The Comparison of Various Correlation Network Models in Studying
Mobility Data for the Analysis of Depression Episodes
Rama Krishna Thelagathoti
a
and Hesham H. Ali
b
College of Information Science and Technology, University of Nebraska Omaha, Omaha, NE 68182, U.S.A.
Keywords:
Depression, Mobility, Population Analysis, Correlation Network.
Abstract:
Depression is a serious mental health disorder affecting millions of people around the world. Traditional
diagnostic approaches are subjective including self-reporting feedback from patients and observational eval-
uation by a trained physician. However, altered motor activity is the central feature for depressive disorder.
Moreover, recent studies show that the analysis of motor activity is the best predictor in characterizing psycho-
logical disorders including depression. With the advent of wearable devices, an individual’s motor activity can
be monitored naturally using body worn sensors and feasible to distinguish depressed persons from healthy
individuals. In this manuscript, we hypothesis to apply a methodology that takes advantage of motor activity
recorded from wearable devices and process mobility patterns for a given group of subjects. Besides, em-
ployed a population analysis approach using correlation networks that evaluates mobility parameters of the
population and identify subgroups that exhibit similar motor complexity. We have analyzed the mobility data
of the given group by extracting three different sets of features using hour-wise, day-wise, and hybrid mobility
data. Also, a comparison study of three models is presented by constructing a correlation graph and finding
a cluster of individuals exhibiting similar mobility patterns. We found that mobility data using hour-wise
features provides the best results compared to the other two models.
1 INTRODUCTION
According to World Health Organization (WHO), ap-
proximately 280 million individuals suffer from de-
pression around the world which is equivalent to 3.8%
of the total world population (The World Health Or-
ganization(WHO), 2021). Moreover, depression may
impact any person regardless of their age, race, and
socio-economic background. However, it is likely
to affect adults than children. The onset of depres-
sion may not trigger by normal mood fluctuations
or temporary emotional disturbance, rather when the
sadness becomes recurrent with intense severity that
leads to a major depressive disorder (Abuse, 2018).
Depression is a serious mental health condition that
may cause frequent mood swings which result in a
deprived quality of life. Furthermore, recent studies
show that there is a surge in suicides in depressed
patients due to feelings of loneliness (Curtin et al.,
2016). Depression often may influence the work-life
balance and cause poor performance in studies. It is
due to the symptomatic nature of illness which causes
a
https://orcid.org/0000-0002-4986-5027
b
https://orcid.org/0000-0002-8016-6144
a gloomy mind, lack of pleasure in doing routine ac-
tivities, feeling worthlessness, and hopelessness (The
National Institute of Mental Health (NIMH), 2021).
It is known that there is no precise pathology test
such as a blood routine test to accurately diagnose
depression, yet most of the existing clinical progno-
sis is largely dependent on visual observation. Nu-
merous subjective diagnostic scales were proposed
to measure the severity of the disease. For exam-
ple, the Center for Epidemiologic Studies Depres-
sion Scale (CES-D) is a self-reporting method that
measures the severity on a 4-point scale (Radloff,
1977). Similarly, the Montgomery-Asberg Depres-
sion Rating Scale (MADRS) measures the serious-
ness of the disorder on a 7-point scale which is ex-
clusively designed for adults over 18 years of age
(Montgomery and
˚
Asberg, 1979). The drawback of
these approaches is that these methods merely de-
pend on human perception and comprehension skills.
Therefore, it is important to develop a sophisticated
methodology that is observer independent.
Although the main cause of depression is an ab-
normality in neurological functioning, altered motor
activity is one of the common symptoms that ap-
200
Thelagathoti, R. and Ali, H.
The Comparison of Various Correlation Network Models in Studying Mobility Data for the Analysis of Depression Episodes.
DOI: 10.5220/0010844500003123
In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 4: BIOSIGNALS, pages 200-207
ISBN: 978-989-758-552-4; ISSN: 2184-4305
Copyright
c
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
pear in patients suffering from depression. Besides,
previous studies demonstrate that analysis of motor
skills is an important aid in classifying depressed pa-
tients from healthy counterparts (Sobin and Sackeim,
1997). Moreover, the depressed group compose lower
body reaction time and decreased body movements
than healthy persons. This opens a door for new pos-
sibilities to categorize depression by utilizing their
mobility data. Altered or lessened motor activity al-
lows to distinguish depressed patients from healthy
subjects. Proliferation in sensing technologies cre-
ated tiny wearable devices to record motor behavior
unobtrusively in a natural setting without disturbing
the daily activities. Wearable devices are proved to
be efficient, affordable, unobtrusive, and more conve-
nient for they can even fit a newborn child and collect
the data for several days (Heinze et al., 2010).
