigate the latent neural dynamics of (top-right corner of the color bar) corresponds to high
classical music listening and emotional states in this arousal and high valence, which such as a color like
study. There were three trial types; we selected the red or orange, associated with emotions like
music-and-reporting trial type in which subjects were excitement; intermediate values would represent a mix
tasked with reporting their continuous emotions on a 2- of valence and arousal levels. For example, a color that
D FEELTRACE interface49-50 for valence and arousal appears in the middle of the color bar might represent
while listening to any of four classical music pieces. moderate levels of arousal and valence, corresponding
Two of the four pieces were selected to demonstrate with an emotion of contentment.
similarities and differences in latent neural dynamics
between similarly-structured music pieces, and Musical phrases and pieces. From the dataset, out of
subsequently, different emotional states with the 4 classical music pieces we selected the 2 musical
implications in mental well-being; we below lay the pieces that were most similar in structure (e.g.,
groundwork of our pipeline/procedure, as well as detail similarity in key, length, tempo, repetition of specific
findings. musical ideas, or motifs, musical texture), as such
musical similarity has been shown to evoke similar
Music and frequency bands. Over the past two emotions65-66 and could lead to more tangible results
decades, the intersection of music-neuroscience across subjects. The two musical pieces were
research has leveraged a wide array of neuroimaging “Chopin’s Etude No. 3 in E Major, Opus 10,” and
techniques, primarily electroencephalogram (EEG) and “Rachmaninoff’s Prelude No. 5 in G Major, Opus 32”;
functional magnetic resonance imaging (fMRI). The they will be referred to as “the Chopin” and “the
seminal work “The Mozart Effect”51 that first Rachmaninoff,” respectively. Their similar qualities as
introduced the the brain’s natural responses to follows: both are in a major key, with the Chopin
periodicity, or consistent patterns in music, has changing key (modulating) more often, whereas the
evolved52 with numerous electrophysiological (EEG) Rachmaninoff is more steady and stays in G major
and brain connectivity (fMRI) studies investigating (with some oodles of E minor) throughout the piece;
how music influences brain activity and emotional both are similar in length, with the Chopin being 2
processing53-61. Notably, the joint EEG-fMRI study that minutes and 30 seconds long and the Rachmaninoff 2
our dataset is came from demonstrated that music minutes and 7 seconds long; both have long musical
perception reflects the asymmetry of the alpha ideas/motifs that are slurred across bar lines; and the
frequency band50, approach-withdrawal behavior and musical textures, with both being piano pieces that are
emotional attention during listening; similarly, the not too loud, are largely similar. Without the
theta band has been shown to be heavily involved in confounding factor of drastically-different musical
music and emotion60. style or 21st century’s extremely dissonant/atonal music
For our study, we selected 3 out of 21 subjects in our (the Chopin was composed in the 19th century, and the
dataset based on similarities in EEG patterns during Rachmaninoff the early 20th) that might cause
initial EEG channel visualizations, reducing inter- discomfort and/or distort valence/arousal responses in
subject variability; 25 out of 31 EEG channels for each the subjects67, we posited that similarities in the
subject were used and we focused on the delta, theta selected pieces would allow us to find tangible
and alpha bands for their role in emotional processing, similarities and/or differences in the population-level
meditation/relaxed physical states and music cognition latent neural dynamics across subjects, related to
emotional experience.
and emotion, respectively60,62-64.
Subsequently, we selected one phrase from each
piece that were inherently different in nature; both
Continuous valence-arousal color scheme for
phrases are 22 seconds in length. With the
emotion representation
characteristics from Fig. 1 in mind, we posited the
To represent each subjects’ emotional experience in
following: musical phrase 1 (from the Chopin) would
our data, we constructed a continuous 2-D color
elicit more high arousal (red in our color spectrum as
scheme (Fig. 2g) for representation of the subjects’
excitedness) emotional responses across subjects, with
continuous FEELTRACE-reported valence and
no assumption on valence, whereas musical phrase 2
arousal. A 0 on both axes (bottom-left corner of the
would elicit more emotions of low-arousal and low-to-
color bar) corresponds to a color representing low
high valence emotional responses (yellow or green in
arousal and low valence, which such as a color like
our spectrum as happiness or a feeling of contentment,
dark blue or green, typically associated with emotions
respectively).
like sadness or extreme calmness; a 1 on both axes

3
 The full score of each piece was not included for experiences, we used the UMAP algorithm, an
consideration of space; both phrases are shown, unsupervised manifold learning technique that
alongside detailed characteristics that should elicit preserves both local and global structures in complex
common emotional responses, in Fig. 1. data68 (see Methods for details on parameters). As it is
UMAP discovers shared neural activity across a newer technique, complex data analysis done with
subjects. To capture the latent neural dynamics of UMAP in recent years has often only consisted of
classical music listening and subsequent emotional

a Musical Phrase 1: Chopin’s Etude No. 3 in E Major, Opus 10

b Musical Phrase 2: Rachmaninoff’s Prelude No. 5 in G Major, Opus 32

Fig. 1 | The musical phrases. a, Musical phrase 1 is an excerpt in the development section of Chopin’s Etude in E Major.
This is the first section in the Etude to move away from E Major, characterized by frequent changing of key (modulating)
between measures; syncopations in rhythm that might increase arousal; and rising cascades of notes (denoted by the long
slurs over notes) that are dissonant in nature and should elicit moment-to-moment (as the dissonances resolve themselves on
the scale of milliseconds (ms)) emotional changes and increases in arousal across subjects. b, Musical phrase 2 is the opening
of Rachmaninoff’s Prelude in G Major, which has a continuous left-hand (bottom line of each staff) note pattern that
establishes a consistent, flow-like musical texture, and a simple right-hand (top line of each staff) step-wise melody.

datasets with long time series69-74, ranging from hours bring this up as a point in Discussion), and thus we
to days to weeks of data. Still, the precise mathematical reasoned that the clustering capabilities of UMAP
foundations laid by the original authors of UMAP do would be largely successful at extracting meaningful
not restrict application to only large datasets68, results on our smaller-scale music-and-reporting trial
although they do mention care with interpretation (we

