Mobile Netw Appl (2018) 23:216–226

https://doi.org/10.1007/s11036-017-0929-3

emHealth: Towards Emotion Health Through Depression

Prediction and Intelligent Health Recommender System
Shiqi Yang1 · Ping Zhou2 · Kui Duan3 · M. Shamim Hossain4 ·
Mohammed F. Alhamid4

Published online: 30 September 2017

© Springer Science+Business Media, LLC 2017

Abstract Depression is an important mental disease of of users psychological and physiological conditions by
global concern. Its complicated etiology and chronic clin- pushing personalized therapy solutions to patients with
ical features make it difficult for users to be conscious of emotional distress. Specifically, this paper first proposes the
their own depression emotion and seriously threaten the system architecture of emHealth. Then, we design person-
patient’s life safety. With the development of e-commerce, alized mobile phone Apps to collect emotional data of users
intelligent recommender system has brought new opportu- with tendentious depressive mood, and find the five main
nities to personalized health monitoring for the users with external characteristics of depression by Pearson correlation
emotional distress. Therefore, this paper puts forward the analysis. We divide 1047 volunteers data into training set
emHealth system, which is an intelligent health recom- and test set, and construct prediction model of depression
mendation system with depression prediction for emotion using decision tree and support vector machine algorithms.
health. This paper explores the monitoring and improvement For the different external factors that lead to depression, we
give personalized recommendation and intelligent decision-
making solution, and push related emotional improvement
suggestions to guide users behavior. Finally, a specific
 Kui Duan application scene is demonstrated where patient’s family
[email protected]
member carry out psychological counseling for the patient,
Shiqi Yang to verify the practicability and validity of the system. The
[email protected]
beneficial effects of this system can meet the needs of the
Ping Zhou electronic market and can be promoted and popularized.
[email protected]
M. Shamim Hossain Keywords E-commerce · Smart devices · Depression
[email protected] detection · Recommender systems
Mohammed F. Alhamid
[email protected]

1 State Grid, Hubei Jingzhou Power Supply Company,

1 Introduction
Jingzhou, China
Due to the rapid development of social economy and the
2 School of Computer Science and Technology, Huazhong remarkable improvement of people’s living standard, atten-
University of Science and Technology, 1037 Luoyu Road,
Wuhan, 430074, China
tion has been paid to the field of medical health care,
especially doctors, psychologists and researchers. Under the
3 School Hospital, Huazhong University of Science pressure of heavy life and work, more and more people
and Technology, Wuhan, China have been confronted with emotional handicap. This kind
4 Department of Software Engineering, College of Computer
of disease which is mainly characterized by emotional dis-
and Information Sciences, King Saud University, turbance is called emotional disorder. World Health Orga-
Riyadh 11543, Saudi Arabia nization pointed out that more than 70% of people will
Mobile Netw Appl (2018) 23:216–226 217

