=Paper=
{{Paper
|id=Vol-2786/Paper53
|storemode=property
|title=Detection of Bipolar Disorder Using Machine Learning with MRI
|pdfUrl=https://ceur-ws.org/Vol-2786/Paper53.pdf
|volume=Vol-2786
|authors=R Sujatha,K Tejesh,H Krithi,H Rasiga Shri
|dblpUrl=https://dblp.org/rec/conf/isic2/SujathaTKS21
}}
==Detection of Bipolar Disorder Using Machine Learning with MRI==
https://ceur-ws.org/Vol-2786/Paper53.pdf
445
Detection of Bipolar Disorder Using Machine Learning
with MRI
R Sujatha, K Tejesh, H Krithi and H Rasiga Shri
School of Information Technology and Engineering, Vellore Institute of Technology, Vellore, Tamilnadu, India
Abstract
Bipolar disorder is a mental ailment caused by maximal mood swings with emotional highs and lows. Nowadays, this has
become the most common abnormality related to mental health and furthermore it is ignored by people of all age groups.
Bipolar disease is generally heritable but not all siblings of the family will be having it though, and will be having same
genetics and the factors which can be risky. Here we use random forest algorithm, along with the Mag- netic Resonance
Imaging (MRI) information. The utility of these irregularities in recognizing individual bipolar disorder patients from state
of mind issue or health controls define patients dependent on their illness. Here we use machine learning algorithms like
Random forest algorithm and CNN-mdrp(multimodal disease risk prediction) for the accuracy .We give the risk factor and
stage of the healthy patient with the attributes we collected from the MRI. We use a trained dataset and machine learning
algorithms mentioned above to get the output. Voxel-Based Morphometry (VBM) will be used to dividing and pre-processing
the MRI infor- mation obtained. To see the changes in Gray Matter (GM) and White Matter (WM) of the different data groups
individually, a simple equation is use and also the Principle Component Analysis will be used and The project gives you the
output showing that CNN MDRP with random forest has high accuracy than other algorithms in bipolar disease prediction.
Keywords
Magnetic Resonance Imaging, Random forest algorithm, Gray mat- ter, White matter, Voxel based morphometry, CNN-mdrp
(multimodal disease risk prediction)
1. Introduction that detects Bipolar diseases with more accuracy and
less cost using CNN-MDRP algorithm and random forest
Mental sickness is one of the most dangerous and life- classification which is proven to be giving more accurate
threatening ailments. One must always take good care of results.
their mental state of mind. There are various proposed There is no accuracy in the present algorithms. Im-
system to identify the mental conditions of an individual. provising the algorithms to include more productivity
These systems were developed using some combination of the framework subsequently improving its working
of machine learning algorithms working with collected isn’t finished. There is no appropriate online bipolar de-
data sets to train and test the model.[2] termination framework that are utilized in clinics and
However, this existing system has some defects that hospitals as of now. [8]
are to be rectified. The collected samples for the data set The proposed random forest Machine Learning Algo-
is insufficient. The attributes that are used in the data rithm for classifying the data not only will predict the dis-
sets cannot be a fixed one, it changes with person and eases but also its sub diseases, predicting sub diseases in-
the disease they suffer from. [12] creases the accuracy of the system and the Deep learning
Detection of Bipolar disease at an early stage enables algorithm CNN-mdrp is used for the accuracy prediction
patients to have a much higher chances of recovery and and from the results it shows that CNN-mdrp has high
healthy survival. Therefore, even though we can’t pre- ac- curacy in fact better than CNN-udrp algorithm.
