A Risk Factors Screening Method in the Context-aware System of
Hypertension
Duoyi Xie
1,2
, Guixia Kang
1,2
and Longfeng Chen
1,2
1
Key Laboratory of Universal Wireless Communications, Ministry of Education,
Beijing University of Posts and Telecommunications, No.10 Xitucheng Road, Beijing, China
2
Wuxi BUPT Sensory Technology and Industry Institute CO.LTD, Wuxi, China
Keywords: Hypertension, Context-aware System, Feature Selection.
Abstract: Hypertension has become a health problem that seriously endangers human life and is the leading cause of
cardiovascular disease. Many patients do not know exactly whether their blood pressure is well controlled
or not, which makes their conditions worse. A context-aware intelligent system can help patients to analyse
their control situation of blood pressure (BP) and provide feedback. It is especially important to determine
whether the risk-factors input in the context-aware system of hypertension is appropriate. The choice of risk
factors will affect the classification performance and accuracy of the system. The risk factors screening
method for hypertension proposed in this paper combined the random forest algorithm and stability
selection (RFSS). It can remove the redundant context information, and leave the key factors of BP control
situation. Experimental results showed that the prediction accuracy achieved more than 77% prediction
accuracy, and dimension of risk factors reduced by 59%. The results indicated that RFSS is an effective
method in the screening of risk factors and the prediction of hypertension.
1 INTRODUCTION
Hypertension is a leading cause of death and
disability-adjusted life-years worldwide (Lim et al.,
2012). There are 270 million hypertensive patients
in China. Annual patient increases by more than 10
million. More than 1 million people die from high
blood pressure every year in China. Three-quarters
of the survived patients were disabled due to
hypertension. However, the awareness rate and
control rate of hypertensive population are only 36%
and 28%, respectively (Chen et al., 2014). Blood
pressure of 120/80 mm Hg or higher is linearly
related to risk for fatal and nonfatal stroke, ischemic
heart disease, and noncardiac vascular disease, and
each increase of 20/10 mm Hg doubles the risk for a
fatal cardiovascular disease event (Carey and
Whelton, 2012). Evidence shows that hypertension
can be prevented and controlled through monitoring
and treatment (Whitworth and Chalmers, 2004).
Wireless eHealth (WeHealth) has developed
rapidly in China in recent years, and remote
hypertension monitoring is one of its most important
applications (Kang and Zhang, 2010). The
hypertension monitoring system can increase the
awareness rate and control rate of hypertensive
population by feeding back their situation of blood
pressure (Sandi et al., 2013). The management of
hypertensive patients is mainly based on the blood
pressure levels. However, the threshold of
hypertension will be different due to the differences
in patient function. Decisions on the management of
hypertensive patients should not only take blood
pressure levels into account, but also the other
cardiovascular risk factors (Whitworth and Chalmers,
2004). Therefore, in addition to blood pressure,
comprehensive analysis of other risk factors in the
context information can help people better recognize
their blood pressure condition. The formation of
hypertension is related to a variety of factors,
including age, lifestyle, family history, etc. (Hong-
Tao et al., 2007; Kunes and Zicha, 2009). The
context information obtained by the monitoring
system contains risk factors and redundant
information that affects the classification accuracy.
Therefore, we need to screen the risk factors of
hypertension in the context information to improve
the classification effect of the monitoring system. In
this way, the patients who take medicine can be
properly guided to continue medication to prevent
Xie, D., Kang, G. and Chen, L.
A Risk Factors Screening Method in the Context-aware System of Hypertension.
DOI: 10.5220/0007672500490056
In Proceedings of the 4th International Conference on Internet of Things, Big Data and Security (IoTBDS 2019), pages 49-56
ISBN: 978-989-758-369-8
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
49
the deterioration of the disease. At the same time,
the study can assist doctors to diagnose patients'
blood pressure condition individually.