4
data, showing different clustering and distance patterns cluster of points in the top right (subject 2’s embedding
with different emotions. has a similarly-structured cluster, but in the bottom
Likewise, we were inspired by the work of Fortunato right of the embedding), a cluster in the middle-top and
et al., 2023, who found that tasks of increasing a cluster in the bottom left in a sort of curved shape.
complexity were paired with increasing nonlinearity of Tackling emotional experience, we can see that
discovered neural manifolds36.Thus, we posited that generally each subject's emotions range from sad (dark
music listening, being such a complex phenomenon blue manifold points) to peacefulness (light blue
with emergent emotional experience as discussed, manifold points) to excitement (red manifold points).
would be better represented with nonlinear manifold Importantly, there are more red manifold points than
learning techniques, such as UMAP, but future blue (either dark or light), indicating a higher arousal
investigations should try comparing our results with level across subjects during musical phrase 1 which is
UMAP with those of a simple, linear dimensionality important to note as earlier we mentioned that musical
reduction (DR) technique, like principal components phrase 1 introduced new musical ideas in the piece and
analysis (PCA). This is a simple next step and could was generally characterized by lots of dissonance and
give intriguing insights into the nonlinearity of the tension in the music—which reflects in listeners'
phenomena we tackle here—that is, showing that linear emotional experience. Thus, the above-mentioned
DR yields less satisfactory results compared to minor differences between cluster locations on each
nonlinear DR, such as UMAP as we used, through the subject’s UMAP manifold embedding during musical
observation that linear DR modes cover less variance. phrase 1 are accounted for by similarities in emotions;
In order to understand how the musical content of a furthermore, we found that subjects’ manifold
piece relates to latent neural dynamics and the embeddings each had a similar Spearman rank-
emergent emotional experiences of listeners, we correlation coefficient, ρS (see Methods for
focused on the qualitative qualities of each subject's calculation), with -0.03, -0.03 and -0.01 for each
UMAP embedding: overall structure/shape of the subject respectively. Although these values are low,
manifold, clustering and colors/emotions of manifold they show near-similarity in magnitude and thus give
points, which are all visually examined as properties of us a generalization of how the latent neural dynamics
the UMAP-embedded neural manifold space’s of each subject during musical phrase 1 are shared in
geometry30,38. For this section, we calculated the relation to emotion: there is a weak, yet statistically
UMAP embedding for every 100 ms epoch in each significant correlation between latent neural dynamics
subject's musical phrase 1 and 2 EEG data, and emotional experiences when embedded with
respectively, which was pre-processed, filtered and UMAP, and future research could investigate further
epoched prior using UMAP (see Methods). with different metrics (as the relationship is highly
Our first finding was that the latent neural dynamics nonlinear and complex, so monotonicity and
of subjects 2, 10 and 17 during musical phrase 1 (the Spearman's rank-correlation coefficient is really low).
Chopin) were similar in nature, as shown in Fig. 2a,b,c. Our next finding substantiates our first’s: Subject 2,
Although there were minor differences in where 10 and 17’s UMAP embeddings of latent neural
clusters were exactly located, one can observe similar dynamics during musical phrase 2 also shared similar
point-to-point distances overall when comparing across point-to-point distances within clusters (Fig. 2d,e,f,),
subjects, which is what matters when looking at even if clusters were located in slightly different spots
UMAP embeddings; similarly, UMAP is often used as in each embedding, just like that for musical phrase 1.
a clustering algorithm, so seeing similar point-to-point To be briefly specific, for all three subjects during
distances in each cluster, despite differences in exact musical phrase 2 (Fig. 2d,e,f) there is a circle-like
cluster location, is still indicative of shared neural cluster in the middle, a generally tail-like cluster in the
dynamics. In detail, in subject 2’s neural manifold right and a line-like cluster in the left-middle of the
during musical phrase 1 we can see a cluster of points UMAP embedding.
in the middle of the manifold, as well as other clusters Circling back to emotion, like in musical phrase 1’s
of points to the left, right, and bottom of this middle subject embeddings, in musical phrase 2 we observe
cluster. Similarly, during musical phrase 1 subject 10’s similar emotions throughout (generally, there are red,
latent neural dynamics have clusters of points in the dark blue and light blue manifold points in each
middle, left, right and top of the UMAP embedding, subject’s embedding), however there are more yellow
although these clusters are positioned slightly different manifold points in subject 10’s UMAP manifold
to those of subject 2’s embedding: instead, subject 10’s embedding. This is, as we posited, likely due to
clusters are more similarly-positioned to subject 17’s, musical piece 2 and musical phrase 2’s (as musical
with both having a cluster of points in the middle, a phrase 2 is the opening of musical piece 2) steady,

5
soothing musical texture and consistency/steadiness of Finally, the embeddings had ρS values of 0.02, -0.01
harmony, leading to subject 10's UMAP embedding and 0.01 for subjects 2, 10 and 17 respectively, with
capturing these more soothing musical qualities in the similar magnitudes of correlation similar to those of
emergence of more yellow manifold points and more musical phrase 1.
relaxed emotional experience. And importantly, this These results further substantiate our hypothesis that
occurrence of more yellow manifold points and a more a manifold learning technique like UMAP is able to
relaxed emotional experience leads to subject 10’s capture (albeit weakly) the correlation between latent
UMAP embedding being more confined to strict neural dynamics and music-listening emotional states;
clusters/paths of manifold points; more on this later. this is all reflected in the similarity of clusters, point

a Subject 2’s latent neural dynamics during musical phrase 1

b Subject 10’s latent neural dynamics during musical phrase 1

6
c Subject 17’s latent neural dynamics during musical phrase 1

d Subject 2’s latent neural dynamics during musical phrase 2

e Subject 10’s latent neural dynamics during musical phrase 2

7
f Subject 17’s latent neural dynamics during musical phrase 2

g Valence-arousal color spectrum.

Fig. 2 | Shared neural dynamics across subjects 2, 10 and 17 during musical phrase 1 and musical phrase 2,
respectively. (For a-f, horizontal axis is UMAP component 1 and vertical axis is UMAP component 2.) a, Subject 2’s
UMAP-embedded latent neural dynamics during musical phrase 1 (shown in Fig. 1a). b, Subject 10’s UMAP-embedded
latent neural dynamics during musical phrase 1. c, Subject 17’s UMAP-embedded latent neural dynamics during musical
phrase 1. Subjects 2, 10 and 17 had a ρS of -0.03, -0.03 and -0.01, respectively, during musical phrase 1; all statistics
significant at p < 0.05. d, Subject 2’s UMAP-embedded latent neural dynamics during musical phrase 2 (shown in Fig. 1b)
shows a characteristic pattern. e, Subject 10’s UMAP-embedded latent neural dynamics during musical phrase 2. f, Subject
17’s UMAP-embedded latent neural dynamics during musical phrase 2. Subjects 2, 10 and 17 had a ρS of 0.02, -0.01 and
0.01, respectively, during musical phrase 2; all statistics significant at p < 0.05. g, The valence-arousal emotion color
spectrum used for projecting onto each manifold point in a-f (and 3a-c).

distances and manifold point colorings for valence- second Rachmaninoff, "musical piece 2" (Fig. 3a,b,c).
arousal as discussed above. It is likely that the We calculated the UMAP embedding for every 500 ms
dynamics we are trying to uncover are not a monotonic in the EEG neural data, as musical piece 2 is longer in
function, which the Spearman rank-correlation length than either musical phrases 1 or 2, and applied
coefficient requires for a strong value, so future the same pre-processing and filtering steps (see
investigations should use different correlation or even Methods).We posited that since musical piece 2 is has
causality measures as a means to understand this an overall consistent structure, with a smooth texture,
complex dynamical (emotions smooth articulation and very few tension points in the
unfold over time, hence the different positioning of piece, that latent neural dynamics for the
colors in the UMAP space) and spatial (the qualitative Rachmaninoff would not only be similar across
qualities like clusters and distances between points of subjects, but also the UMAP embedding would have
UMAP space as discussed) relationship further. happier and more relaxed emotional states (represented
by yellow and green dots, respectively).
Differed latent neural dynamics across subjects in a We found that this held for subjects 10 and 17: there
musical piece through differences in emotional was a wide range of colors, ranging from red and
experience. Taking a similar qualitative quantities light/dark blue, like in musical phrases 1 and 2, to
approach to the above UMAP manifold embeddings yellow and green. Like in the cases of musical phrases
for musical phrases 1 and 2 for subjects 2, 10 and 17, 1 and 2’s subject
respectively, we applied UMAP on the 2 minute and 7

8
a Subject 10’s latent neural dynamics during musical piece 2

b Subject 17’s latent neural dynamics during musical piece 2

9
c Subject 2’s latent neural dynamics during musical piece 2

Fig. 3 | Subjects 10 and 17 share neural dynamics; subject 2 has different dynamics. (For a-c, horizontal axis is UMAP
musical piece 2, which is the entire length of Rachmaninoff’s Prelude No. 5 in G Major, Opus 32. There is a combination of
red, yellow, green, light blue and dark blue manifold points, representing a wide range of emotional states: excited (red for
high arousal), happiness (yellow and orange for high arousal, high valence), peaceful (light blue for low arousal and high
valence) and sadness (dark blue for low arousal and low valence). b, Subject 17’s UMAP-embedded latent neural dynamics
during musical piece 2 show a characteristic pattern similar in color (projected valence/arousal emotions), shape and spread
to subject 2’s dynamics. c, Subject 17’s UMAP-embedded latent neural dynamics during musical phrase 2. Although there
are blue, yellow and red colorings, the majority of manifold points are green, which represents calmness or content and
reflects in the little variance in manifold points/distances and neural dynamics. Subjects 10, 17 and 2 had a ρ S of 0.04, 0.04
and -0.01, respectively; all statistics significant at p < 0.05.