attack their body organs to digest their emotions. Emotional necessary medical information and improve their psycho-
disease will not only bring trouble on patient’s daily life, logical and physiological conditions is lack [23, 24].
such as work/learning performance, social or family rela- Another challenge in emotional diseases research is how
tionships and so on [1, 2]. And if not found timely or not to improve and adjust emotions. Khan [25] monitors the
to take effective treatment measures, this kind of disease emotion and mental state of users through wearable com-
will also bring more serious physical and psychological dis- puting and machine learning algorithms. Tacconi et al. [26]
ease, threatening the patient’s life safety [3–5]. The primary provides support for the early diagnosis of mental illness
challenge in studying emotional disorders is to master the by monitoring the user’s daily activities and emotions.
patient’s emotion changes. At present, the research on user Shang [27] aims to build an easily implementable predic-
emotion detection mainly includes three categories, that is, tion model for individual patient’s depression trajectory
emotion recognition based on audio-visual signals, physio- and design a personalized monitoring scheduling system
logical signals, and multi-modal data [6–10]. These studies from electronic health record (EHR). Paradiso et al. [28]
focus on the emotional detection level, which is designed to proposed a close loop approach between patients and physi-
improve the accuracy of emotion recognition, but the depth cians to monitor mental health and facilitate disease man-
of emotion analysis is not enough, e.g., sad, nervous and agement. These personalized monitoring or management
other passive emotion recognition result is likely to be a systems are relatively simple and have little or no feedback
precursor to depression or other emotional diseases [11–14]. on emotion improvement and regulation.
Therefore, this paper first investigates the research sta- In the face of emotional disorders threatening personal
tus quo of emotional disease, including research methods, safety such as depression, a long-term and effective intel-
research platform and research results. The causes of emo- ligent health recommendation system is urgently needed to
tional disease include stress in work and life, as well as monitor and track the mental status of users. And the sys-
personality, environmental factors, heredity and other pos- tem needs to meet all users of all ages. The occurrence
sibilities [15, 16]. In all emotional disorders, depression of depression is not instantaneous, it is a chronic gradual
has become an important mental and psychological disease process. Therefore, effective monitoring of recommenda-
globally. Depression occurs frequently and is often over- tion system can detect signs of depression in advance [29,
looked and mainly characterized by a marked and persistent 30]. After discovering users tendency of depression timely,
depression. If the treatment of depression is not timely or the system will push the corresponding treatment measures
unreasonable, it may cause the suicide attempt or halluci- for effective regulation to reduce the degree of disease
nations, delusions and other mental illness, causing serious in patients with emotional illness, which is of great sig-
psychological disease and threatening personal safety. Data nificance in the suppression and treatment of depression.
show that about 11 million people suffer from clinical With the development of e-commerce [31–33], intelligent
depression each year in the United States, and the patients recommender system has brought new opportunities for
suicide rate is 10% 15%. Affected by the disease, about $20 personalized monitoring and management of patients with
billion US dollars lose a year [2, 17, 18]. It is predicted that emotional distress [34–37].
by 2020, depression will become the second most disabling Therefore, with the support of Internet of things, mobile
factor in humans, only after cancer [19]. communication, big data and cloud computing [38–40], this
Many studies have made great progress in predicting paper proposes the emHealth system, which is an intelligent
depression. Bhakta and Sau [20]compared five machine health recommendation system for patients with emotional
learning classifiers with respect to three test options for distress. The system uses intelligent mobile devices to col-
depression prediction in aged persons. Ma et al. [21]
lect and analyze user data anytime and anywhere, and
presents a novel and effective approach to depression recog-
predict emotional diseases, then provide effective interven-
nition in the visual modality of videos, which automatically
tion measures for the treatment and health care of patients.
predicts the depression level through two cost-sensitive
In order to address the challenges mentioned above, the
stages. Zhu et al. [22] study on automated depression pre-
main contributions of our research are as follows:
diction based on the deep networks, capturing both the facial
appearance and dynamics from video. These studies are – We proposed a system architecture for intelligent health
very promising in the accuracy of predicting depression, recommendation for patients with emotion disorders.
but the present study does not provide effective interven- – In the framework of emHealth system, we designed
tion measures for the rational improvement of the depressive the personalized mobile phone Apps to collect the
patients after giving predictive diagnosis. Namely, after emotional data of users with tendentious depression
obtaining a lot of medical information of users, how to mood, and find the five main extrinsic characteristics of
analyze these medical features and then give these patients depression by Pearson correlation analysis.
218 Mobile Netw Appl (2018) 23:216–226