In previous studies, feature selection algorithms
were often used to screen for risk factors. Feature
selection methods are divided into selection
mechanisms: Filter, Wrapper, and Embedded. Filter
relies on the data samples themselves, with a low
time complexity. Ding and Peng (2005) proposed
mRMR (minimum-redundancy maximum-
relevancy) feature selection method. The algorithm
considers not only the correlation of each feature
attribute with respect to the class label, but also the
redundancy of the internal relationship of the feature
attributes set. Wrapper algorithms select features by
using learning algorithm results, with a better result
and a higher time complexity (Kohavi and John,
1997). Hu and Bao (2015) proposed a wrapper
feature selection algorithm for short-term load
forecasting (STLF) data, which has excellent
selection effect. The embedded algorithm takes
feature attribute selection as part of the training
process and has balanced efficiency and accuracy. In
recent years, it has become a research hotspot.
Koenigstein and Paqued (2013) proposed an
embedded algorithm based on Bayes matrix
factorization. The classification model can
automatically identify and use informational
features, and useless features can be subtracted.
For a single method, the screening rate and
accuracy are often not balanced. In this paper, we
propose a context-aware system to assess the
treatment effect in the intelligent monitoring of
hypertensive patients. For the decision making
module of the system, we propose a filter-embedded
feature selection method to screen risk factors of
hypertension. By using the support vector machine
(SVM) classification algorithm to verify the
screening effect, this method can improve the
screening rate and ensure the accuracy of this
system. The entire system therefore improves
performance, which can service hypertensive
patients preferably.
2 SYSTEM ARCHITECTURE
To make a comprehensive evaluation of
hypertensive patients’ treatment effect, blood
pressure data should not be the only reference. The
proposed monitoring system is context-aware and
takes into account other risk factors. Risk factors
screening method and data mining techniques are
utilized in classifying the treatment effect. The
overall architecture is shown in Figure 1. It consists
of three modules: data acquisition (DA), decision
making (DM), and diagnostic feedback (DF).
Figure 1: The proposed system architecture.
2.1 Data Acquisition Module
DA module gathers and stores patients’ medical
information. The way to obtain the context is
divided into measurement data acquisition, text data
acquisition and sample data acquisition according to
different acquisition paths and uses. Measurement
data is collected by hardware sensors. Text data is
obtained by textual input. Sample data is stored in
medical databases.
2.2 Decision Making Module
This is the core module of the entire system. It
determines the level of patient treatment
effectiveness. In the intelligent monitoring of
hypertension, it is necessary to evaluate the level of
blood pressure control status of hypertensive
patients, and make corresponding diagnostic
feedback according to the level. Data drawn from
the DA module must be preprocessed first. The data
preprocessing process is the core step and the focus
of this paper. It includes data cleaning, analysis,
feature selection, and evaluation. The specific
method will be detailed in Section 3. Then the
preprocessed data is put into the data mining
algorithms for training and classification to judge the
condition. By analyzing the patient's blood pressure,
age, weight and other contextual information, the
patients are divided into two groups by professional
doctors: Good Treatment (GT) and Poor Treatment
(PT). Finally the hypertensive patients are classified
into these two groups by the system based on their
risk factors. The accuracy of the DA module is
verified by classification results.
IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security
50
2.3 Diagnostic Feedback Module
DF module is a small inference system to generate
the feedback according to DM output and patient’s
relevant information. Knowledge Base contains all
the knowledge needed for the proper treatment of
hypertensive patients, varying from drug therapy to
non-drug therapy (mainly lifestyle modifications).
Rule Base is a set of rules regarding medical
diagnosis. The inference engine gives an automatic
feedback to the input treatment effect based on the
Knowledge Base and Rule Base. The feedback is
then returned to the patient, either by WeChat,
Email, or on the web.
3 SCREENING METHODS
The DM module extracts the hypertension risk
factors from the context information acquired from
the DA module and inputs them into the classifier to
obtain the patient's hypertension control condition.