For our tests, we have chosen the ’Depresjon’
dataset downloaded from the public database (Garcia-
Ceja et al., 2018). It consists of 55 subjects catego-
rized into two groups: the first group has 23 patients
diagnosed with either unipolar or bipolar depression
(the condition group) and the second group contains
32 healthy control subjects (the control group). The
main objectives of this study are as follows
1. Extracting three different categories of features
that represent motor activity segmented by the
hour, day, and combination of an hour as well as
day.
2. Employing the population analysis-based correla-
tion network approach to construct a graph where
the group of persons with similar mobility profiles
are strongly connected in the resultant graph.
3. Applying an appropriate clustering algorithm to
obtain potential subgroups in which each sub-
group represents a set of individuals exhibiting
similar motor complexity.
4. Comparison study of results obtained from three
different categories of features namely hour-wise,
day-wise, and hybrid.
The rest of the paper is organized as follows. Sec-
tion 2 covers the previous studies conducted on the
dataset. Section 3 describes a brief description of the
dataset, feature extraction, and correlation graph con-
struction. Whereas experimental results are shown in
section 4 and post hoc analysis of the obtained results
is elaborated in section 5.
2 RELATED WORK
In the past, several investigators have performed dif-
ferent experiments with the dataset. The dataset
was created by Garcia-Ceja et al. (Garcia-Ceja
et al., 2018) and published baseline performance re-
sults. They tested with different machine learning
algorithms but finally obtained higher accuracy of
73% with Linear Support Vector Machine (SVM).
In another research carried by Rodr
´
ıguez-Ruiz et al.
(Rodr
´
ıguez-Ruiz et al., 2020) processed motor data
and divided it into three sets as day, night, and full-
day activity data. The fundamental objective of this
work is to compare the motor activity patterns across
three different times of a day and draw profound in-
sights. They concluded that the features used to build
the nighttime motor data produced promising results
compared to the other two features. They obtained the
highest sensitivity and specificity of 99.4% and 99.9%
respectively.
On the other hand, Zanella-Calzada et al.
(Zanella-Calzada et al., 2019) extracted hour-wise
features by segmenting the overall motor activity into
the one-hour interval, trained the model using a Ran-
dom Forest classifier. Their model achieved 87% ac-
curacy while the sensitivity was 87% and specificity
was 92% . Similarly, Galvan-Tejada et al. (Galv
´
an-
Tejada et al., 2019) mined 38 statistical features be-
longing to the time and frequency domain. They have
employed a genetic algorithm-based feature selection
approach to identify the best features. They also used
Random Forest to predict between healthy and de-
pressed and obtained a sensitivity of 68% and a speci-
ficity of 61% .
Most of the researchers applied machine learning
techniques such as Random Forest and support vector
machines. Furthermore, they analyzed the data by us-
ing supervised machine learning methods by adding
a class label manually for each subject (0/YES for
condition group and 1/NO for control group or vice
versa). The novelty of our approach is that we do not
include known labels in the study, but we identify the
group of subjects by utilizing their mobility. In such
groups, condition subjects are gathered into a single
cluster and control subjects into another cluster.
3 MATERIALS AND METHODS
3.1 The Pipeline
Fig. 1 depicts the processing pipeline for correlation
graph analysis for mobility data acquired from De-
pressed patients. In the first step, a dataset is acquired
from the public repository. In the preprocessing step,
data is cleaned, normalized, and eliminated outliers.
In the third step, three different types of features are
extracted namely hour-wise (Model M1), day-wise
The Comparison of Various Correlation Network Models in Studying Mobility Data for the Analysis of Depression Episodes
201
Figure 1: The pipeline for correlation network model.