– Based on the collected volunteer data, we build the storage can use the cloud platform to collect all the user
predicting model of depression using decision tree and information together.
SVM machine learning algorithm respectively, and the Data analysis layer: After obtaining a large number of
prediction results are compared. user information, we use a powerful cloud platform and
– In view of the main external factors leading to depres- combine multiple machine learning algorithms to analyze
sion, we give personalized recommendation and intel- user characteristics, and then give users a level of disease
ligent decision-making solutions, and push the relevant and an evaluation system.
emotional improvement therapies to guide the user Personalized recommendation and decision making
behavior. layer: Through monitoring users for a long time, reasonable
– Through the designed mobile phone apps, we carried improvement suggestions are given to users’ psychological
out real scene testing where patient’s family member status, and effective treatment solutions are given to patients
conduct psychological counseling for the patient. with higher prevalence.
The architecture of an intelligent health recommendation
The framework of this article is as follows. Section 2 system for patients with emotional disorders is shown in
puts forward the system architecture of emHealth. Section 3 Fig. 1. And the advantages of our proposed intelligent health
presents the collection method of emotional data in sys- recommendation system are as follows:
tem. Section 4 introduces the acquisition of training set
and test set. Section 5 introduces two kinds of prediction – Personalization: By analyzing the specific user’s per-
model of decision tree and support vector machine. Section 6 sonal behavior characteristics, the system provides
gives personalized recommendation and intelligent decision- more information in line with user needs and a very
making solution for depression monitoring and manage- user-friendly experience from apps data collection to
ment. Section 7 presents our practical system platform. And intelligent feedback.
Section 8 summarizes this paper. – Recommendation: Making use of users behavior char-
acteristics, the system infers the information resources
that users need and pushes them, which avoids the
2 System architecture problem that users cant find a reasonable and effective
coping method when they understand their illness, thus
The emHealth system for patients with emotional distress giving users the most timely and effective treatment.
needs to accomplish the following tasks: users information – In consideration of the heterogeneity, scalability, real-
acquisition, information transmission, information storage, time, complexity, and privacy of big data, we shall
information analysis and modeling, behavior decision mak- effectively “mine” the datasets at different levels during
ing, intervention and regulation. So we divide the system the analysis, modeling, visualization, and forecasting,
into five layers according to the required tasks, namely data so as to reveal its intrinsic property and improve the
collection layer, data transmission layer, data storage layer, decision making.
data analysis layer, personalized recommendation and deci- Based on the intelligent health recommender system
sion making layer. Each layer needs to complete tasks as for patients with emotional problems, this paper realizes
follows: long-term effective monitoring and prediction of users
Data collection layer: Intelligent health recommendation depression, and it can also recommend treatment solutions
system first needs to obtain the user’s data which reflects the matched with the level of depression for users. Next, this
emotional diseases, and the data collection forms are var- paper builds and analyzes the actual application platform
ious, including intelligent mobile devices, such as mobile from the aspects of emotional data collection, data analy-
phones, web pages and so on. In this paper, we design the sis, personalized recommendation and intelligent decision-
mobile phone Apps to complete the collection of user data. making.
Data transmission layer: After obtaining the user infor-
mation, it is necessary to transfer data to the analysis plat-
form, and users can use WLAN, Bluetooth, 3G/4G and other 3 Emotional data collection
networks, such as heterogeneous vehicular networks [41,
42] or cooperative multi-agent networks [43] to transmit This paper applies an APP application to collect volunteers
data. emotional data, so as to carry on analysis and detection
Data storage layer: Data storage is divided into local stor- of depressive disorder in the next step. We collected emo-
age and remote storage. Local storage includes personal tional information through the mobile APPs. Questionnaire
information storage, such as mobile phones, and the remote is shown on the mobile Apps, the user can answer the
Mobile Netw Appl (2018) 23:216–226 219

Fig. 1 System Architecture of emHealth

questions about depression on the phone. The psychomet- 3.2 External factor evaluation
ric questionnaires consist of two parts: (1) the assessment of
depressive mood; and (2) the evaluation of external factors We concluded the possible external factors of depression to
that affect depression. 9 aspects, which include family life, external competition,
interpersonal relationship, self-promotion burden, economic
3.1 Depressive mood assessment burden, work pressure, individual personality, coping style,
and social support. External factor evaluation is done by
This paper uses popular Self-Rating Depression Scale [44] analyzing volunteers data from mobile phone APPs ques-
to evaluate depression mood, including 20 questions. tionnaire.
Each question has a score of one to four, and the total
score can directly reflect the students’ depression degree. – Three aspects of family life status, external competi-
According to the score baseline, the score below 53 is tion state and interpersonal relationship are evaluated by
defined as ”no depression,” and the score ranged from 53 personal living environment questionnaire. The ques-
to 62 is mild depression, with a score of 62 72 for moder- tionnaire has 30 questions in all, and each question has
ate depression, and a score of over 72 for severe depression. a score of one to four. The evaluation result is divided
Table 1 shows the specific content of the self-assessment into 5 grades.
table. The greater the score, the higher the depressive ten- – Three aspects of self-promotion burden, economic bur-
dency. den and work pressure are assessed by the personal living
220 Mobile Netw Appl (2018) 23:216–226

Table 1 Self-rating depression

scale Question A little of Some of Good part of Most of
the time the time the time the time