There is a large amount of multi-dimensional
redundant information in the context information.
Redundant information will decrease the speed of
the system and interferes with normal decision
making. Therefore, it is particularly important to
filter risk factors from multi-dimensional context
information correctly and effectively. Accurate
extraction of features can increase the accuracy of
the classifier and reduce data dimensions. After each
step of screening, we use the same classification
algorithm to evaluate the screening results to test the
effect of the methods.
3.1 Data Preprocessing
This study screened 8619 people who met the
criteria as the data set from Beijing We-Health
Platform. The data included carcinoembryonic
antigen (CEA), sex, age, height, weight, white blood
cells, red blood cells, haemoglobin (HGB), red
blood cell specific volume (HCT), erythrocyte mean
corpuscular volume (MCV), mean corpuscular
haemoglobin (MCH), mean corpuscular
haemoglobin concentration (MCHC), red cell
distribution width coefficient variation (RDWCV),
red cell distribution width standard deviation
(RDWSD), platelet count (PLT), mean platelet
volume (MPV), platelet distribution width (PDW),
monocyte (MON), MON%, Granulocytes (GRA),
GRA%, lymphocyte (LYM), LYM%, alanine
aminotransferase (ALT), UREA, serum creatinine
(Cr), uric acid (UA), serum total cholesterol (TC),
triglyceride (TG), fasting plasma glucose (FPG),
specific gravity (SG), PH, a total of 32 dimensions.
Body mass index (BMI) can accurately reflect the
combined indicators of height and weight, so we
combined these two features into BMI. These data
were labelled to indicate the type, GT and PT.
According to statistics, there were about 47% of GT
data and 53% of PT data. The data contained
diastolic pressure and systolic pressure, but we did
not consider them to ensure the objectivity of the
risk factor screening process. We randomly divided
the data set into 70% training set and 30% test set. In
the following steps, the models were trained by 10-
fold cross validation. A total of 2586 data were used
for the tests. To guarantee the same ratio of the two
types of data, there were 1372 cases of GT data and
1214 cases of PT data.
3.2 Filter by PCC
The continuous variable correlation measure based
on Pearson’s correlation coefficient (PCC)
significance test can quickly test whether there is
linear correlation between data. The source of
context information used in this paper was the
physical examination information. The evaluation
index was mostly linear, so we used the PCC for
preliminary screening. For the given sample points,
we can calculate the PCC. This method can screen
out features that have little correlation with blood
pressure.
The calculation formula is as follows:

(1)
3.3 Random Forest Screening
Random forest algorithm has many advantages,
including high accuracy, robustness and ease of use,
making it one of the most popular machine learning
algorithms (Madan et al., 2008). It randomly selects
samples from the sample set and attributes from
all attributes. After that, it selects the best
segmentation attribute node to establish the CART
decision tree. The above steps are repeated m times
to establish m decision trees and obtain m classifiers
(Breiman, 2001).
However, this kind of impurity-based screening
method has a bias. Once a feature is selected, the
importance of other features will drop sharply
because the impurity has been lowered by the
selected feature. So other features are difficult to
reduce so much impurity. Therefore, the feature
A Risk Factors Screening Method in the Context-aware System of Hypertension
51
selected at first gets the high score while the score of
other related features gets low score, which is easy
to cause misunderstanding.
3.4 Random-forest Stability Selection
Stability selection is a method based on a
combination of subsampling and selection
algorithms. This method provides finite sample
control for some error rates of false discoveries and
hence a transparent principle to choose a proper
amount of regularization for structure estimation
(Nicolai and Peter, 2010).