(Model M2), and hybrid (Model M3). The models
M1, M2, and M3 represent the average motor activity
data segmented by the hour, day, and the combina-
tion of an hour as well as day respectively. Then, a
pair-wise correlation is applied for each of the mod-
els to construct a correlation graph. Then, strongly
connected clusters are detected from the correlation
graph. Finally, resultant graphs are analyzed and dis-
cussed
3.2 Dataset Description
In this study, we have used the ’Depresjon’ dataset
(Garcia-Ceja et al., 2018). It is a public dataset con-
sisting of motor activity collected from 55 partici-
pants including 23 persons belonging to the condi-
tion group and 32 subjects belonging to the control
group. The persons in the condition group were di-
agnosed with either unipolar or bipolar disorder and
they are under antidepressant medications. Whereas
the 32 participants in the control group are healthy
individuals. In this document, motor activity data
and mobility data are used interchangeably through-
out this document. Their motor activity was recorded
using a body-worn wearable sensor embedded in an
Actigraph watch (Name: Actiwatch, Manufacturer:
Cambridge Neurotechnology Ltd, England, Model
AW4). For the comfort of all participants, the acti-
graph watch was worn on the right wrist and their mo-
bility data was continuously monitored in the natural
environment. None of the participants were called to
a pathology lab or followed any specific instructions.
The actigraph measures the activity with a piezoelec-
tric accelerometer that is designed to record the in-
tensity, quantity, and duration of movement in all di-
rections. The Motion data was captured with a sam-
pling frequency of 32Hz and movements over 0.05g
for every minute in the form of activity count. The
actigraph records the motor activity in the form of an
activity count. The higher activity count resembles
the higher intensity in the motor activity.
The captured mobility data contains activity count
along with its timestamp. Besides, each participant’s
mobility data was stored in a separate data file, and
they can be identified with a unique contributor id.
Moreover, all of them were participated and provided
their data for a different number of days. However, on
average every person has 12 days of motor activity.
In addition to the data file, individuals’ demographic
characteristics are provided in a separate file (scores
file). This file contains the important information of
each person such as person unique id, days (number
of days of data monitored), gender (1:female, 2:male),
age (age range), afftype (1: bipolar II, 2: unipolar
depressive, 3: bipolar I), melanch (1: melancholia,
2: no melancholia), In addition to this, every subject
in the condition group was assessed by MADRS ob-
servational scale (Montgomery and
˚
Asberg, 1979) at
the start of the data collection and also at the end of
the data collection. The MADRS scores are available
under MADRS1 and MADRS2 columns respectively.
Further, a statistical summary of all 55 participants
and their demographic details are described in Table
1.
3.3 Preprocessing
The first step in preprocessing phase is combining the
individual raw sensor data into a single dataset and
preparing for further processing. The motor activity
data was not measured for the same duration. How-
ever, on average 12 days of motor data is available
for all the users. Furthermore, the number of days
the data is available is not consistent between the user
sensor data file and the scores file. Therefore, we have
Table 1: Demographic characteristics.
Condition group Control group
Statistic Mean SD Mean SD
Days 12.6 2.3 12.6 2.7
Age 42.8 11 38.2 13
MADRS 1 22.7 4.8
MADRS 2 20 4.7
Statistic Total % Total %
Gender (Male) 13 57 12 38
Depression
(Bipolar)
8 34
Hospitalized
(Inpatient)
5 22
BIOSIGNALS 2022 - 15th International Conference on Bio-inspired Systems and Signal Processing
202
Table 2: Model-wise features.
Model
name
Feature
name
Features
count
Feature description
M1:
Hour-
wise
model
m0-
m23
24
The Average motor ac-
tivity measured for every
hour for 0-23 hours
sd0-
sd23
24
The standard deviation of
motor activity measured
for every hour for 0-23
hours
M2:
Day-
wise
model
dm1-
dm19
24
The Average motor activ-
ity measured for each day
for 1-19 days
dsd1-
dsd19
24
The standard deviation of
motor activity measured
for each day for 1-19 days
M3:
Hy-
brid
model
m0-
m23
24
The Average motor ac-
tivity measured for every
hour for 0-23 hours
sd0-
sd23
24
The standard deviation of
motor activity measured
for every hour for 0-23
hours
dm1-
dm19
24
The Average motor activ-
ity measured for each day
for 1-19 days
dsd1-
dsd19
24
The standard deviation of
motor activity measured
for each day for 1-19 days
id 1
A unique id to represent
each subject
taken the number of days mentioned in the scores file
as the ground truth and deleted the additional days of
motor data present in the sensor data file for each par-
ticipant. In the next step, the activity signal is nor-
malized, and removed outliers. The activity signal
data is normalized between 0 and 1 using the Z-score
standardization technique. Since the condition and
control groups belong to two different entities, both
groups’ sensors data is normalized separately. In the
next step, outliers are eliminated by utilizing the in-
terquartile range (IQR) property. In this context, a
data point is considered an outlier if it is below the
first quartile or above the third quartile. In this pro-
cess, outliers are not removed rather they are replaced
with either the first quartile or the third quartile de-
pending on whether the data point is above the third
quartile or below the first quartile respectively. The
resultant dataset is normalized and free from outliers.