vent it from affecting us, we can certainly detect this at
an early stage so as to provide the appropriate medical
help at the right time to the right people. [5] 2. Existing System
The main motive of this paper is to develop a system
The current system for Bipolar disorder disease predic-
International Semantic Intelligence Conference (ISIC), February tion system uses random forest algorithm which actually
25-27,2021 New Delhi, India determines the features present in the data set and makes
Envelope-Open r.sujath@vit.ac.in (R. Sujatha); k.tejesh2017@vitstudent.ac.in (K. it a decision factor at the level of dec. [17] Now the ob-
Tejesh); krithi.h2017@vitstudent.ac.in (H. Krithi);
hrasiga.shri2017@vitstudent.ac.in (H. Rasiga Shri)
tained testing samples are deliberately compared with
GLOBE https://github.com/TEJESH-K (K. Tejesh); the decision attributes at each level to test if a patient
https://github.com/krithi0506 (H. Krithi); is screened positive or negative. Machine could predict
https://github.com/Rasiga-Shri (H. Rasiga Shri) only the disease but cannot predict sub types and risk
Orcid stage of the diseases. It fails to anticipate and predict
© 2021 Copyright for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0). all potential states of the people. In the past, System
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
446
took care of just organized information. The standing ence Electrode (RE). The detection of Lithium is based on
associations arrange a blend of ML algorithms which are presence of potentiometric Ion-Selective Electrode (ISE)
reasonably good at predicting illnesses or the disease. accompanied by a nanostructured solid contact. [4]
[13] But the limitations with the common framework In this paper, they raised a SVM model using the NumPy
exists. A machine can predict and explain a disease yet library in python. They considered the structural and
can’t speak about the sub kinds of the diseases caused by functional attributes taken from the MRI report of the
the already existing disease. [18] patients. They also used many unique features collected
from the MRI report of the affected individuals. This new
invention of combining the structural and the function-
3. Literature Survey als attributes of the brain anatomy improved the overall
efficiency of the system. It also increased the accuracy
This paper solves many problems related to Bipolar dis-
of the system.[11]
order detection. Many a times the doctors and patients
In this paper, they examined the neurotrophic factors
mistake Major Depressive Disorder as Bipolar Disorder,
of the individuals, and analysed if these factors will in-
this paper shows up many solutions for the doctors and
fluence in the timely detection of the Bipolar disorder.
patients to get diagnosed and treated with the disease.
They employed the model-based algorithm for the iden-
In another system, the patients with neuroanatomical
tification. As a result of this paper neurotrophic factors
abnormalities were classified using Three- Dimensional
were successfully found to assess the disorder.[26]
magnetic resonance imaging (3DMRI). This system also
In this paper, they have used a sensitized T-shirt that
focuses on the unaffected individuals who are blood- re-
records the patient’s Neurotrophic factors. Here, decision
lated to the patients suffering from Bipolar disorder.[6]
tree algorithm is employed to classify the patients suf-
This paper uses machine learning algorithm to screen
fering from Bipolar disorder and the healthy ones. This
bipolar disorder by using Mood Disorder Questionnaire
T-shirt accurately predicted the Neurotrophic factors of
(MDQ), with decision tree algorithm. the data set is fed
the affected individuals and predicted the border value
into the decision tree classifier which determines the
range of the factors.[16]
significant features in dataset and make it decision factor
In this paper, the bipolar state relapse disorder is identi-
at that level of decision tree.[14]
fied among the patients using certain attributes collected
In this paper, early detection of bipolar disorder is
from patient’s smartphones. Nowadays, smartphones
done by using screening question- naire data for around
are capable of monitoring the individual’s heart rate,
300 respondents and it served as a knowledge base to be
blood pressure rate etc. This disorder was successfully
processed using back propagation algorithm, which was
detected.[21]
the drawback in the previous research paper.[9]
The other paper, introduced an approach to study the
In this paper, the patients suffering from BD were ana-
mood disorder with respect to the patterns present in
lyzed and the change in their mental states was identified.
the emotions of the affected people. They introduced
A dataset was created by monitoring their characteristics
LASM - Latent Effective model to locate the connections
and the features was gathered to this disorder. [10]
in the emotions of people. They used six videos for the
In this paper ,1 dimensional time domain data of the
detection of the disorder. These videos were the recorded
fNIRS, is acquired while preparing tasks, which is used
emotional videos. The study concluded that this method
to train a set of neural networks for the diagnosis of
gave more accuracy than the existing one.[15]
common mood disorder, the Bipolar Disorder. With this
In this paper, this system detects the mood disorder.