We improved the screening method for the
instability of feature scores in random forest
algorithms. Firstly, the filter method was used. Then
we combined the random forest algorithm and the
subsampling to an improved screening method:
random-forest stability selection (RFSS). Its main
idea is to run random forest algorithm on different
data subsets and feature subsets, repeat constantly,
and finally aggregate feature selection results. We
used the frequency (the number of times selected as
an important feature divided by the number of times
its subset was tested) at which a feature was
considered to be an important feature as an
evaluation indicator. Ideally, important features
would score close to 100%. A slightly weaker
feature score would be a non-zero number, while the
most useless feature score would be close to zero.
3.5 Evaluation Algorithm Selection
For different screening methods, we used a unified
classification algorithm to verify the screening
effect. The support vector machine (SVM)
(Cristianini and Shawe-Taylor, 2000) has been
introduced as an efficient technique for solving
various function estimation problems, especially for
the pattern classification problems (Vapnik et al.,
1997). SVM is among the most robust and accurate
methods in all well-known data mining algorithms.
The final decision function of the SVM is
determined by only a few support vectors, and the
computational complexity depends on the number of
support vectors. SVM can verify whether we have
selected the most appropriate risk factors and avoids
the "curse of dimensionality". Therefore, we used
SVM to verify the screening effect.
4 EXPERIMENTAL RESULTS
We verified the accuracy of each screening method
by SVM algorithm, and judged the screening effect
of risk factors through multiple evaluation indicators
including sensitivity, specificity, accuracy and
dimension.
4.1 Performance Analysis
Table 1 shows the classification result of raw data
without using a screening method through 10-fold
cross-validation. 650 of the 1214 PT patients were
predicted to be accurate and the sensitivity is 54%.
The accuracy of this model is 66%. The risk factor
dimension is 32.
Table 1: Classification result without using a screening
method.
Actual
GT
PT
Total
Predicted
GT
1068
564
PT
304
650
Total
1372
1214
2586
Then we used the PCCs to filter features. We can
see the result in figure 2. We sorted 32 features
based on scores by PCCs. The feature with 0 value
indicates that the probability of a two-tailed test is
greater than 0.05, indicating that they cannot be used
as a risk factor. Meanwhile, the correlation values of
PLT, LYM, LYM%, MCV, MPV, and MON for
blood pressure were less than 0.05. This means that
these parameters, relative to other parameters, do not
directly reflect or affect a patient's blood pressure
condition. So we screened these parameters out
because they were basically not helpful for the
classification models.
Table 2 shows the result using the filter
screening method through 10-fold cross-validation.
The feature dimension reduced from 32 to 23, and
the data indicated that the sensitivity and accuracy of
the classification had increased significantly. 896
patients were predicted to be accurate and the
sensitivity is 73%. The accuracy of this model raised
to 75%. Dimension was 28% lower than before.
However, the number of risk factors was still too
large. Therefore, we used random forest screening
methods to reduce feature dimension and further
determine the risk factors on this basis.
Table 2: Classification result with using PCC.
Actual
GT
PT
Total
Predicted
GT
1073
318
PT
299
896
Total
1372
1214
2586
IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security
52
Figure 2: The scores by using PCC.
Figure 3 shows the scores of the features after
using the random forest method. As mentioned in
Section 3, once a feature was identified in the
random forest screening process, the scores of other
associated features may drop suddenly, leading to
filter out some useful features. If we chose features
with high score by the result, the sensitivity and
accuracy would decrease in the classification
process.
Figure 3: The scores by using random forest method.
Table 3 shows the classification result using random
forest method after filter method with 10-fold cross-
validation. The risk factors were further reduced to
13. The dimension was 59% lower than the original.
But only 742 PT patients were predicted to be
accurate. The sensitivity and the accuracy of this
model decreased to 62% and 71%.
Table 3: Classification result with using random forest.
GT
PT
Total
Predicted
GT
1085
472
PT
287
742
Total
1372
1214
2586
In order to solve this problem, we chose RFSS
method to calculate the key degree of each factor
with more objective scores. Figure 4 shows the
scores result by using the RFSS method.
Figure 4: The scores result by using the RFSS method.