embeddings, we found that musical piece 2 subject 10 continuum of low arousal and high valence, which can
and 17's UMAP embeddings were similar in both be interpreted as an overall emotion of contentment,
manifold point distances and cluster location, reflecting and thus we proved a case in which latent neural
the wide range of emotions in a widely-spread-out dynamics of one subject differed from others due to its
point manifold. different emotional states over the length of one
However, when we applied the UMAP algorithm to musical piece. And going back to our definition of
subject 2’s musical piece 2 music-and-reporting data, emotional experience, this case shows us one situation
we found a different result: there were much more in which differed emotional experience leads to
green manifold points, and consequently, the overall differed emotional experience, which can filter into
shape, pattern, cluster locations and manifold point many therapeutic interventions down the road once we
distances were entirely different from subject 10 and get closer to understanding the underlying neural
17’s respective embeddings. The data is more spread mechanisms (the latent neural dynamics, as
out into definable clusters than subject 10 and 17’s, investigated through topology) of these differences.
and we reason that this is due to the greater of green Similarly, this is directly related to the “neural-
manifold points to other colors; green represents a dynamics space hypothesis,” which states that a

10
diverse range of behaviors can be described via synchrony across UMAP-embedded subject manifold
different trajectories within their own neural spaces during musical phrases 1 and 2, respectively,
spaces38,75, which in our case is the subject variability underscores shared latent neural dynamics across
in emotional experience leading to differences in subjects while listening to the same phrase. We then
qualitative quantities of each UMAP manifold found similar qualitative qualities of UMAP space
embedding spaces, most drastically seen in subject 2 when analyzing subjects 10 and 17, respectively,
musical piece 2 but also in minor differences in every during musical piece 2. However, subject 2’s musical
subjects’ UMAP embeddings’ qualitative quantities piece 2 data revealed strikingly different results: we
across musical phrases 1 and 2 and musical piece 2. observed differed latent neural dynamics through an
This way of viewing our results and complex extreme contrast in shape, clustering and
behaviors, including emotional experience, as a result colors/emotions of its UMAP space (Fig. 3c). We posit
of latent neural dynamics that include the differences that subject 2’s greater occurrence of bright green
in musical pieces, is a topological approach that takes manifold points during musical piece 2—
into account a wide range of properties, from both the corresponding to an overall greater emotional state of
EEG neural data and the “task condition,” 38 or in our contentment (low arousal and high valence)—is
case the musical piece/phrase being listened to. Future heavily related to its differences in qualitative qualities
results could further investigate the relationship of its UMAP space when compared to subjects 10 and
between the “task condition,” or musical piece/phrase, 17.
by converting the musical piece or phrase being played Our findings, although exploratory and preliminary
into forms—like converting the sound file into simple in nature with only visual inspection of UMAP
qualities like loudness and timbre; or, into more qualitative quantities done so far, are extremely
complex, math-music theory computations like the promising; they serve as a backdrop for future research
Fourier Coefficient/phase of a musical phrase76 or on music and well-being from this very topological and
motivic topologies across a musical piece77—that are population-level perspective. Before we go any further
feedable into a technique like canonical correlation on future steps, however, we would like to address a
analysis (CCA)34,78 or its nonlinear variants79-80, to few limitations in our study, along with doable
unravel the direct correlation(s) between changes in adjustments with time. First, the size of our dataset. As
musical features and the latent neural dynamics, which we tried capturing the latent neural dynamics inherent
also reflect emotional experiences as discussed (with to each subject’s individual emotional experience
manifold point colorings). during an isolated phrase or during an entire musical
piece, we chose to only analyze continuous EEG data
Discussion within each music-and-reporting trial, that is, with a
maximum length up to 2 minutes and 7 seconds in
Everyone knows the power of music in bringing
length for musical piece 2, for which we extracted
diverse peoples together and weaving a fabric of
features from 500 ms epochs (as any shorter length of
common human experience81-82, promoting greater
sub-epochs would get us too few UMAP manifold
mental and physical well-being for listeners in the
process83-86. Yet, until now there has not been any fool- points for interpretation); and EEG data up to 22
seconds for musical phrases 1 and 2, respectively, with
proof concept to date on the complex workings of the
100 ms epochs (also for reason of keeping an
brain in relation to music and a clear connection with
interpretable amount of UMAP manifold points). With
its promoting well-being.
these feature epoch lengths in mind, UMAP, while
Our exploration into the qualitative qualities—
robust, may also be sensitive to the size of our
structure, shape/clustering and colors/emotions—of
dataset68, keeping in mind the large datasets that
subjects’ UMAP embeddings and latent neural
UMAP has usually been performed on recently69-74,
dynamics of music listening and emotional experience
and could benefit from future research with larger and
started with analyzing subject 2, 10 and 17’s music-
more granular datasets. One way to obtain larger
and-reporting trial data during musical phrases 1 and 2,
datasets in our case of music listening and emotional
respectively. We observed similar qualitative qualities
experience is extending length of experimental trials to
and thus shared dynamics: all 3 subjects shared overall
UMAP-space structure, clustering distance patterns yield longer musical listening sessions for subjects, so
as to obtain the most naturalistic setting of music
and occurrence of similar emotions across each phrase,
listening (at least for a classical music concert),
respectively, as evinced by similar color types and
quantities in the manifold points (Fig. 2a-c for musical without the constant stops and starts of 2-minute-or-so
trials that we analyzed, while also allowing for more
phrase 1; Fig.2d-f for musical phrase 2). This
data to feed into features for UMAP. This longer set of
11
EEG data per subject would allow for more UMAP those with Parkinson’s100-102, Alzheimer’s103-106 and
manifold points to work with, and subsequently, more schizophrenia107-109; this topological approach is
opportunity to perform robust computations on the especially relevant as such diseases show large inter-
neural manifold space; more points will allow us to subject variability with music therapy treatment, and
look at a more global and emergent, rather than local, our UMAP-embedded manifold space for subject 2
latent neural dynamics manifold structure. Having musical piece 2 (Fig. 3c) showed great variability in
more EEG data would also allow us to raise our qualitative qualities due to different emotional
n_neighbors UMAP parameter (see Methods) to 15, 20 experience, and so our topological approach, when fed
or more, which might allow us to find more common more features with relevance to these diseases, can
topological shapes—more and more sphere-like— inherently help explain the inter-subject variability in
rather than the very distinct UMAP shapes we found. music therapy with these diseases. With the benefit of
Another option is also extending our UMAP to 3 improved computational resources, larger data and
dimensions as is possible with more data. time, computing and inputting these features will
The second point to address is the features we become more feasible.
computed on our data. For the aforementioned 500 ms Directing forward, we can connect our findings to the
and 100 ms epochs, respectively, we extracted the greater fields of computational neuroscience, emotion
power spectral density (PSD) matrix87 from the theta research and bettering mental well-being for all alike:
and alpha bands, respectively, as well as the Pearson By capturing the neural correlates of inter-subject
correlation coefficient (PCC) between each EEG differences/variability during musical listening and
channel pair88. Now, although our findings were subsequent emotional experience with UMAP, we
compelling—the first to note similarities in qualitative define a new experimental modality through
quantities in music listening-emotional experience topological methods like dimensionality reduction
latent neural dynamics across subjects, with one (DR), specifically UMAP here, for unveiling the
difference being due to difference in overall emotional complex relationship between neural dynamics, music
experience (Fig. 3c), yet another promising future and emotion, which can feed into translatable results
avenue to focus on—with merely the two PSD on a subject-to-subject basis for therapeutic
matrices and PCC per pairwise channel, near-future interventions and greater well-being going forward110-
research should indeed move towards inputting 111
. We can—by investigating subjects’ differences in
nonlinear and complexity theory-related features into latent dynamics during emotional experience (as
the UMAP algorithm, including but not limited to: probed by music listening)—gain valuable insights into
Higuchi's fractal dimension, multifractal detrended other neural modalities of thinking (cognition, and its
fluctuation analysis (MDFA), correlation dimension heavy connection to and arguably inseparability from
(CD), largest Lyapunov exponent (LLE) and Shannon emotion112), anxiety-related disorders like
entropy. Inputting such features into UMAP, as had depression113-116 and neurodegenerative diseases100-
been done in Leal et al., 202389, can help align our 109,117-118
, which all see (as mentioned previously with
UMAP manifold space embeddings with the temporal inputting relevant features for UMAP) significant
intricacies and informational richness of latent neural inter-subject variability, have direct connections to
dynamics during music listening and emotional emotion and cognition and report improvements in
experience, which are both temporally rich and will condition with music therapy. Through our outlined
indeed satisfy complexity of the various nonlinear and topological and population-level approach, music-brain
complex measures mentioned above, we posit. research can finally probe into music listening's latent
Furthermore, these features would extend future neural dynamics and nuances in subjects’ differed
UMAP embeddings’ relevance to clinical applications, emergent emotional experiences (as seen in subject 2
as many of the mentioned features have been applied in musical piece 2, Fig. 3c), ultimately tackling therapies
personalized biomedical health monitoring for seizures and interventions into the above different brain
and general stress90-93, as well as neurodegenerative conditions more deeply; our findings and more
diseases, i.e., Parkinson’s disease94-96, Alzheimer’s97-98 importantly, implications are groundbreaking: Whereas
and schizophrenia99. Simply by inputting these tailored the field of neuropathology has always treated inter-
features into UMAP, we can compare the music subject variability in music therapy, as well as other
listening-emotional experience latent neural dynamics treatments, as an open-ended question with no answer,
of healthy subjects and patients with such now we have the answer: subject differences in
neurodegenerative diseases, finally arriving at a proof- topologies during music listening/therapy reflects
of-concept in (computational) neuroscience, through individual conditions in their pathological states,
topology/latent neural dynamics, for why music helps allowing us to understand the dynamics behind these