the healthy individuals and the affected ones were classi-
The system works with LSTM- based approach for mod-
fied and deep learning algorithms were employed in the
eling the long-range speech of people, it analyses the
detection of the disorder.[7]
mood of the person from their speech. The database con-
In this paper, 65 individuals were isolated and exam-
tained of recorded speeches of different people mixed
ined. Their activities were rec- orded. Here, Voxel-based
with different kinds of emotions. The SVM algorithm is
morphometry was used to investigate the brain anatomy
also employed for the detection patients suffering mood
from the patient’s MRI. The study included 26 Bipolar
disorder.[23]
patients and 38 Healthy individuals. The t-test based sta-
This paper identifies the periods of depression of the
tistical method was also employed to group the affected
patients suffering from bipolar disorder using the move-
individuals.[3]
ments recorded from mobile location. They have used
In the following paper, a complete system is been de-
the quadratic linear regression model to monitor the char-
veloped for the Therapeutic Drug Monitoring (TDM).
acteristics of the affected individuals. This also facilitates
Here, the disease is diagnosed from the lithium present
the doctors to identify those people who are in the need
in the sweat of people affected by this disorder. This
of critical care.[20]
platform incorporates paper fluidics and the stable Refer-
In this paper they monitored the patients of BD and
447
concluded that the report generated from the smartphone
data matched up with the depressive states. This paper
resulted that high methodological rigor along with large
sample of patients affected with
bipolar disorder having manic symptom addressing
doze factors before being implemented using monitoring
tool.[19]
4. Methodology
4.1. DICOM
Expanded as Digital imaging and communications in
medicine, is used as a standard for taking medical data
as input .For example when we try to integrate medical
imaging appliances , we use the DICOM standard. And
here for taking the inputs from the MRI scans, we use
this standard for getting the data and using it for next
step. It is an international standard to send , save and
print the medical imaging data, and The National Elec- Figure 1: Principal component analysis of Bipolar dataset
trical Manufacturers Association(NEMA) holds the all
the copyrights of the above standard After the process of
getting the data from the MRI scanning , we are going makes the analysis process of data much easier and it
to covert into NIFTI and segment the data obtained into makes it faster for machine learning algo- rithms as there
white matter(WM) and grey matter(GM) which is later be is no extra variables to process. [22]
undergoing normalization process and then the obtained
output will be going through smoothing will generate a
4.4. Correlation of Matrix of the Dataset
3D Mask finally, which is used in the next process.
The Figure 2 represents the correlation matrix which
4.2. Voxel based Morphometry gives us the information about the correlation coeffi-
cients between the values taken, above colorful matrix
Voxel based Morphometry technique is used to process represents all the values of the considered dataset and
the identified defects in the brain and prepossess them. each cell in the image shows the correlation between the
The Voxel-based morphometry computes the changes values of the dataset. It is mainly used to congregate the
that occurs in the structure of the brain that is obtained data which is taken as the input and analyze how it going
from the magnetic resonance imaging (MRI) . This is used to process and work, this matrix is mainly used in the
to identify the structural changes that occurs in brain of advanced analysis or as the basic fundamental structure
the affected individuals. The brain of the affected patients, for the advanced analysis.
consist of greater gray matter volume in the left temporal
lobe and central gray matter structures bilaterally.[1] So,
Voxel Based Morphometry is used to identify structural
changes in voxel-based comparison of multiple brain MRI
images.