1. I feel down-hearted and blue. 1 2 3 4

2. Morning is when I feel the best. 4 3 2 1
3. I have crying spells or feel like it. 1 2 3 4
4. I have trouble sleeping at night. 1 2 3 4
5. I eat as much as I used to. 4 3 2 1
6. I still enjoy sex. 4 3 2 1
7. I notice that I am losing weight. 1 2 3 4
8. I have trouble with constipation. 1 2 3 4
9. My heart beats faster than usual. 1 2 3 4
10. I get tired for no reason. 1 2 3 4
11. My mind is as clear as it used to be. 4 3 2 1
12. I find it easy to do the things I used to. 4 3 2 1
13. I am restless and cant keep still. 1 2 3 4
14. I feel hopeful about the future. 4 3 2 1
15. I am more irritable than usual. 1 2 3 4
16. I find it easy to ma decisions. 4 3 2 1
17. I feel that I am useful and needed. 4 3 2 1
18. My life is pretty full. 4 3 2 1
19. I feel that others would be better 1 2 3 4
off if I were dead.
20. I still enjoy the things I used to do. 4 3 2 1

pressure questionnaire. The evaluation was divided into sources, i.e., family, friends, teachers or other important
5 levels. The higher the score, the greater the stress in people. Higher scores are considered to represent better
this aspect, and the questionnaire has 30 questions in levels of social support.
all.
Based on the above data collection method, the design of
– About individual personality, we use the internationally
user evaluation interface of App is shown in Fig. 2. Users
renowned Eysenck personality questionnaire to assess
can answer the corresponding questionnaires according to
personality characteristics of individuals. The question-
their own real situation. After the evaluation, the evaluation
naire has two directions of scoring, scores from 43.3 to
data will be stored in the mobile phone, and the evalua-
56.7 is normal personality, scores in the 38.5 43.3 or
tion data will be sent to the remote cloud platform for later
56.7 61.5 is the type of personality tendency, scores
analysis.
less than 38.5 or higher than 61.5 belong to individual
personality.
– The results of the personal coping style questionnaire
4 Acquisition of training set and test set
can reflect the psychological maturity of the person.
There are 60 questions in total. The volunteer can
The effective number (N) of samples in the experiment is
choose yes or no to answer each question, and gets 1 or
1047, each sample represented by a 10 dimensional vector.
0, respectively. The higher the scores, the more promi-
To gather statistics of the collected data, use the formula (1)
nent behavioral characteristics individuals showed in
and formula (2) to calculate the mean and variance, standard
coping styles.
deviation of features. Statistical results are shown in Table 2.
– The part of external social support uses the social sup-
port scale for evaluation, for the use and verification
1 1
N N
of the existing social support. Scale includes 16 items, xi = xij σi2 = (xij − xi )2 (1)
N N
assessing the support volunteers feel from multiple j =1 j =1
Mobile Netw Appl (2018) 23:216–226 221

As shown in Table 3, different variables are related to

depression to some extent. Among them, self-promotion
burden, economic burden, daily work pressure, and other
factors are positively related to depression. Interpersonal
relationships, external social support, coping styles are
negatively correlated with depression. Statistics show that
health accounted for 70%, mild depression accounted for
17%, moderate depression accounted for 9%, and severe
depression accounted for 4%. We divide people with depres-
sion and those who are not depressed into depression and
non-depression. According to the 1047 samples collected,
the data are divided into training set and test set. Train-
ing set accounted for 70% of the total sample, the test
set accounted for 30% of total sample. Depression and
non-depression sample are assigned to the training set and
testing set according to the proportion. The sample alloca-
tion is shown in Table 4. We use decision tree and support
vector machine (SVM) algorithm to establish the depression
prediction model. Through the correlation between features,
Fig. 2 User evaluation interface of APP
we selected five major features related to depression and
established a classification algorithm, and compared the
accuracy of the results predicted by the two algorithms.
Thereinto,i = {1, 2, · · · 10}.
Then, we use Pearson correlation coefficient to analyze
features, and find the correlation between each feature and 5 Decision tree and SVM prediction model
depression. The formula for the correlation between any
feature x and feature y is formula (3). Table 3 shows all We use decision tree and SVM as the prediction model of
the results that features are analyzed using the covariance depression emotion, and make a contrast between the two
matrix, which indicates the correlation results of all feature algorithms of prediction accuracy.
pair.
N    5.1 Decision tree model
1  xij − xi yij − yi
r= (2)
N −1 σx σy The key step of building a decision tree is choosing the opti-
j =1
mal partition of Feature, through the statistical information
Thereinto,i = {1, 2, · · · 10}. entropy and calculating Feature information gain to select
the optimal partition of Feature, the decision tree model of
Table 2 Statistical data of feature distribution depression emotion detection set up by training is shown in
Fig. 3, and the final decision tree is pruned.
Question Mean Variance Standard
deviation 5.2 SVM model
Depressive factors 44.212 98.83 9.941
For the SVM algorithm, we define the optimization objective
1. family life status 17.506 8.764 2.960
that contains soft margin, the optimization objective is:
2. external competition status 15.147 15.777 3.972
3. interpersonal relationship 14.936 17.104 4.136 1  m