The highest four features’ scores were 1.0, which
means that they were selected as useful features (the
score was affected by the regularization parameter
alpha) every time. The next few features’ scores
began to decline, but the decline was not particularly
sharp like the result by using the random forest
method. It can be seen that RFSS method was
helpful in overcoming overfitting and
misunderstanding data. Good features did not have
low scores due to feature correlation. So the method
is better than the method only using random forest.
Table 4 shows the classification effect by RFSS
method through 10-fold cross-validation. The risk
factor dimension was still 13, but at the same time,
we improved the classification sensitivity and
accuracy. 909 of the 1214 PT patients were
predicted to be accurate and the sensitivity raised to
75%. The accuracy of this model was 77%. This
method achieved the desired effect, that is, less
dimension and higher accuracy.
0
0,1
0,2
0,3
0,4
0,5
age
BMI
sex
TC
TG
SG
FPG
UA
Cr
HCT
HGB
RBC
UREA
ALT
CEA
WBC
GRA
GRA_
MCH
MON_
pH
MCHC
PLT
LYM_
_LYM
MCV
MPV
MON
RDWSD
RDWCV
PDW
systolic pressure diastolic pressure
0
0,2
0,4
0,6
0,8
1
1,2
age
sex
TC
TG
BMI
HCT
FPG
RBC
WBC
UA
UREA
Cr
HGB
CEA
MON_
ALT
GRA
MCH
SG
GRA_
pH
A Risk Factors Screening Method in the Context-aware System of Hypertension
53
Table 4: Classification result with using RFSS.
GT
PT
Total
Predicted
GT
1082
305
PT
290
909
Total
1372
1214
2586
4.2 Discussions
Finally, we selected 13 features to be risk factors for
our proposed context-aware system through the
RFSS method. They are age, sex, TC, TG, BMI,
HCT, FPG, RBC, WBC, UA, UREA, Cr and HGB.
The accuracy of the results can be supported by the
classification results and related literature.
In the last few years, there had been many
studies on rick factors of hypertension. The results
showed that age, sex (Virdis et al., 2002), BMI,
genetic factors, waist circumference (Ashwell et al.,
2012), TG (Wang, 2013), Urea (Pearson et al.,
2001), were the most important risk factors affecting
hypertension. This is consistent with the result of
selecting important features by using a random
forest algorithm.
Table 5 shows the comparison of classification
effect by selecting different features as risk factors
according to different screening methods. All
methods had been tested by 10-fold cross validation.
We can see that the sensitivity was low when we
used all the features for classification, and the high
dimension would affect the system performance.
After the filter method, the indicators had improved,
but the number of dimension still cannot make us
satisfied. Therefore, in the method III and method
IV, we adopted machine learning method, and the
dimension was all controlled to the same number 13
to facilitate the comparison of experimental results.
Compared with method I and III, the accuracy and
sensitivity of method IV were greatly improved.
Compared with method II, there was an advantage in
the number of dimensions.
In the study of Maryam Tayefi et al., (2017),
they used decision tree algorithm to study
hypertension related factors, and the accuracy can
reach 73%. Besides that, in the study of Wang et al.,
(2014), they used a logistic regression and artificial
neural network-based approach to predict
hypertension. The sensitivity and accuracy can reach
49% and 77%. For comparison, we used the same
methods to train the data set of this study through
10-fold cross-validation. As shown in Table 6, The
result for testing data set shows that the accuracy,
sensitivity and specificity of using decision tree
training model was 73%, 61%, 78%. And the neural
network algorithm could reach 75%, 72% and 79%.
As can be seen, our method performed better on
these three indicators and this method is most
suitable for feature screening.
Table 6: Results compared with other method.