3.4 Feature Extraction
Each participant has shared their mobility data for a
certain number of days. But all of them were not col-
lected their motor data for the same number of days.
For example, participant 8 in the condition group has
provided mobility data for 5 days, while person 2 has
20 days of motor data. In this manuscript, we propose
to utilize three different types of features: hour-wise
features (Model M1) that represent hourly motor ac-
tivity in a 24-hour cycle, day-wise features (Model
M2) that signifies overall day motor activity, hybrid
features (Model M3) that combine both hour-wise and
day-wise features. The detailed list of features is elab-
orated in Table 2.
In the hour-wise model (M1), motor activity is
segmented by an hourly pattern. Although each par-
ticipant generated motor activity for a variable num-
ber of days, the total activity of a person for all days
is aggregated before extraction of the features. Then
each person’s motor data is divided by an hour inter-
val. Further, the mean and the standard deviation (SD)
are computed for every hour of aggregated data. As
a result, 24 features are generated from mean and an-
other 24 features are generated from SD. In the day-
wise model (M2), a person’s motor data of a day is
aggregated, and this process is repeated for all days.
Then mean and SD is calculated for each day. From
the dataset, it is known that each participant’s motor
activity is collected for a variable number of days, yet
a user has not more than 19 days of activity data. So,
19 features of as day-wise activity means, and 19 fea-
tures of day-wise SD are processed. This process pro-
duces 38 features for each person. Since every par-
ticipant does not possess 19 days of sensor data, the
remaining days where the data is not present are filled
with 0. To eliminate the bias of the number of days
between two persons, the minimum number of days
is considered during modeling. In the hybrid model
(M3), 48 features from the hour-wise model and 38
features from the day-wise model are combined. Ef-
fectively M3 model generates 86 features.
3.5 Construction of Correlation
Network Model
The objective of building a correlation graph is to
understand the interrelationships among the partici-
pants concerning their mobility parameters. In previ-
ous experiments (Garcia-Ceja et al., 2018) (Zanella-
Calzada et al., 2019), researchers have employed ma-
chine learning approaches and classified depressed
patients from the healthy control group. Neverthe-
less, all these studies have utilized a known class la-
bel such as 0/NO for healthy control subjects, 1/YES
for a depressed patient, then try to classify the sub-
jects and measure the accuracy of the prediction al-
gorithm. The inherent downside of this approach is
that the learning algorithm works only if the known
The Comparison of Various Correlation Network Models in Studying Mobility Data for the Analysis of Depression Episodes
203
label is present in the dataset. Besides, these method-
ologies are label-driven. In this manuscript, we intro-
duce a data-driven approach by employing a popula-
tion analysis approach using correlation graphs. This
approach does not require a label to be present in
the dataset rather it analyzes the mobility parameters
of the given group and identifies the subgroups that
demonstrate similar mobility patterns. Our hypothe-
sis is developed on the fact that subgroups in the given
group compose similar motor activity which makes
them distinguishable from other groups. This is fur-
ther exemplified from the motor data of 55 subjects
where the overall mean activity of the condition group
is 284 while the condition group has 187.
The first step in the graph creation is to establish
the relationship between each pair of subjects with re-
spect to their motor activity data. Once the relation-
ships are identified, their interconnections are repre-
sented using a graph. A graph G = (V, E) is an ab-
stract mathematical representation of any system that
depicts the relationships between the objects. In such
a graph, nodes or vertices (V) denotes the elements
of the system, and edges (E) represent the intercon-
nection between the elements (Dongen, 2000). In this
study, all 55 participants are denoted as nodes, and
their relationship regarding their motor activity is rep-
resented as an edge. It implies that two participants
are connected by an edge if they possess a similar mo-
tor activity profile.