4. self-promotion burden 13.841 19.129 4.374 min  ω 2 +C l0/1 (yi (ωT xi + b) − 1) (3)
ω,b 2
5. economic burden 12.198 16.375 4.047 i=1

6. daily work burden 14.523 15.669 3.958 Where C > 0 is a constant, l0/1 is 0/1 loss function the
7. personality characteristics 45.468 57.445 7.579 hinge loss function is used in the algorithm, as shown in
8. coping style 45.062 65.717 8.107 Eq. 5:
9. external social support 32.003 28.634 5.351
lhinge (z) = max(0, 1 − z) (4)
222 Mobile Netw Appl (2018) 23:216–226

Table 3 Feature correlation analysis

1 2 3 4 5 6 7 8 9
Depression External Coping Individual Self- Economic Work Interpersonal External Family
Social style Personality promotion Burden Pressure Relationship Competition Life
Support Burden

Depression 1.000
1 External Social −0.392 1.000
Support
2 Coping Style −0.387 0.889 1.000
3 Individual −0.216 0.165 0.211 1.000
Personality
4 Self-promotion 0.813 −0.361 −0.257 −0.325 1.000
Burden
5 Economic Burden 0.748 −0.229 −0.199 −0.272 0.606 1.000
6 Work Pressure 0.790 −0.198 −0.141 −0.247 0.710 0.555 1.000
7 Interpersonal −0.352 0.257 0.192 0.099 −0.349 −0.301 −0.336 1.000
Relationship
8 External −0.111 0.091 0.101 0.123 −0.113 −0.154 −0.086 0.243 1.000
Competition
9 Family Life −0.123 0.118 0.152 0.065 −0.049 −0.220 −0.145 0.081 0.077 1.000

In the process of solving the model, we use Gauss kernel hierarchical emotional improvement recommendations and
to map the feature space of the model, it is mapped from 5 treatment recommendations based on the ratings of the 5
dimensional space to high dimensional space, and the kernel major characteristics of the depressive disorder and the
function expression, such as formula (6), is shown. intellectual prediction result, and pushed to intelligent ter-
minal devices via the cloud platform, users can choose the
 
 xi − xj 2 regulation according to their own circumstances, or treat
k(xi , xj ) = exp − (5) from simple to complex based on the hierarchical scheme.
2σ 2
At the same time, the treatment plan was scored to enhance
and improve the related treatment according to the user’s
The contrast results between the two algorithms in the
score in the later study. When the emotion prediction result
training set and the test set are shown in Fig. 4. The com-
is non-depression, we pushed for an emotional improvement
parison results show that the accuracy of SVM in the test set
solution at Level 0. When the emotion prediction result is
and training set is better than that of Decision Tree.
depression, based on the differences in the ratings of the 5
main features of depressive emotion, we pushed for treat-
ment at the Level 1 to Level 5. First of all, the scoring
6 Personalized recommendation and intelligent
of the 5 main features is standardized, including the self-
decision making
promotion burden, economic burden, daily work pressure,
external social support, and coping style, unify all feature
After obtaining the user’s personal Apps data and the intel-
scores with a dimension of 10. The standardized formula is
ligent prediction results of the cloud platform, we designed
shown in Eq. 7.

SF i
Table 4 Feature Correlation Analysis SSFi = × 10 (6)
T S Fi
Depression sample Non-depression sample

Training set 235 498

Among them, SFi , T SFi , SSFi respectively expresses the
score of feature Fi , total score and standardized score. Con-
Test set 100 214
sidering the correlation between each trait and the outcome
Mobile Netw Appl (2018) 23:216–226 223

Fig. 3 Decision tree prediction Root

model
F1<=16.5 F1>16.5

F3<=50.5 F3>50.5 F2<=35.5 F2>35.5

F1<=12.5 F1>12.5 N Y N

F1 Self-promotion Burden
F4<=13.5 F4>13.5 F5<=7.5 F5>7.5
F2 External Social Support
F3 Coping Style
F4 Daily Work Pressure
N Y N Y
F5 Economic Burden
N Non-depression
Y Depression

of depression, the 5 features should be weighted. After pressure, the system pushes the music to decompress the
weighting, the final score of any feature Fi is user and instructs the user to do some exercise; When
the user’s depression is mainly caused by the economic