Measure
Decision
Tree
Neural
Network
Method
RFSS
Sensitivity
61%
72%
75%
Specificity
78%
79%
78%
Accuracy
73%
75%
77%
Compared to other studies, the advantage of this
study is that the sample size is large and 8619 people
were used for modelling. Therefore, the method has
strong applicability. Another advantage of this
method is that the preselected physical indicators
had 32 indicators, which was large enough to screen
features. There are more iterations to increase the
amount of experiment, which can improve the
stability of the screening. However, there are some
shortcomings in the evaluation. As a medical data
study, we should also use statistical hypothesis
testing to consider the random fluctuation in the
results, so as to improve the accuracy and statistical
significance of the results. We will improve this part
in future studies.
The focus on this paper is to study how to screen
features, it is necessary to analyse why the accuracy
of method IV is not particularly high. Firstly, raw
data set was acquired from physical examination,
without specific data indicators like TOD, FHCD.
Table 5: The comparison of classification effect by using different screening methods.
Method
Feature Selection
Dimension
Algorithm
Accuracy
Sensitivity
Specificity
I
None
32
SVM
66%
54%
78%
II
PCC
23
SVM
75%
73%
78%
III
Random Forest
13
SVM
71%
62%
79%
IV
RFSS
13
SVM
77%
75%
78%
IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security
54
With the development of Internet of Things
technology, patients will have more complete data,
and the accuracy rate will increase. On the other
hand, this SVM classifier was only used to judge the
screening effect of each method. As mentioned in
Chapter 2, the system classification module used
weighted algorithm. Therefore, the classification
effect of the whole system will be more accurate.
5 CONCLUSION
In this paper, we designed a risk factor screening
module using different screening methods based on
the proposed context-aware system for hypertension.
After comparison and improvement, we selected the
RFSS method combined by random forest and
stability selection in four methods. We gradually
filtered 32 parameters in context information
obtained from DA module to 13 hypertension risk
factors, and performed by SVM classification
algorithm. Accuracy, sensitivity and specificity have
been improved. This method can improve the
screening rate and ensure the accuracy of this
system. Therefore, the context-aware system of
hypertension will improve performance by using this
screening method. The output of the system can
assist doctors and patients to have a comprehensive
understanding of their blood pressure condition
according to risk factors.
In the future research, we will increase the
dataset and data parameters with the improvement of
the performance of portable devices. On this basis,
we will extend this work by applying DL
technologies such as CNN, in order to see whether
the accuracy can be increased. In addition, statistical
hypothesis testing will be added to experimental
verification and the results will be compared with
clinical practice. Finally, the automated hypertension
risk assessment approach will be improved based on
future study. The diagnosis of doctors will be more
accurate and comprehensive.
ACKNOWLEDGEMENTS
This work was supported by National Natural
Science Foundation of China (No.61471064), and
National Science and Technology Major Project of
China (No.2017ZX03001022).
REFERENCES
Lim S. S., Vos T, Flaxman AD, Danaei G, Shibuya K,
Adair-Rohani H. et al., 2012. A comparative risk
assessment of burden of disease and injury attributable
to 67 risk factors and risk factor clusters in 21 regions,
19902010: a systematic analysis for the Global
Burden of Disease Study 2010. The Lancet,
380(9859), pp.2224-2260.
Chen, W., R.Gao, L.Liu, 2014. China Cardiovascular
Disease Report 2013. Chinese Circulation Journal,
29(07), pp.487-491.
Carey R. M, Whelton P K, 2018. Prevention, Detection,
Evaluation, and Management of High Blood Pressure
in Adults: Synopsis of the 2017 American College of
Cardiology/American Heart Association Hypertension
Guideline. Annals of Internal Medicine.
Whitworth J. A, Chalmers J., 2004. World health
organisation-international society of hypertension
(WHO/ISH) hypertension guidelines. Clinical and
Experimental Hypertension, 26, pp. 747-752.
Guixia Kang, Meikui Zhang, 2010. Recent Advances of
Wireless eHealth (WeHealth) in Terms of ’Sensing
China’. Proceedings of the 5th International
Conference on Cooperation and Promotion of
Information Resources in Science and Technology,
pp.517-519.