The degree of similarity between each pair of sub-
jects is measured using the Pearson pair-wise correla-
tion coefficient (ρ). The Pearson pair-wise correlation
coefficient measures the linear dependence between
a pair of objects. Usually, the value ranges between
-1 and +1 where -1 indicates a negative correlation
and +1 signifies a strong positive correlation. To con-
struct the correlation graph, the ρ value is computed
between each pair of users by utilizing their motor
activity data. This operation outputs a correlation ma-
trix with pair-wise correlation coefficient values. The
ρ value between a pair of users signifies the degree
of similarity with regards to their motor activity. The
higher the ρ value the stronger the relationship be-
tween the pair of users. To create the graph from the
correlation matrix, strongly correlated pairs are iden-
tified by using the significance matrix. A significance
matrix is obtained by setting a predefined threshold k
using equation 1.
significance matrix(i, j) = 1, i f (ρ(Pi, P j)) k
= 0, i f (ρ(Pi, P j))<k
(1)
A predefined threshold k indicates the correlation
at which a pair in the matrix is significant. Intuitively,
when 55 participants are represented by a significance
matrix then two persons (Pi, Pj) are said to be asso-
ciated if their correlation constant exceeds or is equal
to k. Therefore, Pi and Pj are connected by an edge in
the resultant correlation graph. Since the significance
matrix will have either 0 or 1, it is equivalent to the ad-
jacency matrix. As the last step in graph creation, an
adjacency matrix is translated to a correlation graph.
3.6 Clustering
Even though the correlation graph is built, it is nec-
essary to find the potential clusters in the resultant
graph. Often, the terms Clustering and Community
discovery are used interchangeably by the scientific
community. In biological networks, clustering or
community discovery is a method of classifying the
elements into groups (clusters) wherein members of
each group are similar by means of certain character-
istics (Girvan and Newman, 2002) (Ali et al., 2019).
In the current study, clusters are identified accord-
ing to the motor complexity of all the subjects un-
der the study. So, all the persons in a cluster are ex-
pected to have similar mobility patterns. A cluster
in the correlation graph signifies a group of subjects
that are strongly interconnected through mobility pat-
terns. Also, the discovered clusters naturally hold two
principles: Homogeneity and Separation. Homogene-
ity alludes to the similarity among persons within the
same cluster while separation indicates persons in dif-
ferent clusters exhibit different characteristics.
To uncover the hidden communities in the cor-
relation graph, MCL (Markov Clustering) technique
is applied. The MCL algorithm is a popular unsu-
pervised clustering algorithm that is well suitable for
extracting communities in biological networks (Don-
gen, 2000). The MCL algorithm works by a ran-
dom walk property of a graph where all nodes are
randomly visited to find the strongly connected com-
ponents in the graph. A good clustering algorithm
typically produces high-quality clusters with distinct
non-overlapping boundaries. Yet, achieving perfect
separation is practically not possible.
4 RESULTS
This study includes 55 participants with 23 from the
condition group (patients suffering from depression)
and 32 from the control group (healthy counterparts).
The final dataset used for correlation analysis has 55
observations where each observation corresponds to a
person. However, the number of feature variables is
different for each model. The hour-wise model (M1)
has 48 features, the day-wise model (M2) consists of
BIOSIGNALS 2022 - 15th International Conference on Bio-inspired Systems and Signal Processing
204
(a) M1:Correlation graph.
(b) M1:Clusters.
(c) M2:Correlation graph. (d) M2:Clusters.
(e) M3:Correlation graph. (f) M3:Clusters.
Figure 2: Correlation graphs and discovered clusters.
The Comparison of Various Correlation Network Models in Studying Mobility Data for the Analysis of Depression Episodes
205
38 features, and the hybrid model (M3) has 86 fea-
tures. The Pearson correlation coefficient is applied to
three models then M1 outputs 55x48 matrix, M2 out-
puts 55x38 matrix, and M3 yields 55x86 matrix. In
the next step, a predefined threshold of 0.7, 0.6, and
0.55 are set to the models M1, M2, and M3 respec-
tively, to get the significance matrix. A correlation
graph is generated from the three significance matri-
ces. The obtained correlation graphs of M1, M2, and
M3 models are shown in Figure 2 (a, c, e). To rec-
ognize each person a unique id is used where con-
trol subjects are numbered from 1 to 23 and condition
groups are numbered from 24-55. Furthermore, con-
trol group participants are colored in green whereas
condition group subjects are colored in red. Each ver-
tex in the resultant correlation network represents an
individual while the edge between two vertices signi-
fies the degree of similarity in terms of their move-
ment pattern.