SSFi × rFi rFi >= 0 burden, the system pushes users to learn how to manage
SSFi  = (7)
(SSFi − 10) × rFi rFi < 0 their finances; When the user’s depression is mainly caused
by the burden of self-promotion, system push gives users
Among them, rFi , SSFi , SSFi  respectively expresses the encouragement and praise, while suggesting that users cul-
correlation between feature Fi and depression, standardized tivate a variety of hobbies. The personalized hierarchical
scores and weighted scores. recommendation service is shown in Fig. 5.
Finally, according to the score of 5 features, different
treatment solutions were pushed, and when the subjects
were non-depression, the system pushed music to the user 7 A demonstration system for depressive
to relax when the user’s depression is mainly caused by prediction and therapy recommendation
personal coping style, the system pushes the game or inter-
esting stories to transform the user’s way of thinking, while We designed and implemented an emotion counseling
advising the user to try something new; When the user’s demonstration system based on the theories and methods
depression is mainly caused by external social support, the mentioned above. The following were introduced to our
system pushes the current user status to the surrounding development platform and testing scene.
family, work partners and friends, gives incentives; When
the user’s depression is mainly caused by the daily workload 7.1 Development platform

We developed an app on android mobile phone platform.

App front-end to collect user’s personal data, and the back-
end transmits the user data to the cloud platform center
to analyze and return the result. According to the analysis
result of depression emotion, the App push correspond-
ing therapy solutions to testing users. The cloud platform
adopted is Inspur SmartData Appliance with computing
power of 84 core CPU and 336G RAM.

7.2 Testing scene of emotion counseling

The testing scene is carried out between the patient Brian

and the patients family member Angel, that Angel conduct-
Fig. 4 Comparison of prediction accuracy between decision tree and ing emotion counseling for Brian. When user Brian with
SVM a tendency to depression has completed the questionnaire
224 Mobile Netw Appl (2018) 23:216–226

Level Recommended Treatment Sulution

Level5 Self treatment of patients: giving users

Self-promotion Burden encouragement and praise, suggesting
users to cultivate a variety of hobbies

Economic material burden Psychotherapy: it is recommended that users do

Level4 not blindly pursue economic and material
standards, and learn how to manage money

Music therapy: push music to the user to relax;

Level3 Daily workload pressure Exercise therapy: guide users to release daily stress
through exercise;

Interpersonal psychotherapy: pushing the current

Level2 External social support status of the user to the surrounding family, work
partners, friends and giving users incentives

Persuasion: Converting the user's way of thinking through

Level1 games or interesting stories;
Personal coping style
Psychological Tactics: advise users to try something new;

Level0 Non-depression Music therapy: push music to the user to relax;

Fig. 5 Personalized hierarchical recommendation service

in the mobile phone App, the App returns the results of the users and their family members and friends to adjust the
depression test (as shown in left figure of Fig. 6), and trig- patient’s emotion status, and grasp the patient’s mental sta-
gers the alert mechanism to send the current user’s depres- tus and physical condition, indicating the feasibility of the
sion status to the patient’s family. When Angel received emHealth system.
the reminder of depression for Brian, She could conduct-
ing emotion counseling through the video, voice, phone
call, email and other multimedia channels, to improve 8 Conclusions
the other party’s emotions by encouraging (as shown in
right figure of Fig. 6). Through the interaction between In this paper, we present an intelligent health recommenda-
tion system for patients with depression emotion disorder
to address the challenge of accessing personalized therapies
in the current social status of medical resources shortage.
First of all, in order to solve the five tasks in our emHealth
system, that is, data collection, date transmission, date stor-
age, date analysis and personalized recommendation, we
propose a five layer system architecture of emHealth. Then,
we designed personalized mobile phone Apps to collect
emotional data of voluteers, and found the five main exter-
nal features of depression by Pearson correlation analysis.
We divided the 1047 volunteers data into training set and
test set, constructed prediction model of depression using
decision tree and SVM algorithm respectively, and made
a contrast between the two algorithms of prediction accu-
racy. According to the different external factors that lead
to depression, we give personalized recommendation and
intelligent decision-making solution to push related emo-
tional improvement suggestions to guide users behavior.
Finally, we present a practical scene to verify the feasibil-
ity and effectiveness of our emHealth system. In the future,
we would further improve the recommended content and
Fig. 6 Function graph of APP for emHealth system methods in our system.
Mobile Netw Appl (2018) 23:216–226 225