12
differences, and ultimately, the diseases themselves. the repeated, stochastic/predictable processing of
This leads us into the next frontier of research: AI. positive emotional states37 and clear neurobiological
AI, machine learning and deep learning have started traces (the most commonly employed approaches to
gaining traction in all sorts of fields, not excluding understanding music and well-being in the past few
neuroscience; EEG data is a primary mode of AI decades, which are simplistic and make general
research into being able to classify/recognize emotions assumptions, without regard to inter-subject variability
on new data from previous data119-122; many different which we have shown is essential to understanding
forms of machine learning, deep learning and neural latent neural dynamics (Fig. 3c)), but rather, the well-
networks have been used for this emotion classification being benefits of music are attributed to the persistence
purpose123-131. Leading with AI for music listening- of continuous emotional experience, with a wide range
emotional experience research will allow us to form of emotional states140 (both sad and happy, and
tailored solutions to the mental health crisis we have everything in between) and the awareness of such
currently through personalized, convenient therapeutic persistent emotional experience44-47, during music
interventions in hospitals and related clinical settings listening. Since our results show promising traction
paired with music listening132-138. Specifically, in future with shared and differed neural dynamics across
work we can pry into the topologies of subjects’ latent patients, dependent on their shared and differed
neural dynamics with the most naturalistic snapshot of emotional experiences (e.g., subject 2 piece 2's greater
complex music listening and emotional experience: occurrence of emotions of contentment and thus a
subjects/patients simply listening to the music, with more discretely distributed manifold space with strong
their emotional experience predicted by a high- clustering) which can be predicted with deep learning
accuracy deep learning framework131 , exempt of (as mentioned above to allow a more naturalistic
having to physically report their emotions, which can listening setting), we now propose the following
interfere with the latent neural dynamics of pure music layered topological approach to continue to understand
listening and emotional experience, on a physical music listening-emotional experience latent neural
interface. We will also research different genres of dynamics for well-being: using topological data
music, but for the most part focus on classical music as analysis (TDA) techniques like persistent homology
it leads to fewer lyric/song-based biases in emotional and Betti numbers to unravel persistent features of the
experience responses and has a wide range of musical neural manifolds we construct, such as amounts of
styles to choose from. We anticipate focusing on music holes141-145, which should reflect persistence of
from both Johann Sebastian Bach and Claude Debussy emotional experience as our neural manifolds do.
in a future study, who are two very different Greater well-being, if measured as a persistence of a
composers139; the differences in composing styles, certain amount holes, being a significant topological
soundscapes and overall structure of Bach and feature of the neural manifold, can provide insights
Debussy’s respective musical pieces, we posit, will into the continuous cognitive and emotional processing
lead to noticeable differences in emotional experiences that happens in the brain during music listening,
and subsequently, latent neural dynamics of listeners. emergent in the form of complex emotional
These will be our next steps with future experimental experiences. Essentially, we have laid the groundwork
opportunities and computational resources for probing with a simplistic first-application of a manifold
into the differed emergent emotional experience of learning technique like UMAP to extract a
widely different subjects and widely different brain visualization of the latent neural dynamics; the next
conditions, as mentioned earlier with reference to steps are applying TDA steps to understand persistence
neurodegenerative diseases, depression, seizure of topological features and similarly, persistence of
condition or simply healthy subjects looking for that aspects of emotional experience, alongside applying
well-being boost in our modern mental health crisis. deep learning networks to predict emotion and allow
And finally, having introduced deep learning for for more naturalistic experimental settings (no
getting more naturalistic experimental settings for recording of emotion, just music listening). And to add
investigating music listening-emotional experience the cherry on top, we can also investigate extrinsic
latent neural dynamics, we now end off this study by geometries of our music listening-emotional
laying out the immediate next steps of applied experience latent neural dynamics manifolds,
topology, extending beyond merely DR and finding specifically, extrinsic manifold curvature 146, which can
manifolds/topologies as we did here. First off, from help lend us insights into how music listening-
findings in this study, we posit (and hope future emotional experience latent neural dynamics change
research will further explore) that the well-being alongside the moment-to-moment or phrase-to-phrase
benefits of music listening are not merely attributed to
13
changes of a musical piece, through the manifold 5. Mamun, M. A. & Ullah, I. COVID-19 suicides in
changing in curvature with those musical moments. Pakistan, dying off not COVID-19 fear but poverty? -
This is all a truly exciting enterprise, melding The forthcoming economic challenges for a developing
different aspects of topology and mathematics to help country. Brain Behav Immun 87, 163–166 (2020).
us understand how music listening promotes greater 6. Thakur, V. & Jain, A. COVID 2019-suicides: A
well-being, through topological (TDA) and extrinsic global psychological pandemic. Brain Behav
geometric, Riemannian (curvature) properties of neural Immun 88, 952–953 (2020).
manifolds changing alongside the music, paired with 7. Hao, F. et al. Do psychiatric patients experience
machine/deep learning prediction of emotion for a more psychiatric symptoms during COVID-19
most-naturalistic experimental setting without pandemic and lockdown? A case-control study with
requiring physical emotion reporting. This layered service and research implications for
approaches is glued by a sense of complexity theory in immunopsychiatry. Brain Behav Immun 87, 100–106
that the persistence of intrinsic (TDA) persistence and (2020).
changes in extrinsic curvature of our neural manifolds 8. Tan, W. et al. Is returning to work during the
during music listening will emerge in the form of COVID-19 pandemic stressful? A study on immediate
shared or distinct/differed complex emotional mental health status and psychoneuroimmunity
experiences, that listeners share in or contribute their prevention measures of Chinese workforce. Brain
own unique perspectives on over the course of phrase- Behav Immun 87, 84–92 (2020).
to-phrase and moment-to-moment beauties of the 9. Wang, C. et al. Immediate Psychological
musical piece at play. And like in the case of subject 2 Responses and Associated Factors during the Initial
whose musical piece 2 UMAP embedding was vastly Stage of the 2019 Coronavirus Disease (COVID-19)
different paired with different emotions, we can then Epidemic among the General Population in
further ask ourselves the question: What are we China. IJERPH 17, 1729 (2020).
missing? By starting to input complex/nonlinear 10. Wang, C. et al. A longitudinal study on the
features into UMAP (or other future DR techniques), mental health of general population during the
utilizing deep learning and implementing both TDA COVID-19 epidemic in China. Brain Behav Immun 87,
and extrinsic curvature metrics, we can start to 40–48 (2020).
understand more of these music listening-emotional 11. Gao, J. et al. Mental health problems and social
experience topologies, with the foundational findings media exposure during COVID-19 outbreak. PLoS
of this study. We will start to greater understand the One 15, e0231924 (2020).
complex physical phenomena of both physical and 12. Samadarshi, S. C. A., Sharma, S. & Bhatta, J. An
mental well-being, and with time, get that one step online survey of factors associated with self-perceived
closer to providing personalized solutions for people stress during the initial stage of the COVID-19
suffering from all sorts of neurodegenerative diseases, outbreak in Nepal. Ethiopian Journal of Health
depression or even everyday troubles, leading to a Development 34, (2020).
healthier and happier future, with greater well-being 13. Li, C., Liu, J., Lin, J. & Shang, H. COVID-19
for all. and risk of neurodegenerative disorders: A Mendelian
randomization study. Transl Psychiatry 12, 1–6 (2022).
References 14. Fu, Y.-W., Xu, H.-S. & Liu, S.-J. COVID-19
and neurodegenerative diseases. Eur Rev Med
1. Xiong, J. et al. Impact of COVID-19 pandemic on
Pharmacol Sci 26, 4535–4544 (2022).
mental health in the general population: A systematic
15. Hu, C., Chen, C. & Dong, X.-P. Impact of
review. J Affect Disord 277, 55–64 (2020).
COVID-19 Pandemic on Patients With
2. Kupcova, I., Danisovic, L., Klein, M. & Harsanyi,
Neurodegenerative Diseases. Front Aging Neurosci 13,
S. Effects of the COVID-19 pandemic on mental health,
664965 (2021).
anxiety, and depression. BMC Psychol 11, 108 (2023).
16. Zarifkar, P., Peinkhofer, C., Benros, M. E. &
3. Overhage, L. et al. Trends in Acute Care Use for
Kondziella, D. Frequency of Neurological Diseases
Mental Health Conditions Among Youth During the
After COVID-19, Influenza A/B and Bacterial
COVID-19 Pandemic. JAMA Psychiatry 80, 924–932
Pneumonia. Front Neurol 13, 904796 (2022).
(2023).
17. Foster, B., Pearson, S., Berends, A. &
4. McIntyre, R. S. & Lee, Y. Projected increases in
Mackinnon, C. The Expanding Scope, Inclusivity, and
suicide in Canada as a consequence of COVID-
Integration of Music in Healthcare: Recent
19. Psychiatry Res 290, 113104 (2020).
Developments, Research Illustration, and Future
Direction. Healthcare (Basel) 9, 99 (2021).