Unraveling hidden clusters in a correlation graph
is a crucial step at this stage. MCL algorithm is
utilized to discover the potential clusters from three
graphs as depicted in Figure 2 (b, d, f). In this
graph, nodes with similar colors signify that they be-
long to the same community. we can comprehend
from the graph that controls and condition subjects
are fairly separated into two dense communities (con-
dition group in red color and control group is in green
color). Analyzing these communities provides nu-
merous insights into the connections between the in-
dividuals. Section 5 further elaborates on commonal-
ities between the persons in the same community.
5 DISCUSSION
In this section, a post hoc analysis is carried out on
the results obtained from the three models. The input
dataset consisting of mobility data collected from 55
subjects also provides classification labels that corre-
spond to the diagnosis of the person. It is known from
the dataset that participants numbered from 1 to 23
belong to the condition group and they are diagnosed
with either unipolar or bipolar depression, whereas
subjects numbered from 24 to 55 belong to the healthy
control group. Previous studies that employed ma-
chine learning techniques had obtained higher ac-
curacy in terms of predicting the persons with and
without disorder (Garcia-Ceja et al., 2018) (Zanella-
Calzada et al., 2019) (Rodr
´
ıguez-Ruiz et al., 2020).
However, our hypothesis is not established based on
known labels rather we built the network by taking
advantage of the motor activity data itself. Therefore,
subgroups extracted from the correlation model are
more intuitive in terms of their movement patterns.
The main objective behind creating three models
is to understand the granularity of the mobility that
can best describe the overall movement patterns of
the subjects under study. From Figure 2, results ob-
tained from hour-wise mobility data are more promis-
ing than the other two models built on day-wise and
hybrid mobility data. Comparing three models shown
in Figure 2, M1 and M3 produced 6 clusters and M2
produced 4 clusters. However, all three models have
two dense clusters, and they mostly differ with respect
to the number of singleton or dual node clusters that
are not connected to the network. Model M1 has 6
clusters in which persons P2, P14, P18, and P44 are
isolated from the group. The phenomenon of isola-
tion highlights the peculiarity of these persons. They
are isolated because their mobility is not comparable
with any other person in the group. Nonetheless, from
the available information in the dataset, it is not possi-
ble to determine the exact reason behind their separa-
tion. Hence, we believe that having additional infor-
mation such as clinical parameters might be helpful
in further analysis. Additionally, the hourly features
employed for the M1 model separated condition and
control groups into two well-separated subgroups ac-
cording to their mobility but without using known la-
bels. Even though the M3 model divided condition
and control groups, P15 who is supposed to belong
to the condition group is clustered into the control
group. Similarly, in the M2 communities’ graph, P15
and P18 are classified as condition groups but they are
strongly correlated to control subjects than condition
subjects. By utilizing these rich insights, it is plau-
sible to comprehend the severity of the disorder pro-
vided if there is additional clinical information such
as medical history.
Another aspect of constructing three different
models is to realize the best method that can distin-
guish the subgroups according to the degree of mo-
bility. The creators of the dataset did not mention the
actual setting of the subjects under study. If all the
subjects are residing in the same community and have
the same daily routine, then day-wise segmentation
of the mobility data is helpful than the hourly seg-
mentation. Conversely, if the participants are living
in different communities with diverse daily routines,
then hourly features might produce better results than
day-wise features.
6 CONCLUSION
Mobility is considered one of the important influen-
tial factors that determine the overall health of an in-
BIOSIGNALS 2022 - 15th International Conference on Bio-inspired Systems and Signal Processing
206
dividual. However, certain medical conditions such
as depression can impact the mobility pattern. Conse-
quently, the affected individual’s movements are sig-
nificantly altered compared to their healthy counter-
parts. However, the degradation in mobility can be
used as a vital parameter in characterizing the disor-
der. In the past, physicians assessed the depression
by an observation followed by self-reported feedback
from the patients. Yet, with the latest innovations in
wearable devices, it is possible to diagnose the ill-
ness by collecting mobility data from depressed pa-
tients using wearable sensors. In this study, we pro-
posed and built a correlation network model by utiliz-
ing the movement data collected from the group con-
sisting of depressed as well as healthy subjects. Ear-
lier studies predominantly focused on prediction of
the depression by incorporating known labels. How-
ever, our hypothesis is built on the concept of pop-
ulation analysis and correlation network by utilizing
the mobility data. We treated all the subjects belong-
ing to one group then explored similarities and differ-
ences between each pair of subjects by utilizing their
movement data. Then we constructed a correlation
network model that has the potential to discover the
subgroups of those who are suffering from depression
and healthy subjects. We have extracted three differ-
ent granularity of features and we found that hour-
wise features are the best set of feature parameters
that can fairly identify the subgroups.