4.3. Principle Component Analysis
The Figure 1 represents the principal components of bipo-
lar that contains differenti- ated voxels got by using cer-
tain methods are represented as Voxel of Interests. PCA
is used to decrement the proportions of the Voxel that are
obtained from Interests data. Here PCA is a dimension-
ality reduction method used in large datasets to reduce
dimen- sions by transforming large set of variables into
smaller set which contains almost every information in
the large sets. We are doing this because smaller datasets Figure 2: Correlation Matrix of the Dataset
448
4.5. Scatter and Density plot 5. Proposed System
Figure 4: Architecture diagram of our system
The Figure 4 represents the architecture of our system
Figure 3: Scatter and Density plot
and the application is been developed both in front- end
The above Figure 3 The scatter and density plot rep- and back-end. SQLite is used for storing all the patient
resents above gives the details of the values distributed and doctor information as well as reports. The Random
in the data set we have taken , and it is very useful in forest Machine Learning Algorithm not only will pre-
comparing and plotting the scatter plot and it also shows dict the diseases but also its sub diseases. Map Reduce
us how many dots or the values are being concentrated Algorithm which increases the efficiency of the opera-
on a single d plane or the area considered, by this we tion and also it reduces the retrieval time of the query.
can assess the characteristics of the data and also we can [25] The proposed system is different from the ances-
predict the behavior up to an extent. tor’s thought of execution It uses random forest machine
learning procedure for calculating diseases and its respec-
tive sub diseases. which in turn increase the efficiency
and performance and query response time is reduced
too. Along with that, it gives separate patterns to each
patient which gives the patient personalized experience.
In addition to that, it provides definite rations for specific
patients to pattern his/her condition. Thus, making our
system broadly open by all at moderate cost. The predic-
tion accuracy of the CNN-MDRP algorithm reaches 94%
compared to other prediction algorithms.
All the data obtained about the patients from the hos-
pital management is stored in this particular module like
the MRI scans and the other types of data, mainly here we
will be using both the structured and unstructured values
of the attributes of brain anatomy, structured data refers
to the data which is obtained from the reports of the
patient and the unstructured data is something which we
get from the patient’s medical history and the informal
conversations with the doctor and etc.
Training set and Testing data, for producing sophisti-
cated results we perform training set where initial data
help program how technologies like neural networks are
449
per- formed. In testing data, the obtained data will be 6.1. Creating textual data
checking for the execution of test case and verify the ex-
All the data collected from the clinic or the hospital, we
pected output in any of the software applications. DATA
use the word and insert into the first layer to remove the
Labelling preparation, in this module we will be identify-
false wording, the text will be represented as a vector,
ing raw data and further adding meaning and informative
we can also call it as the preprocessing. Here, each word
to provide the context so that we can prepare the data
will be outlined as Rd dimensional variable, where dia as
that could be enriched further. In this section we try to
50, thus, a text that incorporates n words can be depicted
append or enrich the obtained data with relevant context
as Tx = (tx1, tx2, ···, txn), Tx ∈ Rd × n.
gathered from other additional sources. Further this data
is sent to the random forest classifier for classification
of data ,Will also quantify the values of a out- come by 6.2. CNN text transforming level
providing a framework, CNN MDRP, uses both the struc-
In each case we assess words. In some words, we incline
tured and unstructured data for the prediction process,
towards 2 words from frontal and back of every factor of
we collect the data form the medical clinics or hospitals
words in this issue.
and use CNN-MDRP to process the data and predict the
risk of the disease, whereas CNN-UDRP which is in the
existing system used only the structured data gives less 6.3. CNN text pool layer
accuracy than the proposed one. Taking the convolution layer output or the data as the
grouping or gathering level information, we utilize a
6. CNN MDRP large gathering activity as possible (1-max gathering).
The motivation behind why you pick the maximum or the
Collection CNN-MDRP is performed to foresee who is greatest gathering system is the part of each expression
influenced with the illness in an effective manner by is n’t absolutely equivalent; from this we can decide the
utilizing convolutional neural network that utilizes struc- content or the matter which is useful to the system to
tured and unstructured information from the hospital. To proceed to the next step.