14
 18. Campbell, M. Music therapy methods gain 33. Yuste, R. From the neuron doctrine to neural
traction through local efforts. The Beacon networks. Nat Rev Neurosci 16, 487–497 (2015).
Today https://www.thebeacontoday.com/post/music- 34. Natraj, N., Silversmith, D. B., Chang, E. F. &
therapy-methods-gain-traction-through-local-efforts Ganguly, K. Compartmentalized dynamics within a
(2022). common multi-area mesoscale manifold represent a
19. Rossetti, A. et al. The Impact of Music Therapy repertoire of human hand movements. Neuron 110,
on Anxiety in Cancer Patients Undergoing Simulation 154-174.e12 (2022).
for Radiation Therapy. Int J Radiat Oncol Biol 35. Gallego, J. A. et al. Cortical population activity
Phys 99, 103–110 (2017). within a preserved neural manifold underlies multiple
20. Schiffman, R. The Healing Power of Music. The motor behaviors. Nat Commun 9, 4233 (2018).
New York Times (2021). 36. Fortunato, C. et al. Nonlinear manifolds underlie
21. Krout, R. E. Music listening to facilitate neural population activity during
relaxation and promote wellness: Integrated aspects of behaviour. bioRxiv 2023.07.18.549575 (2023).
our neurophysiological responses to music. The Arts in 37. Vuust, P., Heggli, O. A., Friston, K. J. &
Psychotherapy 34, 134–141 (2007). Kringelbach, M. L. Music in the brain. Nat Rev
22. Pratt, R. R. Art, dance, and music Neurosci 23, 287–305 (2022).
therapy. Physical Medicine and Rehabilitation Clinics 38. Pessoa, L. Neural dynamics of emotion and
of North America 15, 827–841 (2004). cognition: From trajectories to underlying neural
23. Moore, K. S. A systematic review on the neural geometry. Neural Netw 120, 158–166 (2019).
effects of music on emotion regulation: implications 39. Verma, G. K. & Tiwary, U. S. Affect
for music therapy practice. J Music Ther 50, 198–242 representation and recognition in 3D continuous
(2013). valence–arousal–dominance space. Multimed Tools
24. de Witte, M., Lindelauf, E., Moonen, X., Stams, Appl 76, 2159–2183 (2017).
G.-J. & van Hooren, S. Music Therapy Interventions 40. Gunes, H. & Schuller, B. Categorical and
for Stress Reduction in Adults With Mild Intellectual dimensional affect analysis in continuous input:
Disabilities: Perspectives From Clinical Current trends and future directions. Image and Vision
Practice. Frontiers in Psychology 11, (2020). Computing 31, 120–136 (2013).
25. Sharma, A., Moon, E., Kim, G. & Kang, S.-U. 41. Liu, J. et al. Eeg-based emotion classification
Perspectives of Circadian-Based Music Therapy for the using a deep neural network and sparse
Pathogenesis and Symptomatic Treatment of autoencoder. Frontiers in Systems Neuroscience 14,
Neurodegenerative Disorders. Frontiers in Integrative (2020).
Neuroscience 15, (2022). 42. Fuentes‐Sánchez, N., Pastor, R., Escrig, M. A.,
26. Raglio, A. et al. Effects of music and music Elipe‐Miravet, M. & Pastor, M. C. Emotion elicitation
therapy on mood in neurological patients. World during music listening: Subjective self‐reports, facial
Journal of Psychiatry 5, 68–78 (2015). expression, and autonomic
27. Fang, R., Ye, S., Huangfu, J. & Calimag, D. P. reactivity. Psychophysiology 58, e13884 (2021).
Music therapy is a potential intervention for cognition 43. Thammasan, N., Fukui, K. & Numao, M.
of Alzheimer’s Disease: a mini-review. Translational Application of deep belief networks in eeg-based
Neurodegeneration 6, 2 (2017). dynamic music-emotion recognition. in 2016
28. Sacks, O. The power of music. Brain 129, 2528– International Joint Conference on Neural Networks
2532 (2006). (IJCNN) 881–888 (2016).
29. Saxena, S. & Cunningham, J. P. Towards the 44. Barrett, L. F., Mesquita, B., Ochsner, K. N. &
neural population doctrine. Current Opinion in Gross, J. J. The Experience of Emotion. Annu Rev
Neurobiology 55, 103–111 (2019). Psychol 58, 373–403 (2007).
30. Chung, S. & Abbott, L. F. Neural population 45. UWA. Science of Emotion: The Basics of
geometry: An approach for understanding biological Emotional Psychology | UWA. UWA Online
and artificial neural networks. Current Opinion in https://online.uwa.edu/news/emotional-psychology/
Neurobiology 70, 137–144 (2021). (2019).
31. Langdon, C., Genkin, M. & Engel, T. A. A 46. Tye, M. The Experience of Emotion : an
unifying perspective on neural manifolds and circuits Intentionalist Theory. Revue internationale de
for cognition. Nat Rev Neurosci 24, 363–377 (2023). philosophie 243, 25–50 (2008).
32. Seung, H. S. & Lee, D. D. The Manifold Ways
of Perception. Science 290, 2268–2269 (2000).