REFERENCES
Abuse, S. (2018). Mental health services administra-
tion.(2017). key substance use and mental health in-
dicators in the united states: Results from the 2016
national survey on drug use and health (hhs publica-
tion no. sma 17-5044, nsduh series h-52). rockville,
md: Center for behavioral health statistics and qual-
ity. Substance Abuse and Mental Health Services
Administration. Retrieved from https://www. samhsa.
gov/data.
Ali, N., Neagu, D., and Trundle, P. (2019). Evaluation of
k-nearest neighbour classifier performance for hetero-
geneous data sets. SN Applied Sciences, 1(12):1–15.
Curtin, S. C., Warner, M., and Hedegaard, H. (2016). In-
crease in suicide in the United States, 1999-2014.
Number 2016. US Department of Health and Human
Services, Centers for Disease Control and . . . .
Dongen, S. (2000). Performance criteria for graph cluster-
ing and Markov cluster experiments. CWI (Centre for
Mathematics and Computer Science).
Galv
´
an-Tejada, C. E., Zanella-Calzada, L. A., Gamboa-
Rosales, H., Galv
´
an-Tejada, J. I., Ch
´
avez-Lamas,
N. M., Gracia-Cort
´
es, M., Magallanes-Quintanar, R.,
Celaya-Padilla, J. M., et al. (2019). Depression
episodes detection in unipolar and bipolar patients: A
methodology with feature extraction and feature se-
lection with genetic algorithms using activity motion
signal as information source. Mobile Information Sys-
tems, 2019.
Garcia-Ceja, E., Riegler, M., Jakobsen, P., Tørresen, J.,
Nordgreen, T., Oedegaard, K. J., and Fasmer, O. B.
(2018). Depresjon: a motor activity database of de-
pression episodes in unipolar and bipolar patients. In
Proceedings of the 9th ACM multimedia systems con-
ference, pages 472–477.
Girvan, M. and Newman, M. E. (2002). Community struc-
ture in social and biological networks. Proceedings of
the national academy of sciences, 99(12):7821–7826.
Heinze, F., Hesels, K., Breitbach-Faller, N., Schmitz-Rode,
T., and Disselhorst-Klug, C. (2010). Movement anal-
ysis by accelerometry of newborns and infants for the
early detection of movement disorders due to infantile
cerebral palsy. Medical & biological engineering &
computing, 48(8):765–772.
Montgomery, S. A. and
˚
Asberg, M. (1979). A new depres-
sion scale designed to be sensitive to change. The
British journal of psychiatry, 134(4):382–389.
Radloff, L. S. (1977). The ces-d scale: A self-report de-
pression scale for research in the general population.
Applied psychological measurement, 1(3):385–401.
Rodr
´
ıguez-Ruiz, J. G., Galv
´
an-Tejada, C. E., Zanella-
Calzada, L. A., Celaya-Padilla, J. M., Galv
´
an-
Tejada, J. I., Gamboa-Rosales, H., Luna-Garc
´
ıa, H.,
Magallanes-Quintanar, R., and Soto-Murillo, M. A.
(2020). Comparison of night, day and 24 h motor ac-
tivity data for the classification of depressive episodes.
Diagnostics, 10(3):162.
Sobin, C. and Sackeim, H. A. (1997). Psychomotor symp-
toms of depression. American Journal of Psychiatry,
154(1):4–17.
The National Institute of Mental Health (NIMH) (2021).
Depression. [Online; accessed 8-August-2021].
The World Health Organization(WHO) (2021). Depression.
[Online; accessed 6-October-2021].
Zanella-Calzada, L. A., Galv
´
an-Tejada, C. E., Ch
´
avez-
Lamas, N. M., Gracia-Cort
´
es, M., Magallanes-
Quintanar, R., Celaya-Padilla, J. M., Galv
´
an-Tejada,
J. I., and Gamboa-Rosales, H. (2019). Feature extrac-
tion in motor activity signal: Towards a depression
episodes detection in unipolar and bipolar patients.
Diagnostics, 9(1):8.
The Comparison of Various Correlation Network Models in Studying Mobility Data for the Analysis of Depression Episodes
207