begin with, here we use the potential segment model to
discover lacking data from the clinical records, from the 6.4. Fully connected layer of text CNN
dataset which we consider it as repository. Besides, with
the assistance of factual information, we could manage This particular layer is related to the neural network
the fundamental degenerative illness which is available which is fully connected, we have a process of calculating
in the past. Furthermore, dealing with the structured it in this layer, we have a formula which is Hf3=Wa3
data by talking with clinic or the hospital specialists to Hf2=Bv3,where we say Hf3 is considered as the total
get the reusable information and properties. For unorga- connection level, and we can be considered as the artifact
nized area, properties can be precisely abused through and Be can be considered as the deviations part
the CNN rule. In this way, we infer that the CNN- MDRP
works the best for this specific problem. 7. Result and discussions
For the assessment in the examination. To start with,
we indicate TP (the number of occurrences accurately Experimental results of this system indicate that among
anticipated as required), FP, TN and FN as evident pos- the existing algorithms CNN-mdrp gives us the best re-
itive false positive (the quantity of cases erroneously sults with higher accuracy. In this system we used CNN-
anticipated as required), True negative (the quantity of mdrp algorithm and made use of the structured and the
examples accurately anticipated as not needed) and False unstructured data of the patients gathered from the clini-
negative (the number of examples mistakenly anticipated cal records of the hospital. No other system worked on
as not needed), separately. At that point, we can get four both structured and unstructured data. Our proposed al-
estimations: gorithm gives an accuracy of 94.3%. Thus, the introduced
accuracy, precision, recall and F1-measure as follows: combination of the algorithms gave better results when
1. Accuracy = (TruePos + TrueNeg)/(TruePos + False- compared to the existing systems. So this proposed sys-
Pos + TrueNeg + FalseNeg) tem reduces the error rate and simultaneously increases
2. Precision = TruePos/(TruePos + FalsePos) the percentage of accuracy. And below we show the com-
3. Sensitivity = TruePos/(TruePos + FalseNeg) parison and the explanation of the existing algorithms.
4. Specificity = TrueNeg/(TrueNeg + FalsePos)
The process could be divided into in to 5 different parts:
450
7.2. SVM
Support vector machine algorithm (SVM )also works
based on supervised learning for classification of prob-
lems. Here the optimal solution is obtained by trans-
forming data by using techniques like kernel trick. The
obtained accuracy for this algorithm is 88.9
7.3. DECISION TREE
The next algorithm used in the existing system is decision
tree part of supervised learning used for classification of
problem. Works by following a set of if-else condition
to represent the data and categorize them. The obtained
accuracy for this algorithm is 91.3
7.4. RANDOM FOREST
The proposed system uses random forest algorithm since
it consists many decision trees within them, they use fea-
Figure 5: CNN-UDRP Using only Structured Data
ture randomness while building each single tree to create
the uncorrelated forest containing trees so that accuracy
produced by this system automatically increases. The
obtained accuracy for this algorithm is 94.3%.
Figure 6: CNN-MDRP Using both Structured and Unstruc-
tured Data
Figure 7: Graphical representation of Algorithms
7.1. LINEAR
The existing algorithms which have used previously are
linear regression, SVM, Decision tree and the proposed
one is random forest algorithm. Here what linear model
means is it works completely based on supervised learn-
ing which is a part of machine learning. Prediction of
value is done on independent variables by using regres-
sion models. The obtained accuracy for this algorithm is
63.3
451
Table 1
Comparison of Algorithms
Algorithm Accuracy Precision Sensitivity Specificity
Linear 63.3 % 0.6 0.7 0.58
SVM 88.9 % 0.88 0.8 0.86
Decision Tree 91.3 % 0.93 0.90 0.97
Proposed Model 94.3 % 0.94 0.96 0.98
8. Conclusion and Future work controls using MRI. In 2019 Medical Technologies Con-
gress (TIPTEKNO) (pp. 1-4). IEEE.