15
 47. Feinberg, T. E. & Mallatt, J. Phenomenal 62. Eisma, J., Rawls, E., Long, S., Mach, R. &
Consciousness and Emergence: Eliminating the Lamm, C. Frontal midline theta differentiates separate
Explanatory Gap. Frontiers in Psychology 11, (2020). cognitive control strategies while still generalizing the
48. Cunningham, J. P. & Yu, B. M. Dimensionality need for cognitive control. Sci Rep 11, 14641 (2021).
reduction for large-scale neural recordings. Nat 63. Lapomarda, G., Valer, S., Job, R. & Grecucci, A.
Neurosci 17, 1500–1509 (2014). Built to last: Theta and delta changes in resting‐state
49. Cowie, r. Et al. ‘Feeltrace’: an instrument for EEG activity after regulating emotions. Brain and
recording perceived emotion in real time, in Behavior 12, e2597 (2022).
proceedings of the isca workshop on speech and 64. Lee, D. J., Kulubya, E., Goldin, P., Goodarzi, A.
emotion, 19–24 (2000). & Girgis, F. Review of the neural oscillations
50. Daly, I. et al. Electroencephalography reflects underlying meditation. Frontiers in Neuroscience 12,
the activity of sub-cortical brain regions during (2018).
approach-withdrawal behaviour while listening to 65. Webster, G. D. & Weir, C. G. Emotional
music. Sci Rep 9, 9415 (2019). responses to music: interactive effects of mode,
51. Jenkins, J. S. The Mozart effect. J R Soc Med texture, and tempo. Motiv Emot 29, 19–39 (2005).
94, 170–172 (2001). 66. Schaefer, H.-E. Music-evoked emotions—
52. Lahiri, N. & Duncan, J. S. The mozart effect: current studies. Front Neurosci 11, 600 (2017).
encore. Epilepsy & Behavior 11, 152–153 (2007). 67. Ball, P. Why dissonant music strikes the wrong
53. Koelsch, S., Cheung, V. K. M., Jentschke, S. & chord in the brain. Nature (2012).
Haynes, J.-D. Neocortical substrates of feelings evoked 68. McInnes, L., Healy, J., Saul, N. & Großberger,
with music in the ACC, insula, and somatosensory L. Umap: uniform manifold approximation and
cortex. Sci Rep 11, 10119 (2021). projection. JOSS 3, 861 (2018).
54. Chabin, T. et al. Cortical patterns of pleasurable 69. Milošević, D. et al. The application of Uniform
musical chills revealed by high-density eeg. Frontiers Manifold Approximation and Projection (UMAP) for
in Neuroscience 14, (2020). unconstrained ordination and classification of
55. Czepiel, A. et al. Synchrony in the periphery: biological indicators in aquatic ecology. Science of
inter-subject correlation of physiological responses The Total Environment 815, 152365 (2022).
during live music concerts. Sci Rep 11, 22457 (2021). 70. Yu, T.-T., Chen, C.-Y., Wu, T.-H. & Chang, Y.-
56. Scarratt, R. J., Heggli, O. A., Vuust, P. & C. Application of high-dimensional uniform manifold
Sadakata, M. Music that is used while studying and approximation and projection (UMAP) to cluster
music that is used for sleep share similar musical existing landfills on the basis of geographical and
features, genres and subgroups. Sci Rep 13, 4735 environmental features. Science of The Total
(2023). Environment 904, 167013 (2023).
57. Flinker, A., Doyle, W. K., Mehta, A. D., 71. Stolarek, I., Samelak-Czajka, A., Figlerowicz,
Devinsky, O. & Poeppel, D. Spectrotemporal M. & Jackowiak, P. Dimensionality reduction by
modulation provides a unifying framework for auditory UMAP for visualizing and aiding in classification of
cortical asymmetries. Nat Hum Behav 3, 393–405 imaging flow cytometry data. iScience 25, 105142
(2019). (2022).
58. Gan, L., Zhang, M., Jiang, J. & Duan, F. 72. Ujas, T. A., Obregon-Perko, V. & Stowe, A. M.
Electroencephalogram of happy emotional cognition A Guide on Analyzing Flow Cytometry Data Using
based on complex system of music and image visual Clustering Methods and Nonlinear Dimensionality
and auditory. Complexity 2020, e6631400 (2020). Reduction (tSNE or UMAP). Methods Mol Biol 2616,
59. Boasen, J., Takeshita, Y., Kuriki, S. & 231–249 (2023).
Yokosawa, K. Spectral-spatial differentiation of brain 73. Schneider, S., Lee, J. H. & Mathis, M. W.
activity during mental imagery of improvisational Learnable latent embeddings for joint behavioural and
music performance using meg. Frontiers in Human neural analysis. Nature 617, 360–368 (2023).
Neuroscience 12, (2018). 74. Leal, A. et al. Unsupervised EEG preictal
60. Sammler, D., Grigutsch, M., Fritz, T. & Koelsch, interval identification in patients with drug-resistant
S. Music and emotion: Electrophysiological correlates epilepsy. Sci Rep 13, 784 (2023).
of the processing of pleasant and unpleasant music. 75. Gao et al. A theory of multineuronal
Psychophysiology 44, 293–304 (2007). dimensionality, dynamics and measurement. bioRxiv
61. Fang, Y. Research on the mechanism of music 214262 (2017).
perception in the perspective of neuroscience.
Scientific and Social Research 5, 53–60 (2023).