Sandi, G., I. G. B. B. Nugraha and S. H. Supangkat, 2013.
Mobile health monitoring and consultation to support
hypertension treatment. International Conference on
ICT for Smart Society, pp. 1-5.
Hong-Tao, Y. U., J. L. Wang and L. I. LiLua, 2007.
Research on the risk factors of hypertension and
OSAHS in the snore patients with different living
habits. Journal of Clinical Pulmonary Medicine.
Kunes, J. and Zicha, J., 2009. The interaction of genetic
and environmental factors in the ethology of
hypertension. Physiological Research, 58, pp. S33.
Ding C., Peng H., 2005. Feature selection based on mutual
information criteria of max-dependency, max-
relevance, and min-redundancy. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 27(8), pp.
1226-1238.
Kohavi R., John G. H., 1997. Wrappers for feature subset
selection. Artificial Intelligence, 97(1-2), pp.273-324.
Hu Z. Y, Bao Y. K., 2015. Hybrid filter-wrapper feature
selection for short-term load forecasting. Engineering
Applications of Artificial Intelligence, 40, pp.17-27.
Koenigstein N., Paqued U., 2013. Xbox movies
recommendations: Variational Bayes matrix
factorization with embedded feature selection.
Proceedings of the 7th ACM conference on
Recommender Systems.
Madan, A. K., H. Dureja S. Gupta et al., 2008.
Topological models for prediction of pharmacokinetic
parameters of cephalosporins using random forest,
decision tree and moving average analysis. Sci Pharm,
76, pp.377394.
Breiman, L., 2001. Random forest. Machine Learning, 45,
pp. 532.
A Risk Factors Screening Method in the Context-aware System of Hypertension
55
Nicolai M, Peter B, 2010. Stability selection. Journal of
the Royal Statistical Society, 72(4), pp.417-473.
Cristianini N, Shawe-Taylor J, 2000. An introduction to
support vector machines and other kernel-based
learning methods. Cambridge University Press, UK,
1
st
edition.
V.Vapnik, S. Golowich, and A. Smola, 1997. Support
vector method for function approximation, regression
estimation and signal processing, in Advances in
Neural Information Processing Systems. MIT Press,
Cambridge.
Virdis, A., L. Ghiadoni, I. Sudano and et al, 2002.
Endothelial function in hypertension: role of gender.
Journal of Hypertension Supplement Official Journal
of the International Society of Hypertension, 20(2), pp.
116.
Ashwell, M., Gunn, P. and Gibson, S., 2012. Waist-to-
height ratio is a better screening tool than waist
circumference and BMI for adult cardiometabolic risk
factors: systematic review and meta-analysis. Obesity
Reviews, 13, pp. 275286.
Wang, X. X., 2013. Relationship of essential hypertension
prognisis with serum TG, TC concentration. Journal
of Hainan Medical University.
Pearson, D. L., Dawling, S., Walsh, W. F. et al, 2001.
Neonatal pulmonary hypertensionurea-cycle
intermediates, nitric oxide production, and carbamoyl-
phosphate synthetase function. N Engl J Med, 344, pp.
18321838.
Tayefi, M.; H. Esmaeili, K. M. Saberi and et al, 2017. The
application of a decision tree to establish the
parameters associated with hypertension. Computer
Methods & Programs in Biomedicine, 139, pp. 8391.
Wang, A.; N. An, Y. Xia, L. Li and G. Chen, 2014. A
Logistic Regression and Artificial Neural Network-
Based Approach for Chronic Disease Prediction: A
Case Study of Hypertension. 2014 IEEE International
Conference on Internet of Things (iThings), and IEEE
Green Computing and Communications (GreenCom)
and IEEE Cyber, Physical and Social Computing
(CPSCom), pp. 45-52.
IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security
56