The proposed system is completely based on the GUI [7] Evgin, H. B., Babacan, O., Ulusoy, İ., Hoşgören, Y.,
where it focuses on reliability, scalability and how well Kuşman, A., Sayar, D., ... Özgüven, H. D. (2019, April).
it is adaptable to the user and this system can be easily Classification of fNIRS Data Using Deep Learning for
expanded for any future requirements. This model could Bipolar Disorder Detection. In 2019 27th Signal Process-
help many people by ultimately decreasing the cost re- ing and Communications Applications Conference (SIU)
quired for treatment. Also this proposed system would (pp. 1-4). IEEE.
help in satisfying the need for preparing the required [8] Isometsä, E., Suominen, K., Mantere, O., Valtonen,
instruments for clinical studies which is made available H., Leppämäki, S., Pippingsköld, M., Arvilommi, P. (2003).
to doctors also for others who need to know the mental The mood disorder questionnaire im- proves recognition
status of the patient. General doctors can use this instru- of bipolar disorder in psychiatric care. BMC psychiatry,
ment for initial diagnostics of mental condition of the 3(1), 8.
patient. [9] Fitriati, D., Maspiyanti, F., Devianty, F. A. (2019,
September). Early Detection Application of Bipolar Dis-
orders Using Backpropagation Algorithm. In 2019 6th
9. References International Conference on Electrical Engineering, Com-
[1] Sarıçiçek, A., Yalın, N., Hıdıroğlu, C., Çavuşoğlu, B., puter Sci- ence and Informatics (EECSI) (pp. 40-44). IEEE.
Taş, C., Ceylan, D., ... Özerdem, A. (2015). Neuroanatom- [10] Grünerbl, A., Muaremi, A., Osmani, V., Bahle, G.,
ical correlates of genetic risk for bipolar disorder: a voxel- Oehler, S., Tröster, G., ... Lukowicz, P. (2014). Smartphone-
based morphometry study in bipolar type I patients and based recognition of states and state changes in bipolar
healthy first-degree relatives. Journal of affective disor- disorder patients. IEEE Journal of Biomedical and Health
ders, 186, 110-118. Informatics, 19(1), 140-148.
[2] Vita, A., De Peri, L., Sacchetti, E. (2009). Gray [11] Li, H., Cui, L., Cao, L., Zhang, Y., Liu, Y., Deng,
matter, white matter, brain, and intracranial volumes in W., Zhou, W. (2020). Identification of bipolar disorder us-
first‐episode bipolar disorder: a meta‐analysis of mag- ing a combination of multimodality magnetic resonance
netic resonance imaging studies. Bipolar disorders, 11(8), imaging and machine learning techniques.
807-814. [12] Hirschfeld, R. M. (2002). The Mood Disorder Ques-
[3] Cigdem, O., Beheshti, I., Demirel, H. (2018). Effects tionnaire: a simple, patient-rated screening instrument
of different covari- ates and contrasts on classification of for bipolar disorder. Primary care compan- ion to the
Parkinson’s disease using structural MRI. Computers in Journal of Clinical Psychiatry, 4(1), 9.
biology and medicine, 99, 173-181. [13] Van Zaane, J., van den Berg, B., Draisma, S., Nolen,
[4] Occhiuzzi, C., Parrella, S., Camera, F., Nappi, S., W. A., van den Brink, W. (2012). Screening for bipolar
Marrocco, G. (2020). RFID-based Dual-Chip Epidermal disorders in patients with alcohol or sub- stance use dis-
Sensing Platform for Human Skin Monitor- ing. IEEE orders: performance of the mood disorder questionnaire.
Sensors Journal. Drug and Alcohol Dependence, 124(3), 235-241.
[5] Eker, C., Simsek, F., Yılmazer, E. E., Kitis, O., Cinar, [14] Jadhav, R., Chellwani, V., Deshmukh, S., Sachdev,
C., Eker, O. D., ... Gonul, A. S. (2014). Brain regions H. (2019, Janu- ary). Mental Disorder Detection: Bipolar
associated with risk and resistance for Disorder Scrutinization Using Machine Learning. In 2019
bipolar I disorder: a voxel‐based MRI study of patients 9th International Conference on Cloud Computing, Data
with bipolar disorder and their healthy siblings. Bipolar Sci- ence Engineering (Confluence) (pp. 304-308). IEEE.
disorders, 16(3), 249-261. [15] Huang, K. Y., Wu, C. H., Su, M. H., Kuo, Y. T.