16
 76. Yust, J. Harmonic qualities in Debussy’s “Les 90. Kesić, S. & Spasić, S. Z. Application of
sons et les parfums tournent dans l’air du soir”. Journal Higuchi’s fractal dimension from basic to clinical
of Mathematics and Music 11, 155–173 (2017). neurophysiology: A review. Computer Methods and
77. Buteau, C. & Anagnostopoulou, C. Programs in Biomedicine 133, 55–70 (2016).
Mathematical and computational modelling within a 91. Feucht, M. et al. Application of correlation
music analysis framework: motivic topologies as a case dimension and pointwise dimension for non-linear
study. Journal of Mathematics and Music 6, 1–16 topographical analysis of focal onset seizures. Med
(2012). Biol Eng Comput 37, 208–217 (1999).
78. Dabagia, M., Kording, K. P. & Dyer, E. L. 92. Benish, W. A. A Review of the Application of
Aligning latent representations of neural activity. Nat. Information Theory and Shannon Entropy to Clinical
Biomed. Eng 7, 337–343 (2023). Diagnostic Testing. Entropy 22, 97 (2020).
79. Kim, H. J. et al. Canonical Correlation Analysis 93. Conforte, A. J., Tuszynski, J. A., da Silva, F. A.
on Riemannian Manifolds and Its Applications. B. & Carels, N. Signaling Complexity Measured by
Comput Vis ECCV 8690, 251–267 (2014). Shannon Entropy and Its Application in Personalized
80. Cohen, U., Chung, S., Lee, D. D. & Medicine. Front Genet 10, 930 (2019).
Sompolinsky, H. Separability and geometry of object 94. Dutta, S., Ghosh, D. & Chatterjee, S.
manifolds in deep neural networks. Nat Commun 11, Multifractal detrended fluctuation analysis of human
746 (2020). gait diseases. Front Physiol 4, 274 (2013).
81. Bakagiannis, S. & Tarrant, M. Can music bring 95. Ihlen, E. Introduction to Multifractal Detrended
people together? Effects of shared musical preference Fluctuation Analysis in Matlab. Frontiers in
on intergroup bias in adolescence. Scand J Psychol 47, Physiology 3, (2012).
129–136 (2006). 96. Mehdizadeh, S. The largest Lyapunov exponent
82. Reybrouck, M., Vuust, P. & Brattico, E. Neural of gait in young and elderly individuals: A systematic
Correlates of Music Listening: Does the Music Matter? review. Gait Posture 60, 241–250 (2018).
Brain Sci 11, 1553 (2021). 97. Ando, M., Nobukawa, S., Kikuchi, M. &
83. Dingle, G. A. et al. How Do Music Activities Takahashi, T. Identification of Electroencephalogram
Affect Health and Well-Being? A Scoping Review of Signals in Alzheimer’s Disease by Multifractal and
Studies Examining Psychosocial Mechanisms. Front Multiscale Entropy Analysis. Front Neurosci 15,
Psychol 12, 713818 (2021). 667614 (2021).
84. McCrary, J. M., Altenmüller, E., Kretschmer, C. 98. Zorick, T., Landers, J., Leuchter, A. &
& Scholz, D. S. Association of Music Interventions Mandelkern, M. A. EEG multifractal analysis
With Health-Related Quality of Life: A Systematic correlates with cognitive testing scores and clinical
Review and Meta-analysis. JAMA Network Open 5, staging in mild cognitive impairment. J Clin Neurosci
e223236 (2022). 76, 195–200 (2020).
85. Reschke-Hernández, A. E., Gfeller, K., Oleson, 99. Kutepov, I. E. et al. EEG analysis in patients
J. & Tranel, D. Music Therapy Increases Social and with schizophrenia based on Lyapunov
Emotional Well-Being in Persons With Dementia: A exponents. Informatics in Medicine Unlocked 18,
Randomized Clinical Crossover Trial Comparing 100289 (2020).
Singing to Verbal Discussion. J Music Ther 60, 314– 100. Machado Sotomayor, M. J., Arufe-Giráldez, V.,
342 (2023). ; Ruíz-Rico, G. & Navarro-Patón, R. Music Therapy and
86. Croom, A. M. Music, Neuroscience, and the Parkinson’s Disease: A Systematic Review from 2015–
Psychology of Well-Being: A Précis. Front Psychol 2, 2020. Int J Environ Res Public Health 18, 11618
393 (2012). (2021).
87. Power Spectral Density - an overview | 101. Hsu, P., Ready, E. A. & Grahn, J. A. The
ScienceDirect Topics. effects of Parkinson’s disease, music training, and
https://www.sciencedirect.com/topics/computer- dance training on beat perception and production
science/power-spectral-density. abilities. PLoS One 17, e0264587 (2022).
88. Choong, W. Y. et al. Correlation Analysis of 102. Lee, H. & Ko, B. Effects of Music-Based
Emotional EEG In Alpha, Beta and Gamma Frequency Interventions on Motor and Non-Motor Symptoms in
Bands. J. Phys.: Conf. Ser. 1997, 012029 (2021). Patients with Parkinson’s Disease: A Systematic
89. Leal, A. et al. Unsupervised EEG preictal Review and Meta-Analysis. Int J Environ Res Public
interval identification in patients with drug-resistant Health 20, 1046 (2023).
epilepsy. Sci Rep 13, 784 (2023). 103. Gómez Gallego, M. & Gómez García, J. Music
therapy and Alzheimer’s disease: Cognitive,