Acknowledgements The authors would like to extend their sincere 19. Murray CJ, Lopez AD (1997) Alternative projections of mortal-
appreciation to the Deanship of Scientific Research at King Saud Uni- ity and disability by cause 1990-2020: Global burden of disease
versity, Riyadh, Saudi Arabia for funding this research group No. study. Lancet 349(9064):1498–1504
(RG-1437-042). 20. Bhakta I, Sau A (2016) Prediction of depression among senior
citizens using machine learning classifiers. Int J Comput Appl
144(7):11–16
21. Ma X, Di H, Wang Y, Wang Y (2016) Cost-sensitive two-
References stage depression prediction using dynamic visual clues. In: Asian
Conference on Computer Vision, pp 338–351
1. Lin K, Song J, Luo J, Ji W, Shamim Hossain M, Ghoneim A 22. Zhu Y, Shang Y, Shao Z, Guo G (2017) Automated depression
(2017) Green video transmission in the mobile cloud networks. diagnosis based on deep networks to encode facial appearance and
IEEE Trans Circuits Syst Video Technol 27(1):159–169 dynamics. IEEE Trans Affective Comput
2. Chen J, He K, Ruiying D, Xiang Y (2015) Dominating set and 23. Chen M, Shi X, Zhang Y, Wu D, Guizani M (2017) Deep
network coding-based routing in wireless mesh networks. IEEE features learning for medical image analysis with convolutional
Trans Parallel Distrib Syst 26(2):423–433 autoencoder neural network. IEEE Trans. Big Data
3. Wan J, Zou C, Ullah S, Lai C-F, Zhou M, Wang X (2013) Cloud- 24. Wang K, Shao Y, Shu L, Zhu C, Zhang Y (2016) Mobile big data
enabled wireless body area networks for pervasive healthcare. fault-tolerant processing for ehealth networks. IEEE Netw Mag
IEEE Netw 27(5):56–61 30(1):36–42
4. Bultz BD, Carlson LE (2006) Emotional distress: the sixth vital 25. Khan AM (2011) Personal state and emotion monitoring by
sign–future directions in cancer care. Psycho-Oncology 15(2):93– wearable computing and machine learning. In: Bcs-hci
95 26. Tacconi D, Mayora O, Lukowicz P, Arnrich B, Setz C, Troster
5. Chen M, Jun Y, Zhu X, Wang X, Liu M, Song J (2017) Smart G, Haring C (2008) Activity and emotion recognition to support
home 2.0: Innovative smart home system powered by botanical iot early diagnosis of psychiatric diseases. In: International confer-
and emotion detection. Mobile Networks and Applications ence on pervasive computing technologies for healthcare, 2008.
6. Busso C, Deng Z, Yildirim S, Bulut M, Lee CM, Kazemzadeh pervasivehealth, pp 100–102
A, Lee S, Neumann U, Narayanan S (2004) Analysis of emotion 27. Shang W (2016) Depression management using electronic health
recognition using facial expressions, speech and multimodal infor- record. Individual progression prediction. Master’s thesis Univer-
mation. In: International Conference on Multimodal Interfaces, sity of Washington
pp 205–211 28. Paradiso R, Bianchi AM, Lau K, Scilingo EP (2010) Psyche:
7. Chen M, Zhang Y, Qiu M, Guizani N, Hao Y (2017) SPHA: personalised monitoring systems for care in mental health. In:
Smart Personal Health Advisor Based on Deep Analytics. IEEE Engineering in Medicine and Biology Society, vol 2010, pp 3602–
Commun 3605
8. Yang Z, Narayanan S (2016) Lightly-supervised utterance-level 29. Wu C, Yoshinaga T, Ji Y, Murase T, Zhang Y (2017) A rein-
emotion identification using latent topic modeling of multimodal forcement learning-based data storage scheme for vehicular ad hoc
words. In: IEEE International Conference On Acoustics, Speech networks. IEEE Trans Veh Technol 66(7):6336–6348
and Signal Processing, pp 2767–2771 30. Chen M, Jun Y, Hao Y, Mao S, Hwang K (2017) A 5g cognitive
9. Chen M, Zhou P, Fortino G (2017) Emotion communication system for healthcare, big data and cognitive computing. Big Data
system. IEEE Access 5:326–337 Cognitive Comput 1(1)
10. Lin K, Chen M, Deng J, Hassan MM, Fortino G (2016) Enhanced 31. Chen M, Miao Y, Hao Y, Kwang K (2017) Narrow Band Internet
fingerprinting and trajectory prediction for iot localization in smart of Things. IEEE Access 7
buildings. IEEE Trans Autom Sci Eng 13(3):1294–1307 32. Chen M, Qian Y, Hao Y, Li Y, Song J (2018) Data-driven comput-
11. Power M, Dalgleish T (2015) Cognition and emotion: from order ing and caching in 5G networks: architecture and delay analysis.
to disorder. Psychology Press IEEE Wirel Commun 25(1)
12. Chen M, Hao Y, Hwang K, Wang L, Wang L (2017) Disease 33. Chen M, Yang J (2018) 5G-Smart diabetes: towards personal-
prediction by machine learning over big data from healthcare ized diabetes diagnosis with healthcare big data clouds. IEEE
communities. IEEE Access 5(1):8869–8879 Commun
13. Hwang K, Chen M (2017) Big-data analytics for cloud IoT and 34. Schubert P, Ginsburg M (2010) Virtual communities of transac-
cognitive learning. John Wiley & Sons tion: the role of personalization in electronic commerce. Electron
14. Lin K, Xia F, Wang W, Tian D, Song J (2016) System design for Mark 10(1):45–55
big data application in emotion-aware healthcare. IEEE Access 35. Zou C, Zhang D, Wan J, Hassan MM, Lloret J (2017) Using con-
4:6901–6909 cept lattice for personalized recommendation system design. IEEE
15. He D, Chan S, Zhang Y, Yang H (2014) Lightweight and confi- Syst J 11(1):305–314
dential data discovery and dissemination for wireless body area 36. Adomavicius G, Tuzhilin A (2005) Toward the next generation of
networks. IEEE J Biomed Health Inf 18(2):440–448 recommender systems: a survey of the state-of-the-art and pos-
16. Chen M, Ma Y, Li Y, Wu D, Zhang Y, Youn C (2017) Wearable sible extensions. IEEE Trans Knowledge Data Eng 17(6):734–
2.0: enable Human-cloud integration in next generation healthcare 749
system. IEEE Commun 55(1):54–61 37. Zhang Y, Chen M, Huang D, Di W, idoctor YL (2016) Per-
17. Neugebauer R (1999) Mind matters: the importance of mental dis- sonalized and professionalized medical recommendations based
orders in public health’s 21st century mission. Am J Public Health on hybrid matrix factorization. Futur Gener Comput Syst
89(9):1309–1311 66:30–35
18. Ap Association (2013) The diagnostic and statistical manual of 38. Dai W, Qiu L, Ana W, Qiu M (2016) Cloud infrastructure resource
mental disorders fifth edition: Dsm-5. Psychiatry Res 189(1):158– allocation for big data applications. IEEE Transactions on Big
159 Data
226 Mobile Netw Appl (2018) 23:216–226