[6] Cigdem, O., Soyak, R., Aydeniz, B., Oguz, K., Demirel, (2018). Detecting uni- polar and bipolar depressive disor-
H., Kitis, O., ... Unay, D. (2019, October). Classification of ders from elicited speech responses using latent affective
healthy siblings of bipolar disorder patients from healthy structure model. IEEE Transactions on Affective Com-
puting.
452
[16] Migliorini, M., Mariani, S., Bianchi, A. M. (2013,
July). Decision tree for smart feature extraction from
sleep HR in bipolar patients. In 2013 35th Annual Inter-
national Conference of the IEEE Engineering in Medicine
and Biology So- ciety (EMBC) (pp. 5033-5036). IEEE.
[17] Castelo, M. S., Carvalho, E. R., Gerhard, E. S.,
Costa, C. M. C., Ferreira, E. D., Carvalho, A. F. (2010). Va-
lidity of the Mood Disorder Questionnaire in a Brazilian
psychiatric population. Brazilian Journal of Psychiatry,
32(4), 424- 428.
[18Nederlof, M., Heerdink, E. R., Egberts, A. C. G.,
Wilting, I., Stoker, L. J., Hoekstra, R., Kupka, R. W. (2018).
Monitoring of patients treated with lithium for bipolar
disorder: an international survey. International journal
of bipolar dis- orders, 6(1), 12.
[19] Faurholt-Jepsen, M., Bauer, M., Kessing, L. V.
(2018). Smartphone- based objective monitoring in bipo-
lar disorder: status and considerations. Inter- national
journal of bipolar disorders, 6(1), 1-7.
[20] Palmius, N., Tsanas, A., Saunders, K. E. A., Bilder-
beck, A. C., Geddes, J. R., Goodwin, G. M., De Vos, M.
(2016). Detecting bipolar depression from geographic lo-
cation data. IEEE Transactions on Biomedical Engineer-
ing, 64(8), 1761-1771.
[21] Schleusing, O., Renevey, P., Bertschi, M., Koller, J.
M., Paradiso, R. (2011, March). Monitoring physiolog-
ical and behavioral signals to detect mood changes of
bipolar patients. In 2011 5th International Symposium
on Medical Information and Communication Technology
(pp. 130-134). IEEE.
[22] Conus, P., Macneil, C., McGorry, P. D. (2014). Pub-
lic health signifi- cance of bipolar disorder: implications
for early intervention and prevention. Bi- polar disorders,
16(5), 548-556.
[23] Yang, T. H., Wu, C. H., Huang, K. Y., Su, M.
H. (2016, October). De- tection of mood disorder us-
ing speech emotion profiles and LSTM. In 2016 10th In-
ternational Symposium on Chinese Spoken Language
Processing (ISCSLP) (pp. 1-5). IEEE.
[24] McCormick, U., Murray, B., McNew, B. (2015). Di-
agnosis and treat- ment of patients with bipolar disorder:
A review for advanced practice
nurses. Journal of the American Association of Nurse
Practitioners, 27(9), 530- 542.
[25] Rocha-Rego, V., Jogia, J., Marquand, A. F., Mourao-
Miranda, J., Sim- mons, A., Frangou, S. (2014). Exam-
ination of the predictive value of structural magnetic
resonance scans in bipolar disorder: a pattern classifica-
tion ap- proach. Psychological medicine, 44(3), 519-532.
[26] Zheng, Y., He, S., Zhang, T., Lin, Z., Shi, S., Fang, Y.,
... Liu, X. (2019). Detection Study of Bipolar Depression
Through the Application of a Model- Based Algorithm
in Terms of Clinical Feature and Peripheral Biomarkers.
Fron- tiers in psychiatry, 10, 266.