17
psychological, and behavioural effects. Neurologia 32, Brain (eds. Cuddy, L. L., Belleville, S. & Moussard,
300–308 (2017). A.) 333–349 (Academic Press, 2020).
104. Matziorinis, A. M. & Koelsch, S. The promise 119. Chi, Y. M. et al. Dry and Noncontact EEG
of music therapy for Alzheimer’s disease: A review. Sensors for Mobile Brain–Computer Interfaces. IEEE
Ann N Y Acad Sci 1516, 11–17 (2022). Transactions on Neural Systems and Rehabilitation
105. Moreira, S. V., Justi, F. R. D. R. & Moreira, M. Engineering 20, 228–235 (2012).
Can musical intervention improve memory in 120. Huang, Y.-J., Wu, C.-Y., Wong, A. M.-K. &
Alzheimer’s patients? Evidence from a systematic Lin, B.-S. Novel Active Comb-Shaped Dry Electrode
review. Dement Neuropsychol 12, 133–142 (2018). for EEG Measurement in Hairy Site. IEEE
106. Leggieri, M. et al. Music Intervention Transactions on Biomedical Engineering 62, 256–263
Approaches for Alzheimer’s Disease: A Review of the (2015).
Literature. Front Neurosci 13, 132 (2019). 121. Li, X., Hu, B., Sun, S. & Cai, H. EEG-based
107. Mössler, K., Chen, X., Heldal, T. O. & Gold, C. mild depressive detection using feature selection
Music therapy for people with schizophrenia and methods and classifiers. Computer Methods and
schizophrenia-like disorders. Cochrane Database Syst Programs in Biomedicine 136, 151–161 (2016).
Rev CD004025 (2011). 122. Liu, Y.-J. et al. Real-Time Movie-Induced
108. Jia, R. et al. The effectiveness of adjunct music Discrete Emotion Recognition from EEG
therapy for patients with schizophrenia: A meta- Signals. IEEE Transactions on Affective Computing 9,
analysis. Psychiatry Res 293, 113464 (2020). 550–562 (2018).
109. Ulrich, G., Houtmans, T. & Gold, C. The 123. Craik, A., He, Y. & Contreras-Vidal, J. L. Deep
additional therapeutic effect of group music therapy for learning for electroencephalogram (EEG) classification
schizophrenic patients: a randomized study. Acta tasks: a review. J. Neural Eng. 16, 031001 (2019).
Psychiatr Scand 116, 362–370 (2007). 124. Bahari, F. & Janghorbani, A. EEG-based
110. Alexander, R. et al. The neuroscience of emotion recognition using Recurrence Plot analysis
positive emotions and affect: Implications for and K nearest neighbor classifier. in 2013 20th Iranian
cultivating happiness and wellbeing. Neuroscience & Conference on Biomedical Engineering (ICBME) 228–
Biobehavioral Reviews 121, 220–249 (2021). 233 (2013).
111. Chiao, J. Y. Current emotion research in 125. Orhan, U., Hekim, M. & Ozer, M. EEG signals
cultural neuroscience. Emot Rev 7, 280–293 (2015). classification using the K-means clustering and a
112. Swain, M. The inseparability of cognition and multilayer perceptron neural network model. Expert
emotion in second language learning. Language Systems with Applications 38, 13475–13481 (2011).
Teaching 46, 195–207 (2013). 126. Wang, Y. et al. EEG-Based Emotion
113. Singh, M. K. & Gotlib, I. H. The Neuroscience Recognition with Similarity Learning Network.
of Depression: Implications for Assessment and in 2019 41st Annual International Conference of the
Intervention. Behav Res Ther 62, 60–73 (2014). IEEE Engineering in Medicine and Biology Society
114. Rebecchini, L. Music, mental health, and (EMBC) 1209–1212 (2019).
immunity. Brain Behav Immun Health 18, 100374 127. Fu, Q. et al. Improving learning algorithm
(2021). performance for spiking neural networks. in 2017
115. Maletic, V. et al. Neurobiology of depression: IEEE 17th International Conference on
an integrated view of key findings. Int J Clin Pract 61, Communication Technology (ICCT) 1916–1919
2030–2040 (2007). (2017).
116. Choi, K. W., Kim, Y.-K. & Jeon, H. J. 128. Liu, Y.-J. et al. Real-Time Movie-Induced
Comorbid Anxiety and Depression: Clinical and Discrete Emotion Recognition from EEG
Conceptual Consideration and Transdiagnostic Signals. IEEE Transactions on Affective Computing 9,
Treatment. in Anxiety Disorders: Rethinking and 550–562 (2018).
Understanding Recent Discoveries (ed. Kim, Y.-K.) 129. Zheng, W. Multichannel EEG-Based Emotion
219–235 (Springer, 2020). Recognition via Group Sparse Canonical Correlation
117. Shankar, N. Music Therapy: A Necessary Analysis. IEEE Transactions on Cognitive and
Addition to the Treatment of Neurodegenerative Developmental Systems 9, 281–290 (2017).
Disorders? Georgetown Medical Review 6, (2022). 130. Zhang, Q., Chen, X., Zhan, Q., Yang, T. & Xia,
118. Hobeika, L. & Samson, S. Chapter 13 - Why do S. Respiration-based emotion recognition with deep
music-based interventions benefit persons with learning. Computers in Industry 92–93, 84–90 (2017).
neurodegenerative disease? in Music and the Aging 131. Liu, J. et al. Eeg-based emotion classification
using a deep neural network and sparse
18
autoencoder. Frontiers in Systems Neuroscience 14, Computer Vision and Pattern Recognition Workshops
(2020). (CVPRW) 610–619 (2023).
132. Sourina, O., Liu, Y., Wang, Q. & Nguyen, M.
K. EEG-Based Personalized Digital Experience. Methods
in Universal Access in Human-Computer Interaction. Pre-processing and epochs. Before further analyses,
Users Diversity (ed. Stephanidis, C.) 591–599 we segmented each of the data we selected into pieces
(Springer, 2011). and phrases, specifically musical phrases 1 and 2 and
133. Zhang, H. et al. Subject-independent EEG musical piece 1, adapting the commonly-used “epoch”
classification based on a hybrid neural for our purposes. We then filtered the data for 1-30 Hz
network. Frontiers in Neuroscience 17, (2023). to conduct independent component analysis (ICA) to
134. You, S. D. Classification of Relaxation and remove artifacts (on average, 11 artifiacts were
Concentration Mental States with removed per subject), and following ICA we filtered
EEG. Information 12, 187 (2021). the raw data with the ICA weights/removed artifacts
135. De Filippi, E. et al. Classification of Complex for the delta, theta and alpha bands with an IIR
Emotions Using EEG and Virtual Environment: Proof bandpass filter.
of Concept and Therapeutic Implication. Front Hum
Neurosci 15, 711279 (2021). Continuous 2-D valence-arousal representation. All
136. Saeed, S. M. U., Anwar, S. M., Khalid, H., valence and arousal values from the EEG files of each
Majid, M. & Bagci, U. EEG Based Classification of selected subject were used for our valence-arousal
Long-Term Stress Using Psychological representation. We normalized the valence and arousal
Labeling. Sensors (Basel) 20, 1886 (2020). values to a 0-1 range so that the values index into a
137. Asif, A., Majid, M. & Anwar, S. M. Human color map, which also operates on a 0-1 scale. By
stress classification using EEG signals in response to normalizing, we ensure that the full range of valence
music tracks. Comput Biol Med 107, 182–196 (2019). and arousal values maps to the full range of available
138. Rakesh, S. et al. Emotions Classification Using colors.
EEG in Health Care. in Computer Vision and Machine
Intelligence (eds. Tistarelli, M., Dubey, S. R., Singh, S. Spearman’s rank-correlation coefficient (ρS).
K. & Jiang, X.) 37–49 (Springer Nature, 2023). Spearman's rank-correlation coefficient, ρS , is a
139. Eliza. How Bach wrote the rules - and how nonparametric measure of rank correlation that
Debussy broke them. https://interlude.hk/why-bach- assesses the monotonic association between two sets of
wrote-the-rules-and-debussy-broke-them-all/ (2020). continuous variables. In our case, ρS between UMAP
140. Adler, J. M. & Hershfield, H. E. Mixed manifold point distances and valence/arousal distances
Emotional Experience Is Associated with and Precedes was computed after we created the UMAP embedding
Improvements in Psychological Well-Being. PLOS and valence-arousal arrays for each epoch,
ONE 7, e35633 (2012). respectively. By obtaining the Spearman correlation
141. Carlsson, G. Topological methods for data and p-value for each epoch of each subject, we were
modelling. Nat Rev Phys 2, 697–708 (2020). able to get a rough idea of how well the UMAP
142. Lum, P. Y. et al. Extracting insights from the embedding represented the valence-arousal
shape of complex data using topology. Sci Rep 3, 1236 relationships in the data; finding consistent, yet weak
(2013). correlations across epochs and subjects was an
143. Sizemore, A. E., Phillips-Cremins, J. E., Ghrist, indication of the relation between the UMAP
R. & Bassett, D. S. The importance of the whole: embeddings and the subjects’ emotional experiences.
Topological data analysis for the network
neuroscientist. Netw Neurosci 3, 656–673 (2019). UMAP parameters and embedding neural activity,
144. Wasserman, L. Topological Data In our study, we employed the Uniform Manifold
Analysis. Annu. Rev. Stat. Appl. 5, 501–532 (2018) Approximation and Projection (UMAP) algorithm to
145. Chaudhuri, R., Gerçek, B., Pandey, B., distill the high-dimensional data of neuronal activity
Peyrache, A. & Fiete, I. The intrinsic attractor into a comprehensible low-dimensional embedding
manifold and population dynamics of a canonical representation. UMAP operates under the premise that
cognitive circuit across waking and sleep. Nat data lie uniformly distributed on a manifold within a
Neurosci 22, 1512–1520 (2019). high-dimensional space. It begins by constructing a
146. Acosta, F., Sanborn, S., Duc, K. D., Madhav, weighted graph, where each node corresponds to a data
M. & Miolane, N. Quantifying Extrinsic Curvature in point, and edges reflect the probability of functional
Neural Manifolds. in 2023 IEEE/CVF Conference on

19
connectivity, based on the proximity of points within min_distance = 0.5, spread = 1.0, metric = ‘euclidean’)
this manifold. This proximity is determined by the with a 100 ms epoch for each feature, as there was less
strength of the connections in the weighted graph, timeseries of data to capture (22 seconds compared to
which in turn is indicative of the similarity in the piece 2’s 2 minutes and 7 seconds); our parameters for
activity patterns of the neurons represented by these musical piece 2 were (n_neighbors = 10, min_distance
points37. Then, it projects this graph into a lower- = 0.5, spread = 1.0, metric = ‘euclidean’) with a 500
dimensional space, aiming to maintain the manifold's ms epoch for feature, as there was more timeseries to
intrinsic geometric properties, or “qualitative qualities” capture. We computed features of power spectral
as we say (for an emphasis of qualitative thinking in density (PSD) for the theta and alpha bands,
our topological approach, as opposed to quantitative respectively, and the Pearson correlation coefficient
reasoning which conventional neuroscience typically (PCC) between subsequent pairs of channels.
leans towards). This projection is fine-tuned through a
cost function optimization—specifically, minimizing Data availability
cross-entropy to ensure the low-dimensional The dataset analysed during the current study is
representation most closely retains the original data's available in the Openneuro dataset “A dataset
structure holes37. recording joint EEG-fMRI during affective music
Our UMAP parameters were selected to balance the listening”:
preservation of local details and global structure. The
number of nearest neighbors parameter, n_neighbors,
was set to capture the nonlinear dynamics of EEG data
Code availability
All pre-processing and analyses in this study were
during music-listening and emotional experience; the
conducted on a local VS Code Anaconda distribution
minimum distance parameter, min_dist, was adjusted
environment with the aid of ChatGPT’s coding
to obtain visible clustering without overcrowding; and
capabilities; all code is immediately available at the
the number of components parameter, n_components,
author’s GitHub repository for this project:
was chosen for a 2-D embedding. Our parameters for
https://github.com/JosephDarcourt/Music-Topology
musical phrases 1 and 2 were (n_neighbors = 5,

20
21