39. Li Y, Gai K, Ming Z, Zhao H, Qiu M (2016) Intercrossed multimode communications in heterogeneous vehicular telemat-
access controls for secure financial services on multimedia ics. IEEE Trans Intell Transp Syst 16(6):3033–3049
big data in cloud systems. ACM Transactions on Multimedia 42. Tian D, Zhou J, Sheng Z, Chen M, Ni Q, Leung VCM (2017) Self-
Computing, Communications, and Applications (TOMM) 12(4s): organized relay selection for cooperative transmission in vehicular
67 ad-hoc networks. IEEE Transactions on Vehicular Technology
40. Li J, Qiu M, Ming Z, Quan G, Qin X, Zonghua G (2012) Online 43. Tian D, Zhou J, Sheng Z (2017) An adaptive fusion strategy for
optimization for scheduling preemptable tasks on iaas cloud sys- distributed information estimation over cooperative multi-agent
tems. J Parallel Distributed Comput 72(5):666–677 networks. IEEE Trans Inf Theory 63(5):3076–3091
41. Tian D, Zhou J, Wang Y, Yingrong L, Xia H, Yi Z (2015) 44. Zung WWK (1965) A self-rating depression scale. Arch Gen
A dynamic and self-adaptive network selection method for Psychiatry 12